Identifying the Origin of Finger Vein Samples Using Texture Descriptors
IIDENTIFYING THE ORIGIN OF FINGER VEIN SAMPLESUSING TEXTURE DESCRIPTORS
Babak Maser † Multimedia Signal Processing& Security Lab ‡ University of Salzburg, [email protected]
Andreas Uhl
Multimedia Signal Processing& Security Lab ‡ University of Salzburg, [email protected]
ABSTRACT
Identifying the origin of a sample image in biometric sys-tems can be beneficial for data authentication in case of at-tacks against the system and for initiating sensor-specific pro-cessing pipelines in sensor-heterogeneous environments. Mo-tivated by shortcomings of the photo response non-uniformity(PRNU) based method in the biometric context, we use a tex-ture classification approach to detect the origin of finger veinsample images. Based on eight publicly available finger veindatasets and applying eight classical simple texture descrip-tors and SVM classification, we demonstrate excellent sen-sor model identification results for raw finger vein samplesas well as for the more challenging region of interest data.The observed results establish texture descriptors as effectivecompetitors to PRNU in finger vein sensor model identifica-tion.
Index Terms — Texture Classification, Sensor Identifica-tion, Image Origin Authentication, Finger Vein Recognition,PRNU
1. INTRODUCTION
Nowadays we encounter a significant surge in the use of unat-tended applications of biometric systems. In many biomet-ric modalities, a digital biometric image sensor is the corecomponent for data acquisition, operating in the near-infrared(NIR) domain in the case of finger vein recognition.Deducing sensor information from the images serves as abasis for different forensic and non-forensic tasks. One of themajor tasks in digital image forensics is establishing an im-age’s origin with the help of the deduced sensor information.This can be performed at different levels: Sensor-technology,brand, model, unit. In the context of biometric systems theextracted sensor information can be used for various applica-tions. In this work we focus on two specific ones: Securing † ORCID iD: 0000-0002-1662-8324 ‡ a finger vein recognition system against insertion attacks andenabling device selective processing of the image data.The authenticity and integrity of the acquired biometricsample data plays an important role for the overall security ofthe biometric system, in particular in the case of unattendedoperation. Among other attacks (e.g. presentation attackswhich present spoofing artefacts to the biometric sensor), an insertion attack bypasses the biometric sensor by insertingdata (i.e. biometric samples) into the transmission from thesensor to the feature extractor / comparison module. The fin-ger vein image inserted during the attack could have been ac-quired with another sensor off-site, even without the knowl-edge of a genuine user, or could be a manipulated image tospoof the biometric recognition system.In large-scale biometric system various sensors from dif-ferent manufacturers and models are deployed and the inter-operability is often affected by specifics of each sensor, suchas the acquisition technique or in-sensor image processing.Selective processing of the images helps to improve the inter-operability by applying a sensor tailored biometric toolchain.Therefore information about the sensor model is required,which can be deduced from the iris images directly utilisingimage forensic methods.This work evaluates the feasibility of deducing sensor in-formation at model level, i.e. the biometric sensor a fingervein sample image is captured with, from the finger vein im-age using image texture based methods. To learn an image’sorigin, various methods have been proposed, the approach ex-ploiting the photo response non-uniformity (PRNU) being themost prominent one [1], as it also allows a sensor identifica-tion at unit level.However, the application of PRNU-based techniques inbiometric sample data authentication has exhibited some dif-ficulties: First, the PRNU fingerprint can be extracted fromimages of a biometric sensor and injected into forged sampleimages [2, 3]. Only under certain restrictive conditions suchattacks can be detected or avoided. Second, authentication re-sults with respect to biometric sensors have been reported tobe widely varying at best (see e.g. [4, 3] for iris sensor identi- a r X i v : . [ c s . C V ] F e b cation) and have been shown to be dependent and influencedby sensor components depicted in the images (see e.g. [5] forfinger vein sensor identification). One reason for the difficul-ties is the requirement to compute the PRNU fingerprint fromuncorrelated data - which is of course hard to satisfy giventhe high similarities among sample images present in biomet-ric datasets [6]. Attempts to clarify this issue have not beenconvincing so far [7]. As a consequence, texture classifica-tion techniques have been proposed to identify iris sensors atmodel level earlier [8, 9]. However, contrasting to this work,the authors of the former work propose rather costly tech-niques, i.e. improved Fisher vector encoding of denseSIFTand dense Micro-block difference features.This work is structured as follows: In section 2 we discussthe properties of the finger vein sample datasets as considered.Section 3 explains the conducted experiments in depth, wheresubsection 3.1 describes the used texture description method-ology. Next, We discuss and analyze the experimental resultsin section 4, and finally, we end this manuscript with a con-clusion in section 5.
2. FINGERVEIN SAMPLE DATA
We considered eight different public finger vein datasets (ac-quired with distinct prototype near infrared sensing devices),taking 120 samples from each dataset. As in finger veinrecognition features are typically not extracted from a rawsample but from a region-of-interest (ROI) image contain-ing only finger vein texture, an insertion attack can also bemounted using such ROI data (in case the sensor does not de-liver a raw sample to the recognition module but ROI datainstead). Thus, we produced cropped ROI datasets out ofthe original ones (description of methodology is given after-wards) to be able to test these data for their distinctivenessas well. Subsequently, we briefly detail the specifications ofeach dataset:•
SDUMLA-HMT [10]:
Original resolution is240 × ×
320 pixel. 120 imagesare selected from the first 30 subjects.•
HKPU-FV [11]:
Original resolution is 256 × ×
390 pixel. 120 images are selected from thefirst 60 subjects.•
IDIAP [12]:
Original resolution is 250 × ×
610 pixel. 120 images are selected from thefirst 60 subjects.•
MMCBNU 6000 (MMCBNU) [13] :
Original reso-lution is 640 × ×
620 pixel. 120images are selected from the first 20 subjects.•
PLUS-FV3-Laser-Palmar (Palmar) [14]:
Originalresolution is 600 × ×
500 pixel.120 images are selected from the first 20 subjects. •
FV-USM [15]:
Original resolution is 480 ×
640 pixels,ROI data is 110 ×
280 pixel. 120 images are selectedfrom the first 30 subjects.•
THU-FVFDT [14]:
Original resolution is 600 × ×
390 pixel. 120 images are selectedfrom the first 120 subjects.•
UTFVP [16]:
Original resolution is 380 × ×
490 pixel. 120 images are selected fromthe first 60 subjects.Original sample images, as shown in Fig. 1, can be dis-criminated easily: Besides the differences in size (which canbe adjusted by an attacker of course), the sample images canbe probably distinguished by the extent and luminance ofbackground. To illustrate this, we display the images’ his-tograms above each example in Fig. 1, and those histogramsclearly exhibit a very different structure. Thus, we expect tex-ture descriptors to have an easy job to identify the origin ofthe respective original sample images.
In finger vein recognition, contrasting to e.g. fingerprintrecognition, feature extraction is not applied to the entire rawsample data but instead to a ROI only [14, 11]. In this ROI,only actual finger texture is contained. Depending on thesetup of the system, the sensor might already extract the ROIfrom the raw sample. In this setting, identification of the fin-ger vein data’s origin has to be based on the ROI, thus, it willbe required under these circumstances that only finger veintexture (the ROI) is used to discriminate sensors. This is notunrealistic, as in the iris recognition case, normalised iris tex-ture has been considered by analogy to be used for sensoridentification [8, 9] instead of raw iris sample data.To detect and segregate the finger vein region and ex-tract a patch consisting of biometric data only, different tech-niques have been applied depending on the properties of eachdataset. For the datasets exhibiting a higher intra-variance offinger positions, we applied the following algorithm based onmorphological snakes (morphological active contour withoutedges [17]) to extract the ROI (FV USM, THU FVDT, UT-FVP, and MMCBNC datasets), also illustrated in Fig. 2:1. Apply morphological snakes to the finger vein image toproduce a segmented image.2. Apply Canny edge detection and contour closing to de-tect the finger vein region.3. Fill the contour.4. Find the mass center of the filled contour and fit a lineto the contour; estimate the angle ( θ ) of the line to thex-axis. a) (b) (c) (d) Fig. 1 : Image and corresponding histogram samples of original sample images of (a) HKPU FV dataset, (b) UTFVP dataset,(c) SDUMLA dataset, and, (d) IDIAP dataset.
Fig. 2 : ROI generation for datasets with finger position vari-ability.5. Rotate the texture area by θ degree.6. Find the new mass center and crop the aligned originalsample image.For the Palmar dataset we also used this method but re-placed the morphological snakes technique by the Chan-Vesesegmentation algorithm [18].To create the ROI for the (easier) datasets HK FV, IDIAPand SDUMLA we applied the following steps:1. Apply Canny edge detection.2. Apply a dilation operator on the detected edges.3. Stack all images of a dataset on top of each other.4. Extract the common patch patch of finger vein texture.Fig. 5 illustrates the results of ROI creation for a sampleof each dataset (sample width has been normalised for betterclarity). It gets immediately clear that discrimination is ob-viously more difficult based on the ROI data only. To inves-tigate the differences between raw sample data and ROI data in more detail, we have investigated the range of luminancevalues and their variance across all datasets. Figs. 3 and 4 dis-play the results in the form of box-plots, where the left box-plot corresponds to the original raw sample data, and the rightone to the ROI data, respectively. We can clearly see that theluminance distribution properties have been changed dramati-cally once we change our focus from original datasets to ROIdatasets. For example, original HKPU FV samples can bediscriminated from FV USM, MMCBNUm, PALMAR, UT-FVP, and THU FVFDT ones by just considering luminancevalue distribution. For the ROI data, the differences are notvery pronounced any more. When looking at the variancevalue distributions, we observe no such strong discrepancybetween original sample and ROI data, still for some datasetsvariance can be used as discrimination criterion (e.g. Palmarvs. HKPU FV in original data, FV USM vs. HKPU FV inROI data). Consequently, we expect the discrimination of theconsidered datasets based on texture descriptors to be muchmore challenging when focusing on the ROI data only. Fig. 3 : Luminance distribution of original and ROI imagesacross all datasets, respectively.The steps to produce the cropping images is shown in Fig-2. The size of cropped images for each dataset has been givenin the subsection 2. ig. 4 : Variance distribution of original and ROI imagesacross all datasets, respectively.
3. EXPERIMENTAL DESIGN3.1. Texture Description Methodology
To discriminate sensors we applied a number of classical yetsimple approaches to produce a texture descriptor of a fingervein image. In the following subsections, we briefly describethe chosen techniques and explain how to cope with differ-ently sized images.
Fourier Ring Filter (FRF)
We generate features in the frequency domain using 2-D FFT.Independent of image size, fifteen band pass filters split thefrequency domain into equally sized bands which are usedto compute mean and standard deviation of each ring [19](which are used as statistical texture descriptors).
Local Binary Patterns (LBP)
We use a variant of the original
Local Binary Pattern (LBP) introduced by Ojala et al . [20]. This approach is calledHistogram-LBP (HLBP [21]), we set the radius to 3 and thenumber of curricular neighborhood pixels is set to 15. TheHLBP is invarant to image size if the output of the histogramfor each image is normalized. Further, the number of his-togram bins is fixed.Additionally, we apply uniform LBP (ULBP) - a LBP iscalled uniform if the binary pattern contains at most two 0-1 or 1-0 transitions, and it has been shown that these patternoccur more frequently in natural texture (and significantly re-duces feature length vectors as the LBP histogram bins arereduced).
Image Histogram (IMHIST)
We simply compute the image histogram and take the outputas feature vector. The IMHIST is invariant to image size bybin entry normalisation and fixing the number of histogrambins.
Wavelet-based Features
We apply 2-D wavelet decomposition using Daubechies 8-taporthogonal filters to generate the coefficients in horizontal h , vertical v , and diagonal d directions. On every decomposi-tion level we compute mean ( µ ) and standard deviation (std)for each of the sub-bands v , h , and d and concatinate thoseto get the mean and variance feature (WMV). We achievedinvariance to image size by fixing the number of wavelet de-composition levels to 3.Similar to WMV, we define wavelet variance (WV) bycomputing the variance per subband, and wavelet entropy(WE) by computing entropy per subband, respectively. Local Entropy (LE)
We slice a given image into 16 blocks (tiles) and compute theentropy from each tile. By taking a histogram of all producedentropies a feature vector is generated. LE also is invariant toimage size by fixing the number of image blocks (tiles) andbin number of the histogram.
The SVM classifier is trained by feeding of all imagesthen the remaining images are used for the testing pur-pose. Images in all finger vein datasets are randomly shuf-fled beforehand to avoid subject-related bias. To optimize theSVM classifier and to obtain the most promising hyperparam-eters such as C , γ , kernel and degree , we employed a GridSearch technique in combination with 4-fold cross-validation[22]. Also, we set the decision function to ” one vs rest ” strategy. We use classical measures to rate our sensor identificationtask, which is basically a multi-class classification problem.The multi-class problem is an extension of binary classifica-tion. We use two approaches for evaluation: First, receiveroperating characteristic (ROC) which relates the false posi-tive rate to the false negative rate, and second, the relation ofprecision and recall. For both relations, the Area Under TheCurve (AUC) can be computed as a single measure.Once we have multi-class problem, the challenging pointis how to get an overall score. Often, one simply takes the av-erage of the AUC ROC metrics. For illustration, to calculateRecall for three-class problems, we sum up three Recalls anddivide them by number of contributing classes, that is classi-cal average. In contrast, another approach is based on sum-ming up individual terms during the computation. The formeris called macro-average and latter is called micro-average ap-proach, respectively [23, 24].In this paper we use the micro-average approach to esti-mate the average metrics of our multi-class problem. We usethe micro-average AUC ROC (mA AUC ROC) and micro-average AUC Precision-Recall(mA AUC Pr-Re) as our perfo-mance metrics. In the multi-class setting, we need to estimatethe aforementioned metrics by computing e.g. micro-average a) SDUMLA ROI(b) HKPU FV ROI(c) MMCBNU ROI (d) IDIAP ROI(e) UTFVP ROI(f) PALMAR ROI (g) FV USM ROI(h) THU FVFDT ROI
Fig. 5 : ROI Samples from different datasetsFPR (False Positive Rate), micro-average Recall, and micro-average Precision as follow: mA P r = (cid:80) c T P c (cid:80) c T P c + (cid:80) c F P c (1) mA Re = (cid:80) c T P c (cid:80) c T P c + (cid:80) c F N c (2) mA F P R = (cid:80) c F P c (cid:80) c F P c + (cid:80) c T N c (3)Where c is the class label, TP is True Positive, FN is FalseNegative, TN is True Negative, and FP is False Positive.In this work, the number of images in the contributed datasetsis balanced. Therefore, the value of macro-average andmicro-average are very close and sometimes even identical. To enhance the sample images, we applied the followingmethods.1.
Wiener Filter and CLAHE (Enh.):
To enhance thequality of the images and remove undesired noise-related artifacts, we apply a Wiener Filter [25] andalso to improve the contrast of images, we use CLAHE(Contrast Limited Adaptive Histogram Equalization).Applying these two filters is done sequentially on allimages of the mentioned datasets.2.
No Enhancement (NoEnh.):
Sample images are usedas present in the datasets or as obtained after ROI com-putation.
4. RESULTS
Table 1 displays the experimental results. As expected, sen-sor identification is easily achieved based on original sam-ples. Image enhancement improves results (mostly slightly)in many cases, we get values > Original Sample ROIDescriptor No.Enh Enh. No.Enh Enh. mA AUC ROC
FRF 0.997 0.999 0.989 0.986 mA AUC Pr-Re mA AUC ROC
HLBP 0.992 0.994 0.941 0.955 mA AUC Pr-Re mA AUC ROC
ULBP 0.995 0.999 0.931 0.932 mA AUC Pr-Re mA AUC ROC
LE 0.919 0.952 0.834 0.858 mA AUC Pr-Re mA AUC ROC
ImHist 0.989 0.961 0.906 0.966 mA AUC Pr-Re mA AUC ROC
WV 0.998 0.998 0.984 0.983 mA AUC Pr-Re mA AUC ROC
WE 0.993 0.985 0.977 0.959 mA AUC Pr-Re mA AUC ROC
WMV 0.999 0.999 0.982 0.994 mA AUC Pr-Re
Table 1 : Sensor identification results.For ROI data, results deteriorate slightly. In particular,UC Pr-Re for spatial domain techniques is no longer accept-able. DFT and wavelet-based descriptors however still resultin values well above 0.9, in most cases above 0.95, which isa very good result, that could not be expected given the highsimilarity of textures and the simplicity of our descriptors.For ROI data, there is no clear trend if enhancement as beingapplied is beneficial or not. In any case, similar to the originalsample case, for well performing techniques the difference isnegligible.
5. CONCLUSION
We have identified simple texture descriptors as being wellsuited for finger vein sensor model identification, being ap-plied to raw sample images as well as to the more challengingfinger vein ROI data. Enhancement techniques turn out to benon-decisive for classification accuracy, at least when consid-ering top performing techniques. Overall, but especially whenconsidering results for ROI data, Fourier and wavelet-domaindescriptors are found to perform superior to spatial domaintechniques.The excellent results suggest the proposed techniques tobe better suited as compared to PRNU-based methods for thetask investigated. Also, a fusion of both approaches seemspromising, which will be subject to further investigations.
6. REFERENCES [1] Mo Chen, Jessica Fridrich, Miroslav Goljan, and Jan Luk´as.Determining image origin and integrity using sensor noise.
IEEE Transactions on information forensics and security ,3(1):74–90, 2008.[2] Miroslav Goljan, Jessica Fridrich, and Mo Chen. Sensor noisecamera identification: Countering counter-forensics. In
MediaForensics and Security II , volume 7541, page 75410S. Interna-tional Society for Optics and Photonics, 2010.[3] Andreas Uhl and Yvonne H¨oller. Iris-sensor authenticationusing camera PRNU fingerprints. In
Proceedings of the 5thIAPR/IEEE International Conference on Biometrics (ICB’12) ,pages 1–8, New Delhi, India, March 2012.[4] Nathan Kalka, Nick Bartlow, Bojan Cukic, and Arun Ross. Apreliminary study on identifying sensors from iris images. In
Proceedings of the IEEE Conference on Computer Vision andPattern Recognition Workshops , pages 50–56, 2015.[5] D. S¨ollinger, B. Maser, and A. Uhl. PRNU-based finger veinsensor identification: On the effect of different sensor crop-pings. In ,pages 1–8, 2019.[6] L. Debiasi, Z. Sun, and A. Uhl. Generation of iris sensorPRNU fingerprints from uncorrelated data. In
Proceedings ofthe 2nd International Workshop on Biometrics and Forensics(IWBF’14) , pages 1–6, Valletta, Malta, 2014.[7] L. Debiasi and A. Uhl. Techniques for a forensic analysis of thecasia-iris v4 database. In
Proceedings of the 3rd InternationalWorkshop on Biometrics and Forensics (IWBF’15) , pages 1–8,Gjovik, Norway, 2015. [8] Luca Debiasi, Christof Kauba, and Andreas Uhl. Identifyingiris sensors from iris images.
Iris and Periocular BiometricRecognition , 5:359, 2017.[9] Susan El-Naggar and Arun Ross. Which dataset is this irisimage from? In , pages 1–6. IEEE, 2015.[10] Y. Yin, L. Liu, and X. Sun. SDUMLA-HMT: A MultimodalBiometric Database. In
The 6th Chinese Conference on Bio-metric Recognition (CCBR 2011) , volume 7098 of
SpringerLecture Notes on Computer Science , pages 260–268, 2011.[11] Ajay Kumar and Yingbo Zhou. Human identification us-ing finger images.
IEEE Transactions on Image Processing ,21(4):2228–2244, 2012.[12] Pedro Tome, Matthias Vanoni, and S´ebastien Marcel. On thevulnerability of finger vein recognition to spoofing. In
IEEEInternational Conference of the Biometrics Special InterestGroup (BIOSIG) , September 2014.[13] Yu Lu, Shan Juan Xie, Sook Yoon, Zhihui Wang, andDong Sun Park. An available database for the research of fin-ger vein recognition. In
Image and Signal Processing (CISP),2013 6th InternationalCongress on Image and Signal Process-ing (CISP 2013) , volume 1, pages 410–415. IEEE, 2013.[14] Christof Kauba, Bernhard Prommegger, and Andreas Uhl. Fo-cussing the beam - a new laser illumination based data set pro-viding insights to finger-vein recognition. In
Proceedings ofthe IEEE 9th International Conference on Biometrics: Theory,Applications, and Systems (BTAS2018) , pages 1–9, Los Ange-les, California, USA, 2018.[15] M. S. M. Asaari and B. A. Rosdi S. A. Suandi. Fusion of bandlimited phase only correlation and width centroid contour dis-tance for finger based biometrics.
Expert Systems with Appli-cations , 41(7):3367–3382, 2014.[16] B.T. Ton and R.N.J. Veldhuis. A high quality finger vascularpattern dataset collected using acustom designed capturing de-vice. In
International Conference on Biometrics, ICB 2013 .IEEE, 2013.[17] T. F. Chan and L. A. Vese. Active contours without edges.
IEEE Transactions on Image Processing , 10(2):266–277,2001.[18] Pascal Getreuer. Chan-vese segmentation.
Image ProcessingOn Line , 2:214–224, 2012.[19] A. V´ecsei, T. Fuhrmann, M. Liedlgruber, L. Brunauer,H. Payer, and A. Uhl. Automated classification of duodenalimagery in celiac disease using evolved fourier feature vectors.
Computer Methods and Programs in Biomedicine , 95:S68–S78, 2009.[20] Timo Ojala, Matti Pietik¨ainen, and David Harwood. A com-parative study of texture measures with classification based onfeatured distributions.
Pattern recognition , 29(1):51–59, 1996.[21] Franc¸ois Boulogne, Joshua D Warner, and EmmanuelleNeil Yager. scikit-image: Image processing in python. 2014.[22] Ron Kohavi et al. A study of cross-validation and bootstrap foraccuracy estimation and model selection. In
Ijcai , volume 14,pages 1137–1145. Montreal, Canada, 1995.[23] C´esar Ferri, Jos´e Hern´andez-Orallo, and Miguel Angel Salido.Volume under the roc surface for multi-class problems. In
European conference on machine learning , pages 108–120.Springer, 2003.24] Vincent Van Asch. Macro-and micro-averaged evaluation mea-sures.
Belgium: CLiPS , 49, 2013.[25] Jacob Benesty, Jingdong Chen, Yiteng Arden Huang, and Si-mon Doclo. Study of the wiener filter for noise reduction. In