Segmentation of Forensic Latent Fingerprint Images ... - IEEE Xplore

1 downloads 0 Views 612KB Size Report
Segmentation of Forensic Latent Fingerprint Images. Lifted Contact-less from planar Surfaces with Optical. Cohererence Tomography. Rethabile Khutlang ...
2015 IEEE 39th Annual International Computers, Software & Applications Conference

Segmentation of Forensic Latent Fingerprint Images Lifted Contact-less from planar Surfaces with Optical Cohererence Tomography Rethabile Khutlang, Fulufhelo V. Nelwamondo

Ann Singh

Biometrics Research Group, Identity Authentication CSIR Modelling and Digital Science Pretoria, Republic of South Africa [email protected]

Laser Sources Group CSIR National Laser Centre Pretoria, Republic of South Africa materials and inorganic salts secreted from eccrine and sebaceous (from touching one’s face) glands. Water evaporation is a major factor in latent fingerprint collection since human glands secretions are aqueous-based. The second set of materials found on finger tips are contaminants that come into contact with fingers. Examples of these contaminants are oil, ink, paint or flour. The quality of the friction ridge impression left on a surface is affected by pliability of the skin, pressure of deposition, slippage and other physical factors.

Abstract—Lifting latent fingerprints through means that do not make contact with the surface where the fingerprint is imprinted, is advantageous in many ways. Some of these advantages include: being able to lift the print multiple times; there is no physical or chemical processing of a substrate required, the substrate can be concurrently analyzed for DNA for instance; and this can provide a non-destructive lifting of the fingerprint, something that can aid in scene preservation. In this paper, we present an automatic segmentation of latent fingerprint images lifted contact-less from planar surfaces using swept source optical coherence tomography. We do not perform any localization scans as we know the position of fingerprint impressions left on a substrate. The 3-D lifted scan is processed on a per cross-sectional image basis. First the cross-sections are filtered to reduce the effects of speckle noise, then the one dimensional Sobel edge detection is applied horizontally. The detected edge represents the substrate surface plus the latent fingerprint impression left on it. They are concatenated together to form a 2-D segmented image of the lifted fingerprint. After enhancement using contrast-limited adaptive histogram equalization, minutiae were extracted from the segmented images as an implicit quality evaluation procedure, on top of the subjective one carried out. Segmented images of latent fingerprints lifted off some substrates like glass and stainless steel were of sufficient quality for minutia extraction.

Latent fingerprints are either visible or invisible to the naked eye. There are contact-based chemical, electronic and physical techniques used to render invisible fingerprints visible and to lift them. The methods are destructive in that the latent fingerprint cannot concurrently be used for other tests; for instance after processing a substrate surface chemically, it cannot be analyzed for DNA. Chemical reagents used in latent fingerprint processing can be either specific to certain materials or not. Ninhydrin reacts with amino acids [2]; while ethyl cyanoacrylate polymerization is non-specific as it catalyzes any water based compound [3], to form a layer of white polymer which increases contrast between a fingerprint and the background. Fresh latent fingerprints are usually processed with powder-based techniques. Lifting latent fingerprints through means that do not make contact with the surface where the fingerprint is imprinted, is advantageous in many ways. Some of these advantages include: being able to lift the print multiple times; there is no physical or chemical processing of a substrate required. Furthermore the substrate can be concurrently analyzed for DNA for instance. On electrically conductive surfaces, the scanning Kelvin probe can be used to visualize latent fingerprints contact-less [4]. The image is formed by mapping the electrical potential from a surface to a 2-D plane. Chromatic white light (CWL) sensors are also a nondestructive technique to lift latent fingerprints. With no localization scans performed, as the position of the fingerprint was known, [5] lifted latent fingerprints off forensically relevant surfaces using a CWL sensor. On the other hand, [6] presented a method based on the optical polarization and specular reflection; they obtained superior image quality on specific surfaces, for instance the sticky side of tape. Infrared spectroscopic imaging has been used to lift latent fingerprints contact-less [7]. The method works with both absorbing and

Keywords— Forensics; Latent fingerprints segmentation; Contact-less lifting; styling; Optical coherence tomography (OCT)

I. INTRODUCTION Fingerprints are used prominently in crime scene forensics. A fingerprint comprises a series of ridges and furrows. Where these ridges end or bifurcates, a feature set called minutiae is defined. There are two general types of fingerprint data, exemplar and latent. Exemplar fingerprints are acquired directly from human finger for enrolment in a system or when under arrest for a suspected criminal offense; using specific fluids or scanners in a controlled environment. On the other hand, fingerprints left inadvertently at a crime scene are called latent fingerprints. When a finger comes into contact with a substrate surface, the dielectric residue corresponding to ridge patterns leaves an impression on the surface [1]. The principal endogenous materials found on finger tips are aqueous-based organic The work is sponsored by the Department of Science and Technology of the Republic of South Africa.

0730-3157/15 $31.00 © 2015 IEEE DOI 10.1109/COMPSAC.2015.166

30

non-absorbing surfaces, by analyzing the spectrum of the reflected light from a surface that was irradiated with near infrared light. As [8] pointed out, the technology is expensive as it needs active infrared light sources to be cooled to emit required frequencies and this need rules out its use in mobile crime scene devices.

gradient |∇u| by its estimate |D‫ܩ‬ఙ * u| in the Perona and Malik model (1). The second step in the pipeline presented is the application of a one dimensional Sobel edge detection in the horizontal direction [11]. At each point in the cross-sectional image A, the gradient approximation ‫ܩ‬௬ is derived from the convolution of that point with a 3 x 3 kernel for the horizontal direction:

We present an edge detection-based segmentation method to produce 2-D latent fingerprint images from signal reflected off ridge impressions left on planar surfaces using swept source optical coherence tomography (SS-OCT). We do not perform any localization scans as we know the position of the fingerprint impression. The remainder of this paper is arranged as follows: Section II presents the materials and methods used in this paper, followed by the presentation of results in section III. Section IV presents a discussion, and Section V concludes.

െͳ െʹ െͳ ‫ܩ‬௬ ൌ ൥ Ͳ Ͳ Ͳ ൩ ‫ܣ כ‬ሺʹሻ ൅ͳ ൅ʹ ൅ͳ The sensitivity threshold for the gradient is determined automatically. If the cross-section has two edges, the bottom one is ignored as it corresponds to the width of a substrate – for substrates for which OCT penetrates their entire width. To compensate for inaccuracies in edge detection, the final substrate plus fingerprint impression edge is taken as an average of 25 pixels across the detected edge; 25 was found empirically. A 2-D latent fingerprint image is obtained by concatenating the edges detected in the 512 cross-sectional images that constitute a 3-D scan together. Finally, the contrast of the lifted 2-D latent fingerprint image is enhanced using contrast-limited adaptive histogram equalization [12].

II. MATERIALS AND METHODS OCT was first used by [9] to lift latent fingerprints, and reported only on a glass substrate surface. In this work, a swept source OCT system (OCS1300SS, Thorlabs, U.S.) was used to lift latent fingerprints off substrate surfaces investigated. The swept laser optical source has a central wavelength of 1325 nm and a spectral bandwidth of 100 nm. It has an average power output of 10.0 mW, and an axial scan rate of 16 kHz. The system has a maximum imaging depth of 3 mm. The process starts by acquiring an A-scan. An A-scan provides information about the reflective or scattering properties of the substrate under investigation as a function of depth at a certain position of the scanned beam. The latent fingerprint is superficial to the substrate surface. A collection of A-scans results in a cross-sectional image (B-scan). A collection of B-scans results in a 3-D volumetric image.

C. Quality evaluation of lifted latent fingerprint images Quantitative evaluation of the quality of lifted 2-D latent fingerprint images was not conducted. Lifted images were evaluated subjectively. In forensics, a dactyloscopic expert analyses a fingerprint and the substrate to conclude whether a fingerprint impression is present and assesses the degree of the detail. As an implicit 2-D lifted latent fingerprint image quality evaluation, we automatically extract minutiae, to augment the subjective evaluation. This is because a dactyloscopic expert would mark minutiae in latent fingerprint matching. From the contrast enhanced images, minutiae were extracted using the Secugen FDx SDK, version 3.7.

A. Materials found on finger tips and substrates investigated We lift latent fingerprints off stainless steel, polypropylene slab, glass, polyvinyl chloride (PVC) synthetic plastic panel, rubber washer and aluminum flat substrate surfaces using a swept source OCT system. The fingerprint impressions are made on a pre-marked area on a substrate. We test for three materials found on finger tips; endogenous aqueous-based eccrine and sebaceous glands secretions; petroleum jelly and used motor oil.

III. RESULTS An area of 10mm by 10mm was scanned for each material on fingertip substrate combination. OCT scanning was performed within the first hour of making a fingerprint impression on a substrate. There was no coarse scan performed to localize a fingerprint impression as an impression was made on a pre-defined area. Investigated substrates were everyday objects that were not cleaned or smoothed in any special way; the only condition was that they be flat. Extracted minutiae are overlaid on the 2-D images segmented using the developed pipeline. Blue cycle annotates a ridge ending; and the arrow points to the direction that ridge came from. Green cycle annotates a bifurcation; and the arrow points to the direction that ridge would have taken had it not split. Results for three cases are given below, when the material on finger tips was nothing but aqueous-based secretions from eccrine and sebaceous glands, petroleum jelly, and used motor oil.

B. Substrate surface edge detection Since the reflection characteristics of the substrates surface and the residual material left by the finger impression are buried in noise in OCT imaging, the first step in the imaging pipeline presented is image denoising. Cross-sectional images of the 3-D OCT scan are filtered using the regularized version of the Perona and Malik PDE filter due to its stability in the presence of speckle noise [10]. The Perona and Malik filter improves on the heat equation (equivalent to the convolution of the signal with Gaussians at each scale) with the signal as initial datum by reformulating it as a nonlinear equation of the porous medium type: u / dt = div (g(|∇u|)∇u),

u(0) = ‫ݑ‬଴

A. Endogenous aqueous-based secretions from glands This case can be considered as the naked fingertip case; the fingerprint impression is made by natural occurring sweat. Fig.

(1)

In this equation, g is a smooth non-increasing function. The stability brought about by [10] is due to the replacement of the

31

Fig. 1. First column is a montage of 3-D OCT latent fingerprint images in the case of a naked fingertip: lifted from i. polypropylene, ii. glass, iii. rubber, iv. aluminum, v. stainless steel, vi. pvc. Second column is a montage of corresponding segmented images with minutiae overlaaid.

similar; which suggests that peetroleum jelly plays a significant role in the quality of images obttained.

1 shows the 3-D images lifted when the imppression was made using a naked finger; and the correspondingg 2-D images with minutiae overlaid. Good images were lift fted off the glass substrate. Stainless steel had comparrable quality to polypropylene even though OCT penetratess polypropylene a lot more than it does stainless steel. No minuttiae were extracted from aluminum, PVC and rubber emphasizinng the poor quality of latent fingerprints lifted from those substraates.

C. Used motor oil Fig. 3 displays 3-D OCT latent fingerprint images lifted from substrate surfaces when the impression was made from fingertips dipped in used mottor oil. For each substrate, the segmented latent fingerprint 2-D D image is shown, overlaid with extracted minutiae. Even thhough no minutiae would be extracted from PVC, rubber and stainless steel, fingerprint ridges are readily visible on all substrates. Visually, glass substrate seems to yield an image of better quality than the rest. However, the difference is minnor, which again suggests quality is more of a function of used motor m oil than it is of the substrate in question. This is a weak deduction as quality was not objectively evaluated.

B. Petroleum jelly Fig. 2 presents 3-D OCT latent fingerpprint images lifted from substrate surfaces when the impressioon was made from fingertips smeared with petroleum jelly. Corresponding C 2-D images obtained using the developed pipelinne are also shown, extracted minutiae overlaid. Rubber waas the most uncooperating substrate. Quality of the 2-D latent fingerprint images from the rest of the substrate surfaces was more or less

Fig. 2. First column is a montage of 3-D OCT latent fingerprint f images in the case of a fingertip smeared with petroleum m jelly: lifted from i. polypropylene, ii. glass, iii. rubber, iv. aluminum, v. stainless steel, vi. pvcc. Second column is a montage of corresponding segmented images with minutiae overlaid.

32

Fig. 3. First column is a montage of 3-D OCT latent fingerprint fi images in the case of a fingertip dipped in used motor oiil: lifted from i. polypropylene, ii. glass, iii. rubber, iv. aluminum, v. stainless steel, vi. pvc. Secoond column is a montage of corresponding segmented images with minutiae m overlaid.

camera and light source. UV-Imaging U was found to be conducive to lifting latent finggerprints from porous substrates, for example copying paperr. It struggled to lift latent fingerprints from the non-porous substrates that were investigated – curved power ouutlet cover, matte metallic paint and furniture surface.

IV. DISCUSSION The developed pipeline segments a 2-D D latent fingerprint image from the 3-D OCT scan of a fingeerprint impression deposited on a substrate surface. A scan is processed p on a per cross-sectional image basis. Firstly the slicces are filtered to reduce the effects of speckle noise. Thenn the Sobel edge detection is applied in the horizontal directtion. The detected edge is averaged across a 25 pixel wide bannd to counter edge detection inaccuracies. The detected edges arre concatenated for all the cross-sections of a scan. The resultaant image, the 2-D image of the lifted latent fingerprint, is contraast enhanced using the contrast-limited adaptive histogram equallization. Quality of the resultant image was not objectively evaluated; rather minutiae were extracted as an implicit indicattion of quality.

The prime challenge with thhe current SS-OCT system is the area than can be scanned. The maximum m area that was scanned was 10mm by 10mm. This willl be a limitation on the matching that can be done after further im mage enhancement and minutiae extraction. Current images do not mostly look like fingerprint images when subjectively evvaluated because of the area restriction. The ability to exttract minutiae was used as an indication of quality, it is weakk especially because accuracy of minutiae is not considered but b rather ability to find any minutiae is used as a quality assessment a in this work. To aid image enhancement in ordder to improve lifted latent fingerprints, objective quality assessment of the segmented images will be investigated in future. f

To our knowledge, glass is the only subbstrate from which OCT has been used to lift latent fingerprints [1], [9]. They used a white fluid to make a fingerprint impressioon. They recorded interferograms for each step in sweeping the t input RF, and stacked them together along the wavellength axis. The corresponding variation of intensity along thhe wavelength was computed, for a fixed lateral position of thhe interferograms. The FFT of the interference fringe signal prrovides peaks that correspond to different depth layers withhin the scan. 2-D images could be reconstructed at different z-axis heights from the substrate and subjectively evaluatedd for fingerprint information presence. The height that correspponds to maximum fingerprint signal correlated with the second-order filtered Fourier peak.

The 2-D images obtained using u the presented pipeline had horizontal lines at times as an image artefact. This was caused by cross-sectional images withhin a 3-D OCT scan for which horizontal Sobel edge detectionn struggled to find the surface of the substrate. This was mostly observed o where not enough light was reflected from the substraate. One way to eliminate that might be to have an array of eddge detection techniques; and the signal to noise ratio of an inpuut 3-D OCT scan can be used to select the appropriate edge deteection technique. No experiments were conduucted to investigate the effects of time on the lifted fingerprints. This T will be investigated, so will the ability to lift latent fingerrprints under a layer of dust as reported in [1]. An avenue woorth exploring is the differential image approach [8]. This is pertinent to lifting latent fingerprints using OCT as soome of the substrates were not smooth. Everyday objects werre used. The differential image approach might remove any suubstrate surface texture from the lifted image; and this might pronounce p the fingerprint ridge structure left by an impression.

From the review conducted in [8], chroomatic white light sensor showed superior detail of fingerprint inntensity images on smooth non-textured substrates, for instancee furniture surface. A differential image approach was used too avoid additional pre-processing methods to visualize a finngerprint residue. Furniture surface was also found to be cooperative c to the camera with polarizer; using the combinnation of multiple images acquired with different polarization filter angles as a processing technique. The method is not suitable for nonplanar substrate surfaces because of the requuired angles of the

33

Chromatic white light sensors have been reported on as a mechanism that lifts good quality images [8]. The images subjectively deemed of good quality using OCT are comparable to those obtained using a chromatic white light sensor. For instance, lifted latent fingerprints from a glass substrate used in conjunction with naked fingertips were of good quality. Images of latent fingerprints lifted when a naked finger had left an impression were generally of a good quality. And there was a big difference in quality between cooperating substrates and those that do not. This suggests that material on the fingertip; petroleum jelly and used motor oil in this case, play a part in the quality of the lifted fingerprint image.

REFERENCES [1]

S. Dubey, D. Mehta, A. Anand, and C. Shakher, “Simultaneous topography and tomography of latent fingerprint using full-field sweptsource optical coherence tomography,” J. Opt. A: Pure Appl. Opt., vol. 10, pp. 1-8, Jan 2008. [2] N Petraco, G. Proni, J. SJackiw and A. Sapse, “Amino acid alanine reactivity with the fingerprint reagent ninhydrin. A detailed ab initio computational study,” J. For. Sci., vol. 51(6), pp. 1267-1275, Nov 2006. [3] J. Day, H. Edwards, S. Dobrowski, A. Voice, “The detection of drugs of abuse in fingerprints using Raman spectroscopy II: cyanoacrylate-fumed fingerprints,” Spectrochimica Acta Part A, vol. 60, pp. 1725-1730, July 2004. [4] H. Dafydd, G. Williams and S. Bleay, “Latent fingerprint visualization using a scanning Kelvin probe in conjunction with vacuum metal deposition,” J. For. Sci., vol. 59, pp. 211-218, Jan 2014. [5] M. Leich, S. Kiltz, J.Dittmann and C. Vielhauer, “Non-destructive forensic latent fingerprint acquisition with chromatic white light sensors,” In IS&T/SPIE Elec. Imaging, pp. 78800S-78800S. Inter. Society for Optics & Photonics, 2011. [6] S. Lin, K. Yemelyanov, E. Pugh and N. Engheta, “Polarization-based and specular-reflection-based noncontact latent fingerprint imaging and lifting,” J. of Opt. Society of America, vol. 23, pp. 2137-2153, 2006. [7] N. Crane, E. Bartick, R. Perlman and S. Huffman, “Infrared Spectroscopic Imaging for Noninvasive Detection of Latent Fingerprints,” J. For. Sci., vol. 52(1), pp. 48-53, Jan 2007. [8] S. Kiltz, M. Hildebrandt, J. Dittmann and C. Vielhauer, “Challenges in contact-less latent fingerprint processing in crime scenes: Review of sensors and image processing investigations,” 20th European Signal Proc. Conf., pp. 1504-1508, Aug. 2012. [9] S. Dubey, T. Anna, C. Chakher and S. Mehta, “Fingerprint detection using full-field swept-source optical coherence tomography,” Applied Physics Letters, vol. 91, pp.181106-181108, Oct 2007. [10] F. Catte, P. Lions, J Morel, and T. Coll, “Image selective smoothing and edge detection by nonlinear diffusion,” SIAM J. Num. Anal., vol. 29, pp. 182-193, Feb. 1992. [11] W. Pratt, Digital Image Processing. New York, NY: Wiley, 1978. [12] Z. Karel, “Contrast Limited Adaptive Histograph Equalization,” Graphic Gems IV. San Diego, CA: Academic Press Professional, 1994.

V. CONCLUSION An automated image processing pipeline has been developed to segment a 2-D latent fingerprint image from a 3D OCT scan of a fingerprint impression deposited onto a substrate surface. The slices of an OCT scan are first filtered, then Sobel edge detection is applied to find the substrate surface plus fingerprint edge, which is averaged over a 25 pixel band. The detected edges are concatenated together and contrast enhanced. Quality of the images was implicitly evaluated by extracting minutiae. The method produces images potentially relevant to crime scene forensics, from a contactless acquisition sensor. The main feature of OCT acquisition to be improved for forensic application is the area that can be scanned. Latent fingerprints could be lifted off some substrates like glass and stainless steel with sufficient quality for minutia extraction.

34