<%BANNER%>

Noise Reduction Techniques Applied to the Detection of Explosive Materials for Improved Airport Screening

Permanent Link: http://ufdc.ufl.edu/UFE0042345/00001

Material Information

Title: Noise Reduction Techniques Applied to the Detection of Explosive Materials for Improved Airport Screening
Physical Description: 1 online resource (97 p.)
Language: english
Creator: Vishwanathan, Karthik
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: differentialreflectometry, explosives, hybridfilter, imaging, limitofdetection, snr, trinitrotoluene
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, M.S.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Science NOISE REDUCTION TECHNIQUES APPLIED TO THE DETECTION OF EXPLOSIVE MATERIALS FOR IMPROVED AIRPORT SCREENING By Karthik Vishwanathan December 2010 Chair: Ant Ural CoChair: Rolf. E. Hummel Major: Electrical and Computer Engineering Counter terrorism is the practice and technique that governments and militaries adopt to prevent terror threats. The illegal use of explosives has become a daily event in the modern world. For terrorist organization and insurgents, explosive devices have become a weapon of choice, which are being used to kill or harm civilians, military personnel and equipment/infrastructure. Hence, the detection of explosives is a paramount issue in security. Keeping this in mind, our group has developed an explosives detector. Our explosives detector can detect trace quantities of explosives on the surface of a baggage using the technique called differential reflectometry. This technique measures the optical absorption of energetic materials. These optical characteristics consist of characteristic shoulders between 370 nm and 490 nm; for example at 420 nm for Trinitrotoluene (TNT) and 480 nm for Triaminotrinitrobenzene. From our prototype, the differential reflectogram of these explosives has results at these characteristic shoulders. Our prototype was automated to increase the overall efficiency of the system. The explosives detection is a combination of several components: light source, camera, spectrometer and a computer with dedicated software. In this study, noise were produced and hence noise characterization study for each of them becomes vital for knowing its exact effect on the signal. Our goal is to reach high sensitivity, i.e., a very good Signal-to-Noise ratio. Understanding the noise of our system will allow us to propose several strategies to improve such a Signal-to-Noise ratio. In order to reduce the noise, several imaging filters (median, average, disk, Gaussian, moving average and hybrid filters) were tested. A comparative study was made among the filters using Signal-to-Noise Ratio (SNR) as a metric. It was discovered that a combination of two filters, namely median filter and moving average filter, having filter size 10 and 4 respectively, lead to the best results. In conclusion, our study was helpful in achieving our goal (increasing the sensitivity of our explosive detection prototype). With improved noise reduction, it is a two-fold advantage. First of all, the limit of detection can be improved and secondly, the false positive alarm rates can be reduced which, in turn, will increase the efficiency of our explosives detector
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Karthik Vishwanathan.
Thesis: Thesis (M.S.)--University of Florida, 2010.
Local: Adviser: Ural, Ant.
Local: Co-adviser: Hummel, R. E.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2011-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0042345:00001

Permanent Link: http://ufdc.ufl.edu/UFE0042345/00001

Material Information

Title: Noise Reduction Techniques Applied to the Detection of Explosive Materials for Improved Airport Screening
Physical Description: 1 online resource (97 p.)
Language: english
Creator: Vishwanathan, Karthik
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: differentialreflectometry, explosives, hybridfilter, imaging, limitofdetection, snr, trinitrotoluene
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, M.S.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Science NOISE REDUCTION TECHNIQUES APPLIED TO THE DETECTION OF EXPLOSIVE MATERIALS FOR IMPROVED AIRPORT SCREENING By Karthik Vishwanathan December 2010 Chair: Ant Ural CoChair: Rolf. E. Hummel Major: Electrical and Computer Engineering Counter terrorism is the practice and technique that governments and militaries adopt to prevent terror threats. The illegal use of explosives has become a daily event in the modern world. For terrorist organization and insurgents, explosive devices have become a weapon of choice, which are being used to kill or harm civilians, military personnel and equipment/infrastructure. Hence, the detection of explosives is a paramount issue in security. Keeping this in mind, our group has developed an explosives detector. Our explosives detector can detect trace quantities of explosives on the surface of a baggage using the technique called differential reflectometry. This technique measures the optical absorption of energetic materials. These optical characteristics consist of characteristic shoulders between 370 nm and 490 nm; for example at 420 nm for Trinitrotoluene (TNT) and 480 nm for Triaminotrinitrobenzene. From our prototype, the differential reflectogram of these explosives has results at these characteristic shoulders. Our prototype was automated to increase the overall efficiency of the system. The explosives detection is a combination of several components: light source, camera, spectrometer and a computer with dedicated software. In this study, noise were produced and hence noise characterization study for each of them becomes vital for knowing its exact effect on the signal. Our goal is to reach high sensitivity, i.e., a very good Signal-to-Noise ratio. Understanding the noise of our system will allow us to propose several strategies to improve such a Signal-to-Noise ratio. In order to reduce the noise, several imaging filters (median, average, disk, Gaussian, moving average and hybrid filters) were tested. A comparative study was made among the filters using Signal-to-Noise Ratio (SNR) as a metric. It was discovered that a combination of two filters, namely median filter and moving average filter, having filter size 10 and 4 respectively, lead to the best results. In conclusion, our study was helpful in achieving our goal (increasing the sensitivity of our explosive detection prototype). With improved noise reduction, it is a two-fold advantage. First of all, the limit of detection can be improved and secondly, the false positive alarm rates can be reduced which, in turn, will increase the efficiency of our explosives detector
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Karthik Vishwanathan.
Thesis: Thesis (M.S.)--University of Florida, 2010.
Local: Adviser: Ural, Ant.
Local: Co-adviser: Hummel, R. E.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2011-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0042345:00001


This item has the following downloads:


Full Text

PAGE 1

1 NOISE REDUCTION TECHNIQUES APPLIED TO THE DETECTION OF EXPLOSIVE MATERIALS FOR IMPROVED AIRPORT SCREENING By KARTHIK VISHWANATHAN A THESIS PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE UNIVERSITY OF FLORIDA 2010

PAGE 2

2 2010 Karthik Vishwanathan

PAGE 3

3 In memory of my loving father

PAGE 4

4 ACKNOWLEDGMENTS I would like to express the deepest appreciation to my advisor and an extremely wonderful person, Prof. Rolf. E. Hummel. I still remember the day when we had our first discussion, after which I prayed that he would select me as his student. His passion in learning science even now has always been my source of inspiration I thank him for providing me with the opportunity to explore all ideas related to my research. I was, am range of subjects. Thank you Prof. Hummel for all your academic and non academic support I am deeply indebted to Dr. Thierry Dubroca for his constant encouragement In addition to being my supervisor and mentor, he has been my biggest motivator throughout my degree program. Whenever I am down and feel exhausted, he would also be there to support me. Without him, I am quite ce rtain that I could not have gone this far in my research work and finished my thesis in time He was instrumental in Dubroca was to me. He has been the best mentor a student could ever dream of. Thierry, you are the best! I would like to thank Prof. Ant Ural who has time and again, bailed me out of problems each semester, starting from my first. During my most difficult times, he readily agreed to be my chair, which I feel was very gra cious of him. I would also like to thank Prof. Martin Uman, who a greed to serve on my committee and provide me with his valuable suggestions and comments. To my mother and sister who have g iven me unconditional and positive support in every endeavor I have attempted I wish to express my deepest love, admiration and

PAGE 5

5 strongest thanks. The confidence I have in my abilities and myself, I owe to thei r unyielding encouragement. I thank all the Go ds for giving me such a wonderful mother and sister I would also like to thank all my family members for their unconditional love. I would like to take this opportunity to thank my brother, who was the first person to make me understand what research mean s after all. He always holds a special place. I am also grateful to my colleagues Michael Croft (He helped me a lot with the experiments carried out in C hapter 4) and Max Lemaitre, who have always given me lot of suggestions during our weekly group meeting s. I would also like to thank all my friends who have been very supp ortive throughout

PAGE 6

6 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 9 LIST OF FIGURES ................................ ................................ ................................ ........ 10 ABSTRACT ................................ ................................ ................................ ................... 14 CHAPTER 1 INTRODUCTION TO THE PRESENTED RESEARCH ................................ ........... 16 1.1 Motivation ................................ ................................ ................................ ......... 16 1.2 Approach: Explosives Detector Baggage Scanner ................................ ........... 16 2 PROPERTIES OF ENERGETIC MATERIALS ................................ ........................ 18 2.1 Explosive Material ................................ ................................ ............................. 18 2.1.1 Definition ................................ ................................ ................................ 18 2.1.2 Physical Properties of 2, 4, 6 Trinitrotoluene ................................ ........... 18 2.1.2 Physical Properties of TriAminoTrinitroBenzene ................................ ..... 19 2.2 Optical Properties of Energetic Materials ................................ .......................... 20 2.3 Detection Method: Differential Reflection Spectromet ry ................................ .... 21 3 EXPLOSIVE DETECTOR BAGGAGE SCANNER: AN OVERVIEW ....................... 22 3.1 Explosives Detector : System Description ................................ ......................... 22 3.2 Description of the Light Source ................................ ................................ ......... 22 3.3 Spectrometer ................................ ................................ ................................ .... 23 3.4 Charge Coupled Device (CCD) Camera ................................ ........................... 25 3.6 Experimental Setup: Explosives Detector Baggage Scanner ........................... 25 3.7 Mathematical Analysis using Image Acquisition ................................ ............... 27 3.8 Differential Reflectogram of Trinitrotol uene ................................ ....................... 29 3.9 Summary ................................ ................................ ................................ .......... 30 4 CHARACTERIZATION OF SYSTEM PARAMETERS ................................ ............ 31 4.1 Signal To Noise Ratio: Metric for Comparative Study ................................ ...... 31 4.2 Camera Exposure Time ................................ ................................ .................... 32 4.3 Camera Gain and Offset Variation Effect ................................ .......................... 34 4.4 Effects of the Distance between Collection Optics and Sample ........................ 41 4.5 Summary ................................ ................................ ................................ .......... 43

PAGE 7

7 5 EXPLOSIVES DETECTOR NOISE CHARACTERIZATION OF COMPONENTS ................................ ................................ ................................ ...... 44 5.1 Charge Coupled Device Camera Noise Characterization ................................ 44 5.2 Sources of Noise from Charge Coupled Device Camera ................................ 44 5.2.1 Readout Noise ................................ ................................ ......................... 45 5.2.2 Dark Noise ................................ ................................ ............................... 45 5.2.3 Noise Equation ................................ ................................ ........................ 45 5.3. Charge Coupled Device Camera Noise Characterization: Experimental Procedure ................................ ................................ ................................ ............ 46 5.3.1 Readout Noise Measurement ................................ ................................ .. 46 5.3.2 Dark Noise Measurement ................................ ................................ ........ 46 5.3.3 Experimental Results ................................ ................................ ............... 46 5.4 Light Source Noise Characterization ................................ ................................ 48 5.4.1 Experimental Procedure for Noise Characterization of Light Sources ..... 48 5.4.2 Lamp Off Characterization Control Experiment ................................ .... 49 5.4.3 Quartz Tungsten Halogen ................................ ................................ ....... 50 5.4.4 75W Xenon Lamp ................................ ................................ .................... 52 5.4.5 White Light Emitting Diode ................................ ................................ ...... 54 5.4.6 Xenon Arc Lamp (300W) ................................ ................................ ......... 56 5.5 Summary ................................ ................................ ................................ .......... 59 6 DENOISING STRATEGIES ................................ ................................ .................... 62 6 1.Method to Reduce Noise during Frame Acquisition ................................ .......... 62 6 2.Post Processing Filtering Methods ................................ ................................ ... 66 6.2.1 Median Filter ................................ ................................ ............................ 67 6.2.1.1 Understanding Median Filter ................................ .......................... 67 6.2.1.2 Experimental Results ................................ ................................ ..... 68 6.2.2 Average Filter and Disk Filter ................................ ................................ .. 69 6.2.2.1 Understanding Average Filter and Disk Filter ................................ 70 6.2.2.2 Experimental Results ................................ ................................ ..... 71 6.2.3 Gaussian Filter ................................ ................................ ........................ 76 6.2.3. 1 Understanding Gaussian Filter ................................ ....................... 77 6.2.3.2 Experimental Results ................................ ................................ ..... 77 6.2.4 Moving Average Filter ................................ ................................ .............. 79 6.2.4.1 Understanding Moving Average Filter ................................ ............ 79 6.2.4.2 Experimental Results ................................ ................................ ..... 81 6.2.4.3 Effect of Spectral Shift on the Differential Reflectogram of TNT .... 83 6.3 Custom based Hybrid Filter ................................ ................................ .............. 84 6.4. Summary ................................ ................................ ................................ ......... 87 7 CONCLUSION and FUTURE WORK ................................ ................................ ..... 88 7.1. Conclusion ................................ ................................ ................................ ....... 88 7.2 Future Work ................................ ................................ ................................ ...... 89

PAGE 8

8 APPENDIX A GRAPHICAL USER INTERFACE ................................ ................................ ........... 91 LIST OF REFERENCES ................................ ................................ ............................... 94 BIOGRAPHICAL SKETCH ................................ ................................ ............................ 97

PAGE 9

9 LIST OF TABLES Table page 6 1 Comparison of image processing filters. ................................ ............................. 87

PAGE 10

10 LIST OF FIGURES Figure page 2 1 Single molecule structure of TNT at room temperature ................................ ...... 19 2 2 Single molecule structure of TATB at room temperature. ................................ ... 20 2 3 Electromagnetic radiation spectrum ................................ ................................ 21 3 1 Model of our explosives detector. ................................ ................................ ...... 22 3 2 300W Xenon light source used for our explosives detector ................................ 23 3 3 Photograph of our custom design spectrometer from Jobin Yvon. ..................... 24 3 4 Schematic of our custom made spectrometer. ................................ ................... 24 3 5 Pictur e of Charge Coupled Device camera (Sarnoff model CAM UV512). ......... 25 3 6 Schematics of the explosives detector. ................................ .............................. 26 3 7 Example of a sample and its placement on the conveyor belt ............................ 27 3 8 Schematic of light (blue) collection from sample on conveyor belt. .................... 28 3 9 Differential reflectogram of TNT obtained by differential reflectometry. .............. 29 3 10 Differential reflectogram of TATB obtained by differential reflectometry. ............ 30 4 1 Frame acquired at (a) 99ms exposure time and (b) 50ms exposure time. ......... 32 4 2 Differential reflectograms of TNT as a function of exposure time ....................... 33 4 3 Signal to Noise Ratio as a function of exposure time. ................................ ........ 34 4 4 Intensity as a function of the wavelength for various gain settings from 100 to 1023.. ................................ ................................ ................................ ................. 36 4 5 Intensity as a function of the gain settings of a point pixel at 410 nm, 420 nm, 480 nm and 520 nm. ................................ ................................ ........................... 37 4 6 Differential reflectogram s of TNT as a function of the gai n settings .................... 38 4 7 Signal to Noise Ratio as a function of the gain settings.. ................................ ... 39 4 8 Differe ntial reflectograms of TNT as a function of the o ffset settings .................. 40 4 9 Signal to Noise Ratio as a function of the offset settings. ................................ .. 40

PAGE 11

11 4 10 Differential reflectogram of TNT as a function of distance between the collection optic and the sample. ................................ ................................ .......... 42 4 11 Signal (Blue diam onds) and Noise (Red squares) as a function of the distance between collection optics and sample.. ................................ ................ 42 4 12 Signal to Noise Ratio a s a function of the distance between collection optics and sample (blue diamonds). ................................ ................................ ............. 43 5 1 Characterization of noise from camera.. ................................ ............................. 47 5 2 Detector voltage as a function of time when no lamp is used (Control experiment).. ................................ ................................ ................................ ....... 49 5 3 Fast Fourier Transform amplitude of detector voltage as a function of time for the control experiment. ................................ ................................ ....................... 50 5 4 Detector voltage as a function of time for the QTH Lamp at 50W. ...................... 51 5 5 Fast Fourier Transform amplitude of detector voltage as a function of time for the QTH Lamp. ................................ ................................ ................................ ... 52 5 6 Detector voltage as a function of time for 75W Xenon light. ............................... 53 5 7 Fast Fourier Transform amplitude of detector voltage as a func tion of time for a 75W Xenon light source.. ................................ ................................ ................. 53 5 8 Detector voltage as a function of time for a white LED.. ................................ ..... 55 5 9 Fast Fourier Transform amplitude of detector voltage as a function of time for white LED. ................................ ................................ ................................ .......... 55 5 10 Detector voltage as a function of time for Xenon 300W light source immediately after turn on. ................................ ................................ ................... 56 5 11 Fast Fourier Transform amplitude of detector voltage as a function of time for Xenon 300W immediately after turn on. ................................ ............................. 57 5 12 Detector voltage as a function of time for Xenon 300W light source 30mins after turn on. ................................ ................................ ................................ ....... 58 5 13 Fast Fourier Transform amplitude of detector voltage as a function of time for Xenon 300W 30mins after turn on ................................ ................................ ..... 59 5 14 Characterization of noise from light sources. ................................ ...................... 60 6 1 Schematic of the image acquisition and storage modes. ................................ .... 62 6 2 Timing of image acquisition on the conveyor belt. ................................ .............. 64

PAGE 12

12 6 3 Proposed methodology to reduce the eff ect of noise and increase SNR.. .......... 65 6 4 Differential reflectograms of TNT when one frame is acquired (blue) and acquisition of two fram es (red). ................................ ................................ ........... 66 6 5 Schematic working of the median filtering.. ................................ ........................ 67 6 6 Differential reflectograms of TNT with increasing sizes of median filter.. ............ 68 6 7 SNR as a function of the median filter size for differential reflectograms of TNT.. ................................ ................................ ................................ .................. 69 6 8 Schematic principle of the average filter (filter size = 5). ................................ .... 70 6 9 Representation of the disk filter with disk size =3. ................................ .............. 71 6 10 Differential reflectograms of TNT with increasing sizes of disk filters.. ............... 72 6 11 Schematic of feature amplitudes as a function of the frequency for the disk filter of various sizes.. ................................ ................................ ......................... 73 6 12 SNR as a function o f the disk filter size for the differential reflectograms of TNT.. ................................ ................................ ................................ .................. 74 6 13 Differential reflectograms of TNT with increasing sizes of average filter.. .......... 75 6 14 SNR as a function of the average filter size for differential reflectograms of TNT. ................................ ................................ ................................ ................... 76 6 15 Differential reflectograms of TNT with increasing sizes of Gaussian filt er ......... 78 6 16 SNR as a function of the Gaussian filter size for differential reflectograms of TNT. ................................ ................................ ................................ ................... 79 6 17 Schematic of the Moving Average Filter (filter size 4). ................................ ........ 80 6 18 Differential reflectograms of TNT with increasing sizes of moving average filter.. ................................ ................................ ................................ ................... 82 6 19 SNR as a function of the moving average filter size for the diff erential reflectograms of TNT. .. ................................ ................................ ...................... 82 6 20 The signal (shoulder) shifting towards the larger wavel engths when the size of the moving average filter increases. ................................ ............................... 83 6 21 Inflection point as a function of the filter size from differential reflectograms of TNT.. ................................ ................................ ................................ .................. 84

PAGE 13

13 6 22 Differential reflectogram of TNT and application of different filters as marked on th e graph.. ................................ ................................ ................................ ..... 86 A 1 GUI window for CCD Camera settings control. ................................ ................... 91 A 2 GUI window for selecting the camera pixel depth (8, 10 or 12 bits). ................... 92 A 3 GUI window for frame acquisiti on from the camera video stream. ...................... 92

PAGE 14

14 Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Science NOISE REDUCTION TECHNIQUES APPLIED TO THE DETECTION OF EXPLOSIVE MATERIALS FOR IMPROVED AIRPORT SCREENING By Karthik Vishwanathan December 2010 Chair: Ant Ural Co C hair: Rolf. E. Hummel Major: Electrical and Computer Engineering Co unter terrorism is the practice and technique that governments and militaries adopt to prevent terror threats. The ill egal use of explosives has become a daily event in the modern world. For terrorist organization and insurgents, explosive devices have become a weapon of choice, which are being used to kill or harm civilia ns military personnel and equipment/infrastructure Hence, t he detection of explosives is a paramount issue in security. Keeping this in mind, our group has developed an explosive s detector. Our explosive s detector can detect trace quantities of explosives on the surface of a baggage using the technique called d ifferential r eflectometry. This technique measures the optical absorption of energetic mat erials. These optical characteristics consist of characteristic shoulders between 370 nm and 49 0 nm; for example at 420 nm for Tr initrotoluene (TNT) and 4 8 0 nm for Triaminotrinitrobenzene From our prototype, the differential r eflectogram of these explosiv es has results at these characteristic shoulders. Our prototype was automated to increase the overall efficiency of the system

PAGE 15

15 The explosive s detection is a combination of several components: l ight s ource, camera, spectrometer and a computer with dedicated software. In this study, noise were produced and hence noise characterization study for each of them becomes vital for knowing its exact effect on the signal. Our goal is to reach high sensitivity, i.e., a very good Signal to Noise ratio U nderst anding the noise of our system will allow us to propose several strategies to improve such a Signal to Noise ratio. In order to reduce the noise, several imaging filters ( median average, disk, Gaussian, moving average and hybrid filters) were tested A co mparative study was made among the filters using Signal to Noise Ratio (SNR) as a metric. It was discovered that a combination of two filters, namely m edian filter and m oving a verage filter having filter size 1 0 and 4 respectively, lead to the best result s In conclusion, our stud y w as helpful in achieving our goal (increasing the sensitivity of our explosive detection prototype) With improved noise reducti on, it is a two fold advantage. First of all the limit of detection can be improved and s econd ly the false positive alarm rates can be reduced which in turn, w ill increase the efficiency of our explosive s detector

PAGE 16

16 CHAPTER 1 INTRODUCTION TO THE PRESENTED RES EARCH 1.1 Motivation For several decades acts of terrorism have been rampant and weapons of mass destructions have been used [1] T he July 7 2005 London attack and the recent November 26 2008, Mumbai India terror attacks have a deep impact in our daily lives. In all these attacks, one of the underlying areas of fault was the security lapse [2 3] Hence it is highly critical to have a robust application which detects explosives to e nsure the s afety of human live s Our research group under the guidance of Prof. Hummel has developed a n explosives detection prototype using a technique called Differential Reflectometry 1.2 Approach: Explosive s Detector Baggage Scanner The purpose of the explosive s detector baggage scanner is to identify and detect explosives on luggage surface s The principal material used for this study is 2 4, 6 Trinitro t oluene (TNT) This system is fully automated with the help of image processing software algorithms Chapter 2 introduc es TNT and Triaminot rinitrobenzene ( TATB) and their respective chemistr y T he concept of d ifferential r eflectometry is reviewed and applied to the detection of explosives In C hapter 3, the var ious components that make up our explosive s detector and the i mage processing algorithms are presented. Additionally a comparison of other techniques with our new prototype is presented.

PAGE 17

17 In C hapter 4, a system c haracterization is performed in order to study the parameters (Exposure time, gain, offset and d istance be tween c ollection o ptics and s ample ) which influences the overall output of the explosive s detector. The presence of noise affects the behavior of our explosive s detector (for example increasing the number of false positives) The various sources of noise are examined and studied in detail in C hapter 5 In order to reduce the noise, various filter design were tested. Our studies show how such filters reduce the influence of noise. Chapter 6 presents the details of each filter and compares them with one another. Finally, C hapter 7 summarizes our study and present s futur e work regarding the possible improve ment s of our system

PAGE 18

18 CHAPTER 2 PROPERTIES OF ENERGE TIC MATERIALS 2.1 Explosive Material T he first widely used explosive in warfare and mining was black powder even though thermal weapons have existed previously During the 19 th century, black powder was replaced by nitr o glycerin and nitro cellulose. Eventually, they were r eplaced by modern explosives such as 2, 4, 6 Trinitro toluene ( TNT ) and Triaminotrinitrobenzene ( TATB ) [ 4 ] 2.1.1 Definition An explosive material is a substance that contains a great amount of stored energy that can produce a sudden expansion of the material after initiation, usually accompanied by the produ ction of light, heat, sound and pressure. They are characterized by their sensitivities [ 5 ] Sensitive materials that can be initiated by a relatively small amount of heat or pressure are called primary explosives ( Lead Azide ) and materials that are relati vely insensitive are called secondary explosives. TNT and TATB are examples of secondary explosives Explosives are also called energetic materials. The l arger the heat of formation, the great er the amount of energy is stored Explosion of the material is caused by the release of energy upon decomposition. 2.1.2 Physical Properties of 2, 4, 6 Trinitrotoluene TNT is one of the most commonly used explosives for military and industrial purposes [6] It is highly valued because o f its inse nsitivity to shock and friction, which reduces the risk of accidental detonation. It is highly stable compared to other high explosives.

PAGE 19

19 Figure 2 1. S ingle molecule structure of TNT at room temperature with monoclinic TNT molecule conformation [7 ] The molecular structure of TNT at room temperature is shown in Figure 2 1 as it appears in the molecular crystals (either monoclinic or orthorhombic) TNT increases the explosive power in improvised or home made explosives made from ammon ium nitrate fuel and is considered an ignition source in this case. 2.1.2 Physical Properties of TriAminoTrinitroBenzene TATB is one of the most powerful e xplosive kno wn. It is extremely insensitive to shock and vibration. The high stability of TATB favors its use in military and civilian applications when insensitive high explosives are required. TAT B is among the many explosive s that belong to a family which are nitrate based [ 8 ]. The molecular structure of TATB is shown in Figure 2 2.

PAGE 20

20 Figure 2 2 Single molecule structure of T ATB at room temperature. 2.2 Optical Properties of Energetic Materials Optical properties are quite significant because they provide a nondestructive means for providing the characteristi cs of each energetic material. It has been observed that due to the chemical compositions eac h energetic material has a unique optical property in the ultraviolet region. T he ultraviolet region of the electromagnetic spectrum is generally defined to be between 10 nm to 400 nm and the visible region from 400 nm to 750 nm as shown in Figure 2 3 Our region of interest is the near UV Visible spectrum between 200 nm and 5 5 0 nm. Explosive materials such as TNT and TATB have characteristics features in the ultraviolet and the visible region [7].

PAGE 21

21 Figure 2 3 Electromag n etic radiation spectrum. Our region of interest is between 200 nm and 55 0 nm 2.3 Detection Method: Differential Reflection Spectrometry There are many methods for the detection of explosives such as ion mobility, Millimeter wave technology and Raman Scanner Even though these tec hniques are in current use, there is a great demand for a simple, fast, accurate and contact less detection system [7]. D ifferential reflection spectrometry (DRS), also called differential reflectometry was developed at the University of Florida in 1970 by Prof. Rolf. E. Hummel It is a fast, simple and non destructive surface analytical technique [9 11] I t electronic structure of materials. differential mode, it removes instrumental varia tions [7 ] [12 15] The first experimental work [16 17 ] carried out on a bench top d ifferential r eflectometry system which consists of a 75 W Xe non light source, a vibrating mirror and a photomultiplier tube ( PMT). The current system which also uses d ifferential r eflectometry technique employs a different geometry tailor made to make multiple measurements at a time, adapt ed to the airport security needs

PAGE 22

22 CHAPTER 3 EXPLOSIVE DETECTOR BAGGAGE SCANNER : AN OVERVIEW The design of our explosive s detector has been desi g n ed so that it can be added to the current X ray airport baggage scanner. A detailed schematic of our prototype is given in this chapter. 3.1 Explosive s Detector : System Description As shown in Figure 3.1., the explosives detector consists of a light source (Ultra Violet Blue light), a spectrometer and a detector ( camera ). These components can be fitted in conjunction with a conventional airport baggage scanner (conveyor belt and X r ay machine ). Figure 3 1. M odel of our explosives detector U ltra Violet Blue light shines on the baggage moving on the conveyor. The detector analyzes the reflected light from the baggage. 3.2 Description of the Light Source We u sed a 300 W Xenon light source as shown in Figure3.2 The majority of the light is generated within a 1 mm cloud of plasma situated at the point where the electron

PAGE 23

23 stream leaves the face of the cathode Electrons passing through the plasma cloud strike the anode, causing it to he at. Figure 3 2 300 W Xenon light s ource used for our explosive s detector which produces light in the UV blue light region. The light source has a bluish white spectrum and high UV output with a cut off wavelength near 300 nm This is of particular interest to our prototype system since the main structure of TNT and other explosives lie in the near UV region. 3.3 Spectrometer A spectrometer ( shown in Figure 3 3 ) is an instrument used to diffract light over a portion o f the ele ctromagnetic spectrum. The design was custom made for our application by Jobin Yvon ( France )

PAGE 24

24 Figure 3 3 Photograph of our custom design s pectrometer from Jobin Yvon. Figure 3 4 Schematic of our c ustom made s pectrometer Figure 3 4 shows a working schematic of our spectrometer. It operates in the range between 200 nm 550 nm and has a resolution better than 6 nm with a 100

PAGE 25

25 micrometer slit width at the entrance. The exit produces a dispersed beam such that 1 nm wavelength is spread ov er 1 8 m (1 pixel width on the camera) 3.4 Charge Coupled Device (CCD) Camera The detector of our prototype is an image sensor in the form of a Charge Coupled Device ( CCD ) Camera. The camera has a resolution of 512X512 pixels with a pixel size of 18 mX18 m and provides a maximum frame rate of 300 frames per second The camera is shown in Figure 3 5 S everal parameters of the camera alter the overall output of the system which includes t he exposure time, g ain setting and o ffset. The influences of such func tions on our sys tem will be explain ed in C hapter 4 Figure 3 5 Picture of our Charge Coupled Device c amera (Sarnoff model CAM UV512) 3.6 Experimental Setup: Explosive s Detector Baggage Scanner The light from a 300 W Xenon source is directed via fiber optic s and focuses on the conveyor belt producing a line of 5 mm width by 50 cm length ( which is a bout the width of airport carry on luggage ) depicted as scanning beam in Figure 3 6 T he

PAGE 26

26 reflected light is then collected and analyzed with the spectro graph and computer system. Figure 3 6 Schematic s of the e xplosive s detector. L ight is pro jected on the conveyor belt. The reflected light is collected and analyzed with a spectrograph and computer. Carbon pad s are used as standard substrate as shown in Figure 3. 7 a C arbon pad s absorb most of the light both in the UV and the visible region of the system and minimize the noise from the background The line of light shun on a mock suitcase that is placed on the conveyor belt is shown in Figure 3 7 b.

PAGE 27

27 Figure 3 7 Example of a s ample and its placement on the conveyor belt a ) Crystalline TNT is placed on c ar b on p ad using a g lass substrate b ) The line of light on a mock suitcase T he operating system ( computer) processes the data and calculat es the d ifferential reflectivity ( explained in the next section) I mage processing techniques are performed to clas sify the spectrum in order to i dentify the material being i n vestigated 3. 7 Mathematical Analysis using Image Acquisition The methodology involving d ata processing is explained in this section T he TNT is illuminated by a light source. At th at instance, a n image frame is acquired The intensity of each pixel is recorded and stored. Let this frame be called R 1 As the conveyor belt moves, a second frame R 2 is captured. For each line of pixels, the differential reflectivity is calculated [15] as follows: W here the numerator (R 1 R 2 ) is the difference of reflectivities of the two sample parts while the denominator (R 1 +R 2 )/2 represents the average of the reflectivity between the two sample parts, thus yielding a normalized difference in reflectivity. Such

PAGE 28

28 normalization decreases the influence from intensity variation of spectral output of the light source, the spectral sen sitivity variation of the detector and the spectral reflectivities of the mirrors or the substrate [ 15 ]. All of these steps are implemented by the us e of a graphical user interface, an interface tool developed to automat e the entire process from image acqu isition to the data processing [See Appendix A for details] Figure 3 8 Schematic of light (blue) collection from sample on conveyor belt. R 1 and R 2 are two frames, each frame containing 512 pixels along the width of the conveyor belt. As shown in Figure 3.8, t he d ifferential reflectivity ( ) calculation is perfo r med 512 times along the width of the conveyor belt, t hus rendering a resolution of 1 m m X 5 m m on the surface to be analyzed. The d ifferential r eflectogram s ( versus wavelength) t hus obtained can be used to identify energetic material

PAGE 29

29 3. 8 Differential Reflectogram of T rinitrot o l uene The differential reflectogram of TNT, shown in Figure 3 9, is obtained by using the method discussed in the previous section It is observed from Figure 3 9 that there is a characteristic shoulder from 380 nm 430 nm which is unique for TNT. This feature is Figure 3 10 which has a unique shoulder like that of TN T but is right shifted by 65 nm to 485 nm. Figure 3 9 Differential r eflectogram of TNT obtained by differential reflectometry The characteristic shoulder around 420 nm is observed. -0.9 -0.8 -0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 270 290 310 330 350 370 390 410 430 450 470 DeltaR/Rbar(a.u.) Wavelength(nm) 420 nm

PAGE 30

30 Figure 3 10 Differential reflectogram of TATB obtained by differential reflectometry The characteristic shoulder is shown at 485 nm 3.9 Summary In this chapter, the different components (light source, CCD camera and spectrometer) that make up our instrument were studied. The working methodology of the explosive detector using the differential reflectometry system was reviewed. It has been observed that TNT and TATB provide distinctly different absorption peaks. -0.3 -0.2 -0.1 0 0.1 0.2 0.3 400 420 440 460 480 500 DeltaR/Rbar(a.u.) Wavelength (nm) 485 nm

PAGE 31

31 CHAPTER 4 CHARACTERIZATION OF SYSTEM PARAMETER S In this chapter, the parameters of the explosives detector ( e xposure time, gain setting, offset setting and distance between the collection optics and sample) are varied and their effect on Signal to Noise ratio is presented in this chapter. 4.1 Signal T o Noise Ratio : Metric for Comparative Study One common metric associated with noise is the Signal to Noise Ratio (SNR) SNR is defined as the ratio of signal power to the noise power corrupting the signal. The h igher the ratio, the greater is the signal st rength If P signal and P Noise correspond to the average power of signal and background noise respectively, then SNR can be written as Since powers are measured across the same impedance (i.e. same sensor) it can be expre ssed in squares of their amplitude as follows, Generally, it is represented in the logarithmic decibel (dB) scale, hence In image processing the amplitude of the signa l is the mean while the noise is characterized by the standard deviation. Thus the SNR is t he ratio of mean to standard deviation of a signal measurement i.e [18 ]

PAGE 32

32 w here is the mean signal and is the standard deviation. The above relation is used in the calculation of SNR throughout this thesis. The differential reflectogram of TNT is characterized by its shoulder at 420 nm. The signal is the mean of this shoulder height. The noise is characterized as the stand ard deviation of the curve. The SNR is computed according to the above imaging equation. The higher the dip in magnitude at the 420 nm region, the greater and better is the quality of the result. This definition is only valid if the associated variables ar e positive, such as photon counts or luminance. 4.2 Camera Exposure Time In our camera, t he amount of time over which the acquisition (or the exposure) takes place can be varied between 5 ms and 99 ms The purpose of exposure adjustment is to control the amount of light coming from the subject that is allowed. A long exposure time signifies more signal increases contrast (if noise does not increase as fast as signal) and adds more sharpness This is shown in F igure 4 1. a) b) Figure 4 1 Frame acquired at ( a ) 99 ms exposure time and ( b ) 50 ms exposure time A l onger exposure time is characterized by a brighter image while a smaller exposure time is characterized by a dark er image

PAGE 33

33 Figure 4 2 depicts various differential reflectograms ( versus wavelength) as a function of exposure time. It is observed that the should er height between 20 ms and 40 ms (curves represented by top most red, green and blue) though present, is lower in magnitude when compared to the shoulder height between 50 ms and 80 ms (curves represented by yellow, pink, cyan and black) Figure 4 2. D ifferential r eflectogram s of TNT as a function of e xposure time as listed o n the curves at g ain set ting of 7 00 a.u. offset = 1 and distance between sample and collection optics = 7 cm from the detector. The curves have been shifted for clarity In Figure 4 3, the SNR is plotted versus the exposure time, using the data contained in Figure 4 2. It can be seen SNR increases considerably up to about 7 5 ms and then saturate s I n other words, l ong exposure time s lead to saturation of pixel s, thereby result ing in overshadowing the actual reflectivity signal from the TNT material Thus long exposure s time setting s (80 ms and 99 ms) are not recommended For th is reason the exposur e time s between 60 ms and 7 5 ms would probably give optimum result s

PAGE 34

34 Figure 4 3 S ignal to Noise Ratio as a function of exposure t ime The data is plotted with blue diamond. The solid blue line is used as a visual aid to model the behaviour of SNR. The point in the circle is not considered in the model (solid line) 4. 3 Camera Gain and Offset Variation Effect The concept of Gain is explained as follows The signal, coming off the CCD Camera i s typically pre amplified which allows the signal to be boosted by a variable amount called the gain We can also shift the signal by another variable amou nt called offset Let s i llustrate this with a n example. For pixel counts correspond ing to 0.1, 0.3, 1.2 and 1.3, the A nalog to D igital C onverter (ADC) (which cannot handle fractional numbers ) would output 0,0,1,1. Thus a quantization error occurs On the other hand if 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 0 20 40 60 80 100 120 SNR(dB) Exposure Time (ms) 75 ms

PAGE 35

35 we scaled up the signal (introduce some gain) for example by 10, these numbers would become 1, 3, 12 and 13, t he quantization errors therefore disappear s Also i n practice, the noise can make the signal negative. Since the ADC cannot take in a negative input, an offset is added to the signal to make sure th at all signals coming off the CCD camera are positive. When the CCD camera captures an image, the output of the CCD camera pixel is proportional to the amount of light falling onto it. Saturation occurs when the recording device just cannot hold more light, then the signal is recorded as the maximum v alue of t he camera (in our case: 4095 ). This is called clipping, and no details are recorded. The gain setting s can be varied from 100 to 1023 units. A relation ship between the wavelength and intensity is shown in Figure 4 4 indicating the clipping phenomenon at ga in setting s of 700 between 460 nm and 500 nm. For gain settings beyond 800, t here is extensive clipping Figure 4 5 shows a relationship between gain setting s and the intensity of a pixel measured at 410 nm (green triangle), 420 nm (red squares) 480 nm (purple diamonds) and 520 nm ( blue cross). As seen in Figure 4 5, for a gain setting of 800, t he intensity quickly reaches the saturation level for the 520 nm pixel wh ereas the intensity is only at 1555 for the 410 nm pixel As a result, reflected data from the 520 nm pixel migh t not be accurate. We observe the best gain setting close to 700 since there are no instances of clipping. Hence, a gain setting of 800 and above should not be employed.

PAGE 36

36 Figure 4 4 Intensity as a function of the wavelength for various gain settings from 100 to 1023.Intensity is the reflected data before the process of differencing. At high gain, cliiping may occur, as marked on the graph. 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 200 250 300 350 400 450 500 550 600 Intensity(counts) Wavelength(nm) 1023 700 600 500 400 300 200 100 900 800 Clipping

PAGE 37

37 Figure 4 5.Inte nsity as a function of the gain s etting s of a poi nt pixel at 41 0 nm (green), 420 n m(red) 480 nm (purple) and 520 nm (blue). At high gain, cl i p ping may occur, as marked on the graph. The differential re flectogram s at various g ain settings are shown in Figure 4 6 and a relationship between the SNR and the gain settings is shown in Figure 4 7. The signal (420 nm feature height )and noise are extracted from the differential reflectograms of TNT (Figure 4 6) for each gain settings. From this, we calculate the SNR and plot it as a function of the gain (Figu re 4 7). From the model (solid red line) it can be seen that the SNR linearly increases until a gain setting of around 750 ; a bove this the SNR decreases. Having observed the clipping effects in Figure 4 5 and seen the SNR 0 500 1000 1500 2000 2500 3000 3500 4000 4500 0 100 200 300 400 500 600 700 800 900 1000 Intensity Gain Setting Allowed range of gain settings 480nm 520nm 420nm 410nm Best Gain Setting Clipping

PAGE 38

38 improvement in Figure 4 7, the op timum gain setting can be chosen to be between 6 5 0 and 75 0. Figure 4 6 D ifferential r eflectogram of TNT with increasing g ain s ettings at e xpos u re t ime= 7 0 ms, o ffset =1 and distance between the collection optics and sample = 7 cm from the detector. The curves have been shifted for clarity.

PAGE 39

39 Figure 4 7 S ignal to Noise Ratio as a function of the g ain s etting s (blue cross) The red solid line is used as a visual aid to indicate the behaviour of th e SNR In addition to the gain, an offset can be added to the signal. The offset settings can be varied from 1 to 255. Differential reflectograms of TNT as a function of the offset are plotted in Figure 4 8. Figure 4 9 displays the SNR extracted from Figure 4 8 (Signal = 420 nm feature height, noise = st andard deviation) as a function of the offset. 0 2 4 6 8 10 12 14 16 0 200 400 600 800 1000 1200 SNR(dB) Gain settings Best Gain settings range 650 750

PAGE 40

40 Figure 4 8 D ifferential r eflectogram s of TNT with d ifferent o ffset settings as listed in legend at g ain s etting = 700, exposure t ime = 7 0 ms and distance between the collection optics and sample = 7 cm from the d etector. The curves have been shifted for clarity. Figure 4 9 S ignal to Noise Ratio as a function of the o ffset settings The red solid line is used as a visual aid to model the behaviour of the SNR. The SNR is maintained constant, indicati ng that the offset has no effect. We observe that the SNR is not affec ted due to a change in the offset as it remains almost constant ( Figure 4 9) T he 420 nm feature also seem to have equal 0 1 2 3 4 5 6 7 8 0 50 100 150 200 250 300 SNR (dB) offset (a.u.)

PAGE 41

41 prominence at all values of offset (Figure 4 8) H owever, it is best to set the offset settings to the smallest value which prevent any nega tives signal to go into the ADC, as it offer s a large r dynamic range. 4. 4 Effects of the Distance between Collection Optics and Sample Since a standoff optical m ethod is emplo yed, the effect of distance at which the material is placed from the collection optic s plays a very important role in determining the signal intensity. The shoulder height around 420 nm feature, from Figure 4 10, is found to be larger in magnitude at short distance s (between 5.1 cm and 11.3 cm) It should be noted that the noise is essentially constant th roughout and only the signal decrease s as seen in Figure 4 11 The camera parameters ( gain, offset and exposure time) are kept constant in this section Thi s confirms that ex ternal noise effect (stray light) is negligible. From the results shown in Figure 4 12 a decreasing linear trend (solid blue line) is fitted to the data (blue diamonds). The SNR decreases from 8 dB at 5 cm to about 2 dB at 30 cm.

PAGE 42

42 Figure 4 10 D ifferential r eflectogram of TNT as a function of d istance between the collection optic and the sample at g ain setting = 7 00, e xposure t ime=7 0 ms and o ffset =1. The curves are shifted for clarity. Figure 4 11 Signal (Blue diamonds ) and N oise (Red squares ) as a function of the d istance betwe en collection optics and sample The blue and red solid lines are used as visual aids to indicate the behaviours of the signal and noise respectively. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 10 20 30 40 Signal and Noise(a.u) Distance(cm) Signal Noise

PAGE 43

43 ` Figure 4 12 S ignal to Noise Ratio as a function of the d istance betwe en collection optics and sample (blue diamonds) The blue solid line is used as a visual aid to indicate the behaviour of the SNR. 4. 5 Summary The parameters used in this study (exposure time, gain settings and offset settings, distance b etween the collection optics and the sample) were studied and compared with a newly intro duced concept in this chapter namely the Signal to Noise Ratio. The exposure time is observed to work be st between 6 5 ms and 7 5 ms. In low light conditions the gain has to be increased. For optimum p erformance, gain setting s between 6 5 0 and 750 units is recommended Though there is not much effect of the offset on the SNR, it is best to keep it at the lowest value possible to get the largest dynamic range Finally, t he shorter the distance between the sample and collection optics the stronger is the signal. 0 2 4 6 8 10 12 0.00 5.00 10.00 15.00 20.00 25.00 30.00 35.00 SNR(dB) Distance(cm)

PAGE 44

44 CHAPTER 5 EXPLOSIVE S DETECTOR NOISE CHARACTERIZATI ON OF COMPONENTS In our explosive s detector prototype, the majority of the noise contribution comes from the camera and the light source. In this chapter, we characterize the noise from these components. 5.1 Charge Coupled Device Camera Noise Characterization The main goal using a Charge Coupled Device ( CCD ) camera is to produce the best image possible. Achieving low noise images from CCD camera requires basic understanding of the working of CCD and the different sources of noise that can reduce the quality of image s CCD camera s operate by converting photons into electrons which they store in each pixel. The number of electrons stored in each pixel well is proportional to t he number of photons that strike that pixel. After an exposure has been completed, the electrons for each pixel are shifted out of the CCD c amera and convert ed to a number, indicating how dark or light each particular pixel is 5.2 Sources of Noise from Charge Coupled Device Camera In the ideal situation, each photon that strikes a pixel would be converted into exactly one electron but there are always some unwanted light variations in pixel (noise). in dark background image and the faint horizontal or vertical lines that become visible in low signal areas. The metrics used in CCD camera should also be considered. Often, CCD manufacturers measure noise as a number of elec trons RMS (Root Mean Square) or counts

PAGE 45

45 5.2 .1 Readout Noise Readout noise is the noise produced when signals are converted from analog to digital data in an A/D converter present in the CCD Camera In the output of the C CD image, readout noise is created in each pix el every time the array is read out Thu s this noise is present in all the images, regardless of the exposure time gain setting s or offset and more importantly it has constant amplitude 5.2 .2 Dark Noise Dark noise is the relatively small electric current that flows through the CCD even when no photons are entering the device. Dark noise is present in long exposure time i mages taken with the CCD camera T his noise shows up as bright pixels in the image where Dark noise is one of the noises which can be decreased by reducing the exposure time. 5.2 .3 Noise Equation The noise sources discussed above could be put together into a n equation to calculate the overall uncertainty or standard deviation In terms of standard deviation, [ 19 ]

PAGE 46

46 5.3 C harge C oupled D evice Camera Noise Characte rization : Experimental Procedure 5.3 .1 Readout Noise Measurement In imaging applications, a bias frame is an image obtained from a CCD Camera with no actual exposure time or the smallest a llowed value pe r the instrument specifications ( in our case 5 ms ) with the shutter closed The bias image so obtained contains the readout noise which has been explained in Section 5.2.1 A total of 32 bias frames in immediate succession were acquired in the dark at 5ms exposure time. The 32 frames were considered pair w ise for the purpose of calculation The difference between the two images w as taken and the standard deviation of the differenced image calculated The standard deviation per image is obse rved to be at 1.6 8 counts. 5.3 .2 Dark Noise Measurement A d ark frame is an image captured with the CCD camera in the dark at a given exposure time. W ith the shutter closed, a series of 8 dark frames were acquired at different exposure times ( 5 ms to 95 ms ) Using a Labview algorithm (Courtesy of Dr. Thierry Dubroc a) t wo dark frame image s were considered in a time interval and its difference computed. By considering two frames at a time, the spatial noise s (external noise from space) are not considered in the calculation; since forming the difference will essential ly remove the spatial noise. The s tandard deviation is calculated from such differential images which allow to correlate the dark noise with the exposure time. 5.3 .3 Experimental Results Figure 5 1 depicts the noise of the camera as a function of the exposure time. T he read out noise was found to be 1.6 8 counts (extra polated at zero exposure time). The equation explained in section 5.2.3 is used to model the noise as follows:

PAGE 47

47 Where For example, at 99 ms exposure time, the noise is 1.83 counts, i.e., a readout noise of 1.68 counts and a dark noise of 0.72 counts. In effect, the dark noise is less than half of the readout noise. Figure 5 1 Characterization of n oise from c amera Read out noise is measured to be 1.68 counts (extrapolated to zero exposure) The dark measurement data is represented in red squares The red solid line is used as a visual aid to model the behaviour of the noise. It is observed that at low exposure time s the readout noise is more dominant over the dark noise The re is a gradual increase in the dark noise as function of the exposure 1.65 1.7 1.75 1.8 1.85 1.9 0 10 20 30 40 50 60 70 80 90 100 Noise (cts) Exposure Time(ms) noise model Readout Noise = 1.68 Counts 1.83 (for 99ms exposure)

PAGE 48

48 time. The plot in Figure 5 1 is an average of several measurements ( 8 frames ) and the read out noise was observ ed to be cons istent for all of them Similarly, the dark noise was measured from an exposure time of 5 ms to 9 9 ms. T he model (red line in Figure 5 1) is well correlated with the collected data ( coefficient of correlation = 0.9 8 ) Finally this is consistent with the manufacturer maximum noise specifications (3 counts). 5.4 Light Source Noise Characterization The choice of a proper light source is critical as it has a major influence on the output of our explosive s detector. Hence, the choice of the light source is critical. In this section, several light sources and their associated noises are characteriz ed and compared. 5.4 .1 Experimental Procedure for Noise Characterization of Light Sources D ifferent light sources were placed at different distance s from the detector so that the detector output was about 5 V The light was directed towards a bright surface and the reflected data recorded The acquisition time was 10 s econds for this experiment at a sampling rate of 10 kHz. The det ector voltage intensity is o btained with respect to the acquired time in milliseconds. Also, a Fast Fourier Transform (FFT) computation was performed on the raw light intensity data to understand the behavior of th e light source in the frequency domain The noise present for all the light sources is measured in the form of standard deviation of the signal intensity Experimental results for each of the light source are shown in the subsequent sections. Note : These experiments were measured in the dark.

PAGE 49

49 5.4 .2 Lamp Off Characterization Control Experiment This experiment was carried with out any light source to determine the actual noise of the measurement system itself Even in the absence of light, the detector and Data Acquisition (DAQ ) card produce noi se and is measured to be 0. 8 mV (standard deviation of the detector voltage) which will be used a s a base for c omparing the different light source s Figure 5 2 shows the detector voltage as a function of time when no lamp is used while Figure 5 3 displays a FFT computation on the detector voltage as a function of time for the control experiment. Figure 5 2. Detector v oltage as a function of time when no lamp is u sed ( Control e xperiment) Noise = 0.8 mV

PAGE 50

50 Figure 5 3 F ast F ourier T ransform amplitude of detector voltage as a function of t im e for the control experiment The presence of harmonics (first, second and fourth) fr om the power outlet and white noise having an amplitude 0. 15x10 4 is observed. There are 4 harmonics produced, wherein, the fundamental harmonics at 60Hz is m ost pronounced. Apart from these spike like noises we observe that the magnitude of the background noise is constant characteristic of white noise White noise is a random signal with a flat power spectral density In th e control experiment, the source s of noise are the white noise and harmonic spike noise The white noise has amplitude of about 0. 15x10 4 This is th e noise floor for our measurements in this section 5.4 .3 Quartz Tungsten Halogen Quartz Tungsten Halogen ( QTH ) lamps are made up of doped tungsten filament inside a quartz envelope. They are often filled with a rare gas and a small amount of halogen such as iodine or bromine. QTH is strong in the high visible near infra red region [20 21] but does not produc e enough light in the UV region. The QTH consumes a power of 50 W. The standard deviation of the detector voltage is the measured noise.

PAGE 51

51 Figure 5 4 Detector voltage as a function of t ime for the QTH Lamp at 50W Noise = 8.9 mV Figure 5 4 shows the variation of the detector voltage with respect to the time for the QTH. The noise of the QTH lamp is measured to be 8.9 mV (standard deviation of the detector voltage) This is 10 times larg er than the noise from our control experiment This confirms the validity of our experiment (i.e. the detector and the DA Q card noise are essentially negligible).

PAGE 52

52 Figure 5 5 shows the Fast Fourier Transform amplitude detector voltage as a function of time for the QTH light source Figure 5 5 Fast Fourier Transform amplitude of detector voltage as a function of time for our QTH Lamp The presence of first, second and fourth harmonics is indicated in red. The presence of first (60 Hz), second (120 Hz) and fourth (240 Hz) harmonics is due to the spikes from the power supply from the wall outlet. From Figure 5 5, the white noise is measured to be around 0.2x10 3 in amplitude. M ost of the white noise is evenly distributed through the frequency range of 0 300 Hz.Th ere is small noise around 5 Hz addin g to the white noise. 5.4 .4 75W Xenon Lamp This experiment is to characterize the noise arising from a 75 W Xenon lamp. The noise is observed to be 66 mV (standard deviation of the detector voltage) This is 7 times noisier than the QTH lamp.

PAGE 53

53 Figure 5 6 Detector voltage as a function of t ime for 75 W Xenon light Noise =65.9 mV. Figure 5 7 Fast Fourier Transform amplitude of detector voltage as a function of time for a 75 W Xenon light source The initial high amplitude irregularity comes from noise in the Xenon light The white noise is now negligible below 25 Hz The amplitude of noise is around 0.2 units below 25 Hz

PAGE 54

54 From Figure 5 7, we observe a periodic variation of the light source. As seen in QTH, the fundamental harmonics are very strong at multiple s of 60 Hz from power supply from the wall outlet and the DAQ card The important aspect to note is the initial increase in the amplitude of a bout 0.02 near 12 Hz This is attributed to the Xenon light bulb T he white noise is negligible below 25 H z. The magnitude of noise from this light source is about 10 times larger in amplitude when compared to the QTH, while it is about 100 times larger in amplitude compared to the control experiment. When this light source is used, ad ditional filters in conju nction with white noise filters and spike noise filters must be employed 5.4 .5 White L ight E mitting D iode A white Light emitting diode (LED) is a semiconductor device which converts electricity into light. This is battery powered and hence runs on a DC source As seen in Figure 5 8 the LED has noise characteristics similar to the QTH light source. The noise amplitude is 10.7 mV (standard deviation of the detector voltage) slightly higher than that of the QTH lamp ( 8.9 mV).

PAGE 55

55 Figure 5 8 Detector voltage as a function of t ime for a w hite LED Noise = 10.7 mV. Figure 5 9 Fast Fourier Transform amplitude of detector voltage as a function of time for w hite LED More pronounced harmonic spikes at multiples of 60 Hz. The noise is also homogenously distributed From the FFT plot in Figure 5 9 we do observe the spike like harmonic peaks at multiples of 60 Hz They are coming from the power supply from the wall outlet through

PAGE 56

56 DA Q card and the detector, as explained in the control experiment (Section 5.4.2). Even though LED behavior is similar to the QTH lamp the important point to be noted is that noise amplitu de at around 5 Hz is about twice that of the QTH lamp The noise ampli tude is about 3x10 4 which is 20 times noisier than the control experiment (0.15x10 4 ) It i s homogenously distributed throughout the spectrum a bove 10 Hz The background is mostly white noise. 5.4 .6 Xenon Arc Lamp ( 300W ) The Xenon light source intensity was measured at two different time intervals. The first measurement is recorded immediately after the source is turned on, while t h e second set is measured 30 min ute s afte r the warm up period Figure 5 1 0 Detector voltage as a f unction of t ime for Xenon 300 W light source immediately after turn on Noise =3 3.9 mV From Figure 5 1 0 we find that the noise (=34 mV) produced by the Xenon lamp at 300 W is lower in magnitude by one half, when compared to the 75 W Xenon l amp (=66 mV) I t is about 30 times larger in magnitude when compared to the control experiment

PAGE 57

57 and about 4 times larger when compared to the non UV light generating source s (QTH and LED). Figure 5 1 1 Fast Fourier Transform amplitude of detector voltage as a function of time for Xenon 300 W immediately after turn on Below 70 Hz, the noise is very large and not white. From the FFT spectrum shown in Figure 5 1 1 we find strong noise from the light source below 70 Hz. It is also non uniform in that region. Similar to the LED and QTH light sources, we see a sudden and a very high amplitude increase between 5 10 Hz. The noise due to the light source is present up to 7 0 Hz, above which the noise is white

PAGE 58

58 F igure 5 1 2 Detector voltage as a function of t ime for Xenon 300 W light source 30mins after turn on Noise is 22 mV Waiting 30 minutes for the Xenon lamp to stabilize makes a difference towards noise reduction which is evident from Figure 5 12. I t is observed that the noise is reduced by a factor of 1.5(22 mV after 30 minutes against 34 mV when taken immediately after turn on ). Now, this reduction brings the noise a little closer to the baseline experiment or the QTH sources when compared to the other types of Xenon sources. The other non UV emitting light sources are only about half the noise magnitude compared to this Xenon lamp after it stabilized.

PAGE 59

59 Figure 5 1 3 Fast Fourier Transform amplitude of detector voltage as a function of time for Xenon 300 W 30 mins after warm up period Inspecting Figure 5 13, the FFT amplitude of the detector voltage tells us that t he initial increase around 5 10 Hz is reduced from 8x10 3 to 5x10 3 After 10 Hz, we also observe that the maximum noise is reduced i n magnitude by half compared to the Xe non source with no waiting time (from 4x10 3 to 2x10 3 ) T he noise from th is stabilized source starts to fade out at 50 Hz (as compared to 70 Hz to the previous section), after which there is only white noise and harmon ic spike noise Reduction of noise amplitude at lower frequencies means that the starting frequency for operation of our explosive detector can be lowered to 50 Hz itself (instead of 70 Hz if we do not wait for the stabilization period ). 5. 5 Summary In thi s chapter, the camera and the light sources noise have been studied. We were able to measure the noise from the camera to be 1. 68 counts. While the readout

PAGE 60

60 noise is present and its amplitude constant in all measurements, the dark noise increases with expos ure time but remains negligible even at the highest exposure time A summary of the noise present in various light sources is plotted in F igure 5 1 4. Figure 5 1 4 Characterization of noise from light source s 75 W X enon produces the largest noise The noise (0.8 mV) from the control experiment is inherent in all sources and is essentially negligible QTH lamps and LED have similar characteristics. It is best to allow the Xenon lamp to stabilize for 30 mins and then perform data acquisition Of all the s ources investigated above the QTH has the lowest noise. Our main region of interest is in the Ultraviolet region but the QTH does not produce much UV. LED has a slightly higher noise but has the same drawback as the QTH. It does not produce light in the UV. The Xenon lamp ha ving higher power produces less noise than the lowered power ed source. It is best to stabilize the Xenon lamp as it is found to reduce the noise 0 10 20 30 40 50 60 70 80 No Lamp(Control Experiment) QTH 50W lamp 75W Xe lamp White LED lamp Xe 300W lamp right after turn on Xe 300W lamp after 30 min Noise (mV) Light Source

PAGE 61

61 by 30% The lowest possible o perating frequency a lso decreases from 70 Hz t o 50 Hz when stable. Now that we have characterized some possible light sources and selected the best candidate (300 W Xenon light source), we can investigate more easily our de noising strategies. De noising strategies are discussed in the following chapt er.

PAGE 62

62 CHAPTER 6 DENOISING STRATEGIES Our goal is to achieve a low limit of detection for explosives and a high sensitivity, i.e., a very good Signal to Noise ratio T he need for proper noise suppressing strategies needs to be incorporated 6 1.Method to Reduce Noise during Frame Acqu i sition Although the source of noise is well known, it is difficult to change the internal metrics of the hardware due to manufacturing restrictions. As shown in Figures 5 12 and 5 13, the Xenon light sources has an inherent high noise factor in the low frequency range of 0 50 Hz. The current rate at which the two image frames, R 1 and R 2 are taken is (as shown in Figure 6 1) 5 seconds. Due to the processing time, in particular data storage, the operating frequ ency is around 0.1 Hz (exposure time is negligible when compared to the processing time), which is in the range where the noise is the strongest. Figure 6 1 Schematic of the image acquisition and storage modes (A) External da ta storage takes considera ble time (5 sec onds ) compared to which the exposure time (5 ms) becomes negligible. (B) Acquisition with internal storage time results in working at low exposure time (5 ms) in order to be outside the high noise region.

PAGE 63

63 In order to be outside this high noi se region (0 50 Hz), a target between 50 100 Hz should be achieved (excluding the frequency at 60 Hz). For this we would have to work at a very low exp osure time between 10 ms to 20 ms. If the data is stored in an external location (Excel Spreadsheet) th e time consumed would be more than when the data is stored internally within the same application. From S ection 4.2 of C hapter 4, we have seen that operating at very low exposure would result in a low signal strength (Figure 4 3). A methodology is proposed and illustrated in Figure 6 2 to improve the SNR At t=0, the light is shun at the sample and a pair of consecutive frames are acquired from the same sample. On averaging over multiple frames, the noise is reduced [3 0 3 1 ]. This ca n be achieved by trigger control of the conveyor belt by means of the compu ter acquisition program. At t=10 ms, the conveyor belt is moved and subsequently anoth er pair of frames are acquired.

PAGE 64

64 Figure 6 2 Timing of image acquisition on the conveyor belt The yellow spot represents the light source shining on the sample. The movement of the conveyor belt is in the forward direction as shown. Frames R 1 1 are acquired after which the belt moves acquiring frames R 2 2

PAGE 65

65 Figure 6 3 Proposed methodo logy to reduce the effect of noise and increase SNR. 2 frames (instead of the conventional 1 frame) of the same sample are acquir ed simultaneously and differential reflectogram computed This is schematically shown in Figure 6 3 The two frames (R 1 1 ) acquired at 5ms are averaged to produce A. Similarly, the next pair of frames (R 2 2 ) is acquired and an averaged B is obtained. A differential reflectogram is performed on A and B and their ef fects on SNR studied. A comparative measurement with this m ethodology was conducted on TNT and is presented in Figure 6 4

PAGE 66

66 Figure 6 4 Differential reflectogram s of TNT when one frame is acquired (blue) having a SNR of 19.7 dB while acqui sition of two frames (red) has a SNR of 23.7 d B The blue curve indicates t he conventional differential reflectogram measurement without any averaging and the exposure was set at 8 0 ms. The measured SNR was 19.7 dB, while the red curve indicates the frames measurement being taken pair wise in quick succession at 2 0 ms each. The S NR was found to be 23.6 dB and is increased. The data w ere stored externally and hence the operating frequency was 0.1 Hz. I f we were not limited by the frequency (working frequency of more than 50 Hz) and stored the data internally we would take more frames and average them. Then we could potentially observe more improvement in the SNR. However, this process is more time con suming and additional software is required for rapid data processing as the number of acquired frames increas es and requires trigger co ntrolling of the conveyor belt. 6 2 Post Processing Filtering M ethods Post Processing fi ltering is defined as the process of improving the perceived quali ty of the acquired fram e. I maging filters (median, average, disk, Gaussian and

PAGE 67

67 hybrid filters) have been applied to our explosive s detector in order to reduce the noise. A comparative f ilter study is presented in this section 6. 2 .1 Median Filter A window is defined as a group of neighboring pixels ( i.e., a set of elements ) arou nd a given one. It is often called a it hides those elements which are not being considered). The median filter [22 23 ] is a windowed, nonlinear digital filtering technique, used to remove destructive noise while preserving the amplitude of the signal. Its performance is particularly good for removing s pike noise. Spike noise are dominant in the 300W X enon light sources (Figure 5 13, C hapter 5). 6. 2 .1.1 Understanding Median Filter In one dim ension, the median filter is r epresented in Figure 6 5 Each vertical bar represents the pixel intensity. Figure 6 5 Schematic working of the median filtering. The filtered output is the median value of the sorted pixel intensities in ascending order. L onsider five pixels in some random order of intensities. The m edian filter sort s through all the pixel intensities in ascending order. Then, the median of the window is taken based on the size of the filter The princip al idea of this filter is to r un through the

PAGE 68

68 entire frame replacing each entry with the median of the ne ighboring entries. In our example ( Figure 6 5 ) th e filtered output is the pixel number 5 as it corresponds to the median of ascending pixel intensities At the edges of the frame, the first and the last values are repeated, in order to be consistent with the filter size. 6. 2 .1.2 Experimental Results Different filter sizes were applied to the raw differential reflectogram of TNT (no filtering represented by the red curve). From Figu re 6 6, i t is noted that with increasing filter size, more smoothing takes place and the noise of the differential reflectograms is reduced. Figure 6 6 Differential reflectograms of TNT with increasing sizes of median filter. The r ed curve is the unfilt ered data while the green curve (filter size 2), blue curve (filter size 4) and black curve (filter size 8) show the filtered data. The amplitude difference of TNT features for red and black curves are shown for comparison. The signal shoulder around 420 nm has fewer perturbations in it which is also very important for achieving the limit of detection T he amplitude of the feature is not reduced

PAGE 69

69 From this data, a SNR calculation was al so performed on the various filter sizes and shown in Figure 6 7 Init ially there is a rapid increase in the SNR from 18 dB to 23 dB for a filter size of 4. From then on, the rate at which the SNR increases reduces. From looking at the graph, the SNR approaches saturation at filter size of 12. T his occurs at 2 8 dB. Thus, t he maximum achievable SNR improvement is 9 dB from 19 dB (unfiltered data) to 2 8 dB (filter size of 12). Figure 6 7 SNR as a function of the median filter s ize for differential reflectograms of TNT. The SNR (= 2 7 dB ) saturates at about 1 0 having a SNR improvement of 8 dB 6. 2 .2 Average Filter and Disk Filter An often used terminology used i n imaging filters is the kernel concept A kernel is a matrix which is applied to every pixel in the image. The kernel contains multiplication 0 5 10 15 20 25 30 35 0 2 4 6 8 10 12 14 16 18 SNR(dB) Filter Size Saturates at Filter size 10.SNR=27 dB 8dB

PAGE 70

70 factors to be applied to the pixels and its neighbors. Once all the values have been multiplied, the pixel is replaced with the multiplied new number. By choosing different kernels, different types of filtering can be applied. The a verage filter [24 ] is a windowed filter of linear class, tha t smoothes the signal and image Average filter is characterized by its kernel size The fi lter works as a low pass filter When the size of the kernel increases the smoothing effect increases. 6. 2 .2.1 Understanding Average Filter and Disk Filter T he basic schematic of the average filter is shown in Figure 6 8 W e consider that there are 5 pixels in an array (1D representation) and the pixel intensities are all different. T he average of the pixel is taken when the kernel is applied to the pixel. Figure 6 8 Schematic principle of the average filter (filter size = 5) The filtered output is the average value of the pixel intensities. For example, in an image, a 3x3 average square kernel filter is applied to the pixels as shown below:

PAGE 71

71 The kernel can be made slightly different For example, by arranging the kernel in a circular fashion it will result in disk orientation compared to a square orientation in th e case of the average kernel Orientation is referred to the kernel matrix is arranged. This orientation is called disk filter [25] It i s represented in Figure 6 9 Figure 6 9 Representation of the disk filter with disk size = 3 Brown rectangles are inc luded in computation while shaded rectangles correspond to pixels that are ignored in filtering process. The kernel matrix for the disk filter is as shown below: 6. 2 .2.2 Experimental Results The differential reflectograms of TNT with the a pplication of disk filter is shown in Figure 6 10 The differential reflectograms become more smooth ed when we apply the disk filter with large filter size s As seen from the figure, the presence of distortion i s clear as the amplitude of feature s (TNT fea ture and Peak A) get reduced by different amount s depending on their frequencies.

PAGE 72

72 Figure 6 10. Differential reflectograms of TNT with increasing size s of disk filter s The r ed curve is the unfiltered data while the green curve (filter size 4), blue curve (filter size 6) and black curve (filter size 8) show filtered data. The filtered curves have been multiplied by 5 for clarity. The TNT feature ( filter size = 8 ) is reduced by 15 times (ratio of TNT feature of the red curve to the TNT feature of the black curve multiplied by 5) in amplitude compared to unfiltered data. Also, let us consider the peak A at 484 nm. We can observe that peak A (higher frequency than the TNT feature) is diminished by 70 times in amplitude when we apply the disk filter with a filt er size of 8. Hence, the amplitude features decrease non uniformly as a function of their frequency and the filter size. The schematic drawing ( Figure 6 1 1 ) help to visualize the distortion of spectral features as a function of filter size. In general, the disk filter is used to reduce frequency noise higher than the feature carrying the information (i.e. signal) Hence the cut off frequency is a critical parameter in the filter size choice.

PAGE 73

73 Figure 6 1 1 Schematic of feature amplitudes as a function of th e frequency for the disk filter of various sizes. f 4 f 6 and f 8 are the cut off frequencies for filter size 4, 6 and 8 respectively. From Figure 6 11 we are able to see that the feature amplitude decreases with increa sing filter size, in addition to the de creasing cut off frequency of the filter. F eatures with much higher frequencies than the cut off frequency will be lost. Peak A decreases much faster in amplitude than the TNT feature because it has a higher frequency. Let us take the example of TN T and TATB, which were presented in an earlier chapter The shoulder of TATB is sharper (higher frequency) than that of the TNT. If the disk filter of size 8 is applied, the TNT feature will not be distorted but the TATB feature will. T his will result in incorre ct classification of the material. Even though applying the disk filter (size 8) shows an improvement of 11 dB (Figure 6 12) it is equally important to have a filter

PAGE 74

74 that has a uniform signal reduction throughout a large portion of the frequency spectrum in order to avoid distortion and data loss. Hence, only a small filter size disk filter (4) can be considered for our application. Since the disk filter of filter size 4 increase the SNR by 6 dB only it is less effective than the median filter presented i n the previous section. Figure 6 1 2 SNR as a function of the disk filter size for the differential reflectograms of TNT. The SNR ( 30 dB ) saturates at size=10, having a SNR improvement of 11 dB 0 5 10 15 20 25 30 35 0 2 4 6 8 10 12 14 16 18 SNR(dB) Filter Size Saturates at Filter size 10. SNR=30dB 11dB

PAGE 75

75 Figure 6 1 3 Differential reflectograms of TNT with increasing size s of average filter. Red curve is the unfiltered data while the green curve (filter size 4), blue curve (filter size 6) and black curve (filter size 8) show the filtered data. The amplitude difference of TNT features for red and black curves are shown for comparison. The filtered curves have been multiplied by 5 for clarity. D iffere ntial reflectograms of TNT for average filter s is shown in Figure 6 1 3 As seen the presence of distortion is clear as the different features get reduced by diffe rent amount s depending on their frequencies. The TNT shoulder feature (filter size = 8) is reduced 7 times (ratio of TNT feature of the red to the TNT feature of the black curve) compared to the unfiltered data. The peak A (higher frequency than the TNT fe ature) is diminished by 17 in amplitude for average filter size 8. We observe a simi lar effect of non homogeneity in amplitude reduction at the TNT feature to that of the disk filter The disk filter ha s a 4 :1 ratio of amplitude reduction while the a verage filter has a ratio of 2 :1 of amplitude reduction. As a result, a higher threshold can be set in order to select the filter size compared to the disk filter.

PAGE 76

76 Figure 6 1 4 SNR as a function of the average filter size for differential reflectograms of TNT. The SNR saturates at 28 dB and filter size 1 3 having a SNR improvement of 9 dB From the SNR comparison in Figure 6 1 4 we find that the SNR increases up to filter size of 1 3 before it saturates. T he maximum achievable S NR improvement is 9 dB from 19 dB (unfiltered data) to 28 dB (filter size of 1 3 ). However, at the saturation filter size, we would definitely observe the distortion effect Hence, a filter size of 5 can be chosen which increases the SNR by 5 dB while limit ing the distortion effect 6. 2 .3 Gaussian Filter As discussed in Section 5.3.3 t he re ad out noise of the camera is the white noise This type of noise model is Gaussian in nature [26 27 ] Hence, the Gaussian filter can be used to reduce this type of noise. The Gaussian kernel filter is defined by the size and the standard deviation. 0 5 10 15 20 25 30 35 0 2 4 6 8 10 12 14 16 18 SNR(dB) Filter Size Saturates at Filter size 13. SNR=28dB 9 dB

PAGE 77

77 6. 2 .3 .1 Understanding Gaussian Filter T he Gaus sian function is characterized by a bell shaped function according to the probability theory. The formula of a Gaussian curve is given by: The typical kernel filter of a gaussian filter of size x= 3 and standard deviation, using the above equation is as follows: 6. 2 .3 .2 Experimental Results D ifferential reflectograms of TNT are plotted for different sizes of the Gaussian filter a s shown in Figure 6 1 5 The pattern s are very similar to the disk and average filters.

PAGE 78

78 Figure 6 1 5 Differential reflectograms of TNT with in creasing sizes of Gaussian filter. The r ed curve is the unfiltered data while the green curve (filter size 4), blue curve (filter size 6) and black curve (filter size 8) show the filtered data. The filtered curves have been multiplied by 5 for clarity. The TNT feature (filter size = 8) is reduced by 7 times in amplitude (ratio of TNT feature of the red curve to the TNT feature of the black curve multiplied by 5) compared to unfiltered data. The peak A is diminished 15 times in amplitude for t he Gaussian fil ter of size 8. The Gaussian filter has a 2:1 ratio in amplitude reduction which is less than the disk filter (4.5:1) and equivalent to the average filter (2 :1).

PAGE 79

79 Figure 6 1 6 SNR as a function of the Gaussian filter size for differential reflectograms of TNT. The SNR ( 27 dB ) saturates at size= 15, having a SNR improvement of 8 dB T he SNR of differential reflectograms of TNT for different Gaussian filter size s is shown in Figure 6 16 T he rate of increase is large at low filter sizes The SNR (27 dB) rea ches saturation at filter size 15. T he maximum achievable SNR improvement is 8d B. From the previous discussion related to the distortion effect presented in this section, the filter size between 7 can be considered although it only improves the SNR by 6dB This is more effective than the disk filter and average filter but is poor in comparison to median filter. 6. 2 .4 Moving Average Filter 6. 2 .4 .1 Understanding Moving Average Filter The prime motive for using a simple moving average filter is that it is abl e to reduce random noise or white noise whil e retaining the spectral shape [28 29] 0 5 10 15 20 25 30 35 0 5 10 15 20 SNR(dB) Filter Size Saturates at Filter size 15. SNR=27dB 8dB

PAGE 80

80 T h e general working principle of this filter can be represented in an equation as follows: where The schematic of the moving average is shown in Figure 6 1 7 Figure 6 1 7 Schematic of the Moving Average Filter (filter size 4) A) Pixel 4 is the reference filter. The average is computed using the reference pixel and the previous 3 pixels (1, 2, 3) (B) The s ubset is shifted forward and pixel 5 is considered the reference a nd an average computed between pixel 2 to 5 If we consider a given series of numbers and a given subset size (the filter size), the moving average can be obtained by first taking the average of the first subset. The

PAGE 81

81 fixed subset size is then shifted forwa rd, creating a new subset of numbers, which is again averaged. This process is repeated over the entire series of pixels The plot connecting all the averages of each subset is the moving average. The difference between the average filter and the moving av erage is that a kernel matrix is not used and the pixels ahead of the reference pixel are not considered. 6. 2 .4 .2 Experimental Results A differential reflectogram of TNT is compared with its filtered differential reflectogram using moving average filter. We observe that the noise around the TNT feature from 356 nm to 470 nm is reduced and the signal is smoothed out considerably as we increase the filter size T here was no observable decrease in the amplitude as it wa s the case for the disk, average and Gau ssian filters However, there is one major drawback to this filter : significant spectral shift. The re is a 7nm shift f rom unfiltered data to filter ed data of size 8. To explain this shift in detail, a nother set of measurements experiments have been carried and are presented in the next section.

PAGE 82

82 Figure 6 1 8 Differential reflectograms of TNT with increasing sizes of moving average filter. The r ed curve is the unfiltered data while the green curve (filter size 2 ), blue curve (filter size 4 ) and black curve (filter size 8) show the filtered data. The amplitude difference of TNT features for red and black curves are shown for comparison Figure 6 19 SNR as a function of the moving average filter size for the differential reflectograms of TNT. The SNR=30 dB and no noticeable saturation is observed. 0 5 10 15 20 25 30 35 0 2 4 6 8 10 12 14 16 18 SNR(dB) Filter Size

PAGE 83

83 The SNR, as shown in Figure 6 19 increases gradually. It does not saturate even up to filter size 16. W e are able to achieve an improvement of 11 dB un t il a filter size of 16. The curves would probably saturate at a larger filter size s 6. 2 .4.3 Effect of Spectral Shift on the Differential Reflectogram of TNT As explained in the previous section, the moving average filter has one major drawback, i.e. it shifts the features spectrally. The shift is seen in Figure 6 2 0 and becomes more pro minent at high er filter size s Even though t he curve becomes very smooth (i.e. higher SNR), one should keep in mind that this shift can be very d isturbing since the spectral shift will cause an increase in false positives and true negatives rate in the classification of materials. Figure 6 2 0 The signal (shoulder) shifting towards the larger wavelengths when the size of the moving average filter increases. -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 250 300 350 400 450 500 550 DeltaR/Rbar(a.u.) Wavelength(nm) Raw Data Filter Size 4 Filter Size 8 Filter Size 16 Filter Size 32 Filter Size 40 Filter Size 2

PAGE 84

84 The wavelength of the inflection point of the shoulder around 420nm i s considered. Figure 6 21 display the effect of filter size on the inflection point of the TNT feature. Large fil ter size will shift the inflection point, while small filter size will minimally shift the spectral features. Figure 6 2 1 Inflection point a s a function of the filter size from differential reflectogram s of TNT. The picture in the inset shows where the infle c tion point is located on the TNT feature. In order to consider the best filter output, we should keep in mind that a maximum spectrum shift of 3 nm to 5 nm is acceptable, therefore the maximum filter size allowed for our application is 4. In this case the SNR improvement is 4 dB. 6. 3 Custom based Hybrid Filter Several imag ing filters have been presented in detail. From the differential r eflectograms of disk, a verage and Gaussian filters amplitude features decrease non 420 425 430 435 440 445 450 455 0 5 10 15 20 25 30 35 40 45 Shoulder's inflection Point(nm) Filter Size

PAGE 85

85 homogenously as a function of the filter size whereas there is no amplitude reduction in the TNT feature for the median and moving average filters. Therefore, we conclude t hat the median and the moving average filters are the best options compared to other filters presented. As discussed in the Chapter 4, the white noise and the spike like noise are inherent to all light sources. The spike like noise and the high frequency n oise observed in the xenon lamps (Section 5.4.4 and 5.4.6) are reduced by the median filters. On the other hand, t he moving average filter is better at reduc ing the white noise. Therefore, it would be the best to take advantage of both filt ers (moving aver age and median) that is to use the median filter of size 1 0 as explained in Section 6.2.1.2. The signal improvement is 8 dB. As concluded in the previous section, it is best to keep the spectral shift to a minimum, therefore we will use the moving average filter of size 4 The SNR increases by 3 dB when filter size of 4 is used From the above discussion we propose that a hybrid filter be used. The hybr id filter works in the following way. The unfiltered data is fi rst passed through the median filter of size 1 0 The spikes at 60 Hz (and its multiples) and the initial high amplitude noise seen from the Xenon lamp are therefore reduced. Now, t he data contains mostly white noise. The data is now passed through the movi ng average filter of size 4. This decreases the random noise. The result of the hybrid filter is shown in Figure 6 2 2 We again look at the two important aspects: spectral features and SNR improvement. There is no reduction in the amplitude feature or spec tral distortion as seen in the Figure 6 23. The SNR improves from 19.0 dB ( unfiltered) to 26. 7 dB (median filter) and then finally to 29.3 dB (hybrid filter). A pplying the hybrid filter help s to increase the explosive s detection sensitivity, i.e., Signal to Noise ratio.

PAGE 86

86 Figure 6 2 2 D ifferential reflectogram of TNT and application of different filters as marked on the graph. The hybrid filter is combination of median filter (size 10) and moving average (size 4).

PAGE 87

87 6. 4 Summary The study of various image processing filters is best summarized below in a table format. Table 6 1. Comparison of i mage processing filter s. Median Filter Average Filter Di sk Filter Gaussia n Filter Moving Average Filter Hybrid Filter Type of noise reduced Spike noise from power outlet, large amplitude in spectral density High frequency noise High frequency noise Read out noise White Noise All types of explained here Maximum SNR(dB) achieved 27 28 30 27 30 29.3 Presence of spectral shift (more than 3nm) No No No No Yes (size 5 and greater) No Filter size at which SNR saturates 10 13 10 15 Very high filter size (not measured) 10 SNR improvement 8 9 11 8 11 10.3 Spectral distortion No Yes Yes Yes No No

PAGE 88

88 CHAPTER 7 CONCLUSION AND FUTURE WORK 7.1. Conclusion Optical properties of energetic materials (TNT and TATB) are quite significant because they provide a nondestructive standoff means for providing the characteristics of these material s It has been observed that due to the chemical compositions, each energetic material results unique optical properties The principle of d ifferential reflectometry serves as a me an for fingerprinting the explosive materials in the near UV Visible spectrum (200 nm 550 nm) For example, TNT an d TATB have features around 420 nm and 48 0 nm respectively settings, offset and distance between the collection optics and the sample ) were discussed and the Signal to Noise Ratio (SNR) compared The optimum al working range for the exposure time is between 65 ms and 75 ms. While the offset has no effect on the SNR a gain setting between 650 and 750 is chosen for optimum performance. The shorter the distance between the sample and collection optics, the stronger is the signal T h e components ( camera, light source and spectrometer ) that make up our explosive detector, were analyzed, in particular the noise they produce From this study we observed that the light source s contain white noise, periodic noise (60 Hz) and other large amplitude noises. Additionally, CCD camera n oise was characteriz ed Dark noise and read out noise are the two main sources The read out noise is inherent and measured to be 1.6 8 counts while the dark current noise is dependent on the exp osure time and was measured to be 0. 72 counts at the maximum exposure time. The control experiment indicated that there are spurious effects at the 60 Hz and its harmonics coming from the power supply of the wall outlet. The QTH lamp and LED produced th e

PAGE 89

89 enough ultra violet light The 300 W Xenon arc lamp after 30 min warm up time was found to be the best among the UV producing light sources. To reduce the characterized noise from the lam p and detector several filters were investigated (m edian, average, disk, moving average and hybrid filter s) T he hybrid filter wa s found t o be the best for our application. It is applied using a two step process. The unfiltered data i s passed through a median filter o f size 1 0 T his result i s passed through a moving average filter of size 4. We observe that the hybrid filter is able to reduce most of the noise Additionally it does not shift the spectrum or reduc e the feature amplitude Finally, the SNR is improved by 11 dB using such a filter. Through this study, it is ensured that smooth, lossless and low noise signal s can be obtained, which in turn, would provide a n effective mean s to identif y explosives especially near the limit of d etection i.e., when the signal becomes very small. An additional consequence is the decrease false positives rate when classifying explosives versus non explosives materials. 7.2 Future Work Having progressed reasonably in achieving our goal (increase sign ificantly the SNR) there is still more to do in o rder to optimize and improve our explosive s detector system A few of the pot ential areas of work are listed below. On the hardware aspect, the present light source produces a lot of noise in the low fre que ncy area and this is the region where our prototype currently functions. It is not possible to work in the high er frequen c ies due to the limitation from the CCD camera (the current camera employed is the best commercial product available for our application) Hence, it is very important to look for other possible light sources From the discussion pres ented in Chapter 5, it has been shown that the usage of LED would be

PAGE 90

90 very promising. output and the absence of arc formation inc reases its efficiency many folds. We are currently developing an at specific wavelengths between 200 nm and 550 nm. On the software aspect, the computation time has to be evaluated at first, in order to comply with the w orking s peed of the prototype. The data acquisition and data processing bottle neck s should be minimized This could be achieved for example, by multi threading processes, wherein several threads could be effectively run in parallel. Finally, t he filter d esign in conjunction with the computation time must be studied Pr ese ntly, the computation time i s not used to choose or optimize our filter. A sys tematic clocking technique should be put in place to accurately clock the speed of each filter and optimize i ts design.

PAGE 91

91 APPENDIX A GRAPHICAL USER INTER FACE A s hort description of the Graphical User Interface (GUI) is provided The GUI developed is custom made for our explosives detection system The motive for developing the GUI is to fully automate the operation of the explosives detector. The GUI was written in C++ using the Microsoft Foundation Class (MFC). There are three main windows for this GUI. The first window is dedicated to control of the CCD camera. The exposure time, gain setting and offset f or each channel tap can be adjusted For our application, the camera runs in the continuous mode. This is shown in Figure A 1. Figure A 1. GUI window for CCD Camera settings control. The next window (Figure A 2) is the Pixel depth adjustment window. Thi s GUI indicates the maximum pixel depth values that the camera can take. It can be varied

PAGE 92

92 from 255 (2 8 bits) to 4095 (2 1 2 bits). It is generally set at 4095 counts for best brightness and contrast. Figure A 2. GUI window for selecting the camera pixel depth (8, 10 or 12 bits ) The third window is the video frame acquisition window, shown in Figure A 3 Figure A 3. GUI window for frame acquisition from the camera video stream.

PAGE 93

93 This is the main GUI window where the data viewing takes place. In this win dow, Once the frame is acquired, video streaming resumes particular frame and waits to process the acquired data Freeze s that particular frame and save s to an extern al storage location (Excel spreadsheet) Repeating the operation is necessary to provide 2 frames, then the differential reflectogram of the two frames are computed and displayed in an external application (Excel spreadsheet) More features are being added to minimize the user interaction. The addition of filters, a method to store data locally and viewing the differential reflectograms dynamically during run time are some examples of future GUI options

PAGE 94

94 LIST OF REFERENCES [1] Comprehensive Convention on International Terrorism in Draft Convention on International Terrorism, Embassy of India New Delhi, India, 2004 [2 ] Report of the Official Account of the Bombings in London on 7 th July 2005 in House of C ommons London, 11 th May, 2006. [3] Report of the Final Report on 2008 Terror Attacks (Under Section 173 Cr.PC) in the Court of ADDL.CH.M.M., 37 th Court Esplanade, Mumbai, 2009 [4 ] George Ingham Brown, The Big Bang: A History of Explosives in Sutton Publications 1998 [5 ] Fundamentals of Hazardous Materials Incident in CRC Press 1991 [6] G. Carlton. Smith, TNT: Trinitrotoluenes and mono and di nitrotoluene, their Manufacture and Propertie s in D.Van Nostrand Company 1918 [7 ] Anna Marie Fuller, Investigation of select energetic materials by differential reflection spectrometry in PhD Dissertation, University of Florida USA, 2007. [8] Jai Prakash Agarwal and R.D. Hodgson, Organic Ch emistry of Explosives in John Wiley & Sons 2007 [9 ] Rolf. E. Hummel, Differential Reflectometry and its application in Materials Science, Surface and Interface Analysis, Vol 12, Issue 1, P 11 14 July 1998 [10 ] Rolf. E. Hummel, Differential reflectometry and its application to the study of alloys, ordering, corrosion, and surface properties in Phys. Stat. Vol (A), Vol 76, No 1, pp 11 44 1983 [11 ] Anna. M. Fuller, Rolf E. Hummel, Claus Schoellhorn and Paul Holloway, Standoff Detection of Explosive Materials by Differential Reflection Spectroscopy in Proceedings SPIE Optics East Boston, MA, October 2 5, 2006. [12 ] Ro lf. E. Hummel, Anna. M. Fuller,Claus Schoellhorn and Paul Holloway in Applied Physical Letters 88, (2006) 231903 [13 ] R.E. Hummel ,A.M. Fuller, C. Sc hoellhorn, and P.H. Holloway, Trace Chemical Sensing of Explosives in Remote Sensing of Explosive Materials using Differential Reflection Spectroscopy in John Wiley N.Y. (2007) Chapter 15, page 301

PAGE 95

95 [14 ] Claus Schoellhorn, Anna M. Fu ller, Julien Gratier, and Rolf E. Hummel, Developments on Standoff Detection of Explosive Materials by Differential Reflectometry in Applied Optics 46 (2007), 6232 [15 ] Claus Schoellhorn, Anna M. Fuller, Julien Gratier, and Rolf E. Hummel, New Developments on Standoff Detection of Explosive Materials by Differential Reflectometry in SPIE Defense and Security Symposium proceedings 9 13 April 2007, Orlando,FL [16 ] Rolf. E. Hummel, Differential Reflectometry in Journal de Physique page C10 105, December 1983 [1 7 ] C.W. Shanley, R.E. Hummel and E.D Verink, Jr, Differential Reflectometry: A New Optical Technique .to Study Corrosion Phenomena n Corr.Sci, 20(1980) 467 [1 8 ] D. J. Schroeder Astronomical optics in 2 nd Ed, Applied Phys ics [19] Rochester Institute of Technology, CLEA Workshop, June 2004 [20] CRC Press, 2005 [21] Toru Yoshizawa, Handbook of Optical Metrology: Principles and Applications in CRC Press, 2009 [22 ] Tong Sun and Yrjo Neuvo, Detail Preserving Median Based Filters in Image Processing in Pattern Recognition Letters, Vol 15, Issue 4, Apr 1994, pp 341 347 [2 3 ] Nodes. T and Gallagher. N, Jr, Median Filters : Some modifications and their properties in IEEE Transactions, Vol 30, Issue 5, Oct 1982, pp 739 746 [24 ] C.F. Hester and D. Casasent, Multivariant Technique for multiclass pattern recognition in Applied Optics, 1980 [25 ] R. Bouzerar, M.Y. Jaffrin, L.Ding and P. Paullier, Influence of Geometry and Angular Velocity on performance of a rotating disk filter in IEEE Transactions, 2007 [26 ] K. Ito and K. Xiong, Gaussian Filters for nonlinear filtering problems in IEEE Transactions, Vol 45, Issue 5, pp 910 927, May 2000 [27 ] R. Gonzalez and R.E. Woods, Gaussian Blur in Detail in Digital Image Processing, Prentice Hall. 2008

PAGE 96

96 [2 8 ] J. H ellerstein, Y. Diao, S. Parekh and D.M. T ilbury, Feedback Control of Computing Systems in Wiley IEEE 2004, p 268 [29 ] International Conference on Signal Processing, Orlando, Florida 2003 [3 0 ] Siu Leong Iu, Bensalem and Pa, Noise reduction system using multi frame motion estimation outlier rejection and trajectory correction Patent Number : 5361105, Nov 1, 1994. [3 1 ] Jeff Medkeff, Using Image calibration to reduce noise in digital images in International conference on Image Processing Jan 6 2006.

PAGE 97

97 BIOGRAPHICAL SKETCH Karthik Vishwanathan was born in Chennai, India in 1985. He graduated from S ri S ivasubramania N adar College of Engineering, Anna University in May 2007 with his Bachelor of Engineering in Electrical and Electronics Engineering specializing in electronic materials. He then began attending the University of Florida, Gainesville in August of 2008 for the M aster s of Science degree program in the department of Electrical and Computer Engineering. During the s pring of 2009, he joined the Opto Electronic laboratory in the department of Material Science and Engineering. He worked under the supervision of Prof. H ummel and Dr. Thierry Dubroca as a research assistant on explosives detection.