Biometric Applications
Prof. Vikram Gadre
Biometrics is a study of automated methods for uniquely recognizing humans based on one or more intrinsic physical or behavioural traits of a human, such as fingerprint, iris, face, ear, gait and so on.
Tremendous growth has taken place in automatic biometric based person authentication systems since the last four decades and today it has numerous applications ranging from person identification, forensics, border security and attendance system to large scale national program like Aadhaar, launched by the Government of India.
Research Areas :
– Smartphone based Touchless Fingerprint Recognition System.
– Latent fingerprint recognition.
– Unconstrained Ear Recognition.
– Iris Recognition: Monogenic wavelet for iris phase encoding.
– Deformation invariant classification of large scale fingerprint database.
The Center focuses on developing advanced signal processing methods in acquiring, processing and analyzing biometric information – especially latent fingerprint information, 2D fingerprint,iris, face and ear information. The technology so developed will lay the foundations for improved biometric identification with applications for security forces.
1. Smartphone based Touchless Fingerprint Identification System :
Conventional touch-based fingerprint systems have issues such as elastic distortion and non-uniform contact area which increases false nonmatching rate which is a serious issue in a negative biometric application like de-duplication. Touch-less fingerprint acquisition is an optimized solution to get rid of the nonlinear distortion problem, but simultaneously it offers new challenges like refection, complex background, and non-uniform illumination.
To get rid of the distortion problem of the touch-based system, we successfully developed a smartphone based touch-less fingerprint recognition system.
Some of the key features of the system are described below:
• Supports smartphone camera based touch-less fingerprint acquisition as well as conventional touch-based
scanner interface (via OTG cable) for the acquisition.
• Modes supported: User Enrolment, Verification and identification.
• A novel monogenic wavelet based touchless fingerprint enhancement algorithm using phase congruency features has been developed and used.
• Biometric templates are stored into a remote server for later matching purposes.
• During identification/verification, stored templates are retrieved instantly from the remote server database and matched with the acquired template.
TECHNOLOGY BRIEF :
It consists of four main blocks:
• Acquisition of finger image using a mobile camera,
• Region of interest detection and segmentation,
• Finger image enhancement using phase congruency in the monogenic wavelet framework, and
• Feature extraction and matching by widely used commercial Verifinger SDK in the case of fingerprints.
Current Status :
• The developed android application for touchless fingerprint identification system can be used to capture and enroll biometric template on the field and send it to remote server for matching instantly.
• Designed a novel monogenic wavelet based touch-less fingerprint enhancement algorithm using phase congruency features.
• We created a touch and touch-less based fingerprint database using the android app consisting of 400 touch and touch-less images of 50 subjects.
• The touch-less fingerprints enhanced using the proposed novel algorithm showed a comparative performance as that of conventional touch-based fingerprints in terms of identification accuracy
Planned Value Additions :
• Addition of face and ear biometric (touch-less) to the existing system.
• Fingerprint liveness detection to tackle spoofing.
• Automated localization and detection of fingerprint from the captured image.
Touch-less and Touch-based Fingerprint Database :
Currently there is no publicly available smartphone captured touchless fingerprint images and equivalent touch based fingerprint database. “IIT Bombay, Touchless and Touch Based Fingerprint Database” is a set of fingerprint data prepared by Indian Institute of Technology, Bombay, Mumbai, India.
Description :
• The dataset consist of 800 touchless fingerprint images of 200 subjects, 4 samples per subject having image size 170 x 260.
• It also consist of 800 touch based fingerprint images of the same 200 subjects having image size 260 x 330.
• The touchless fingerprints are captured using Lenovo Vibe k5 plus smartphone with the developed Android App.
• The images are captured using embedded flash for illumination. Touch based fingerprints are captured using eNBioScan-C1 (HFDU08) Scanner.
• The aim of preparing and sharing such a Database is to help researchers in their endeavors in comparing the performance of touchless and touch based fingerprint biometric systems.
Access request to the IIT-Bombay Touchless Fingerprint database must be directed to the following E-mails vmgadre@ee.iitb.ac.in. Applicants should manually fill, sign, scan and attach the application form to the given email address.Upon receipt of an executed copy of the signed application form, access instructions will be given.
Publication :
P. Birajadar , S. Gupta , P. Shirvalkar, V. Patidar , U. Sharma, A. Naik and V. Gadre, “Touch-less fingerphoto feature extraction, analysis and matching using monogenic wavelets,” 2016 International Conference on Signal and Information Processing (IConSIP), SGGS, Nanded , 2016.doi: 10.1109/ICONSIP.2016.7857436 Parmeshwar Birajadar, Meet Haria, Pranav Kulkarni, Shubham Gupta, Prasad Joshi, Brijesh Singh, Vikram Gadre, “Towards smartphone-based touchless fingerprint recognition”, SADHANA; volume 44, Article number: 161 (2019).
2. Latent Fingerprint Enhancement :
Latent fingerprints provide valuable information in forensic applications to identify criminals. In this center, the focus is on the research in the area of developing novel enhancement and matching algorithms which can maintain the identification accuracy.
The term ‘latent fingerprint’ refers to a partial impression of a fingerprint unintentionally left on any surface. Latent fingerprints on banknotes can provide important evidence for robbery and counterfeit investigations. The richness of partial information contained in latent fingerprints depends on the methodology used in lifting the fingerprint and the complexity of the background surface. Despite using sophisticated chemicalor dusting methods, the ridge structure can be corrupted while lifting it. Therefore, fingerprint enhancement is essential before extraction of minutia features for latent fingerprint matching.
The contextual Gabor filter bank based scheme is widely used for enhancement of live scan fingerprints which mainly utilize local orientation and frequency as a contextual information. Several approaches are proposed in the literature for orientation field estimation. But, in the case of latents, any of these methods will not work due to inherent noise present in latent fingerprints.
Wavelets have been extensively used in the literature for multi resolution pattern analysis and recognition. Our method of latent fingerprint orientation estimation is based on analytic signal representation. The term monogenic signal is used to represent an analytic signal in higher dimensions and is widely accepted by the research community in image processing for AM-FM modeling of images. We proposed a novel algorithm for multi resolution latent fingerprint orientation estimation using the monogenic wavelet and multi resolution orientation field. Out of several existing isotropic wavelets, we have used the Simoncelli isotropic wavelet in order to estimate latent fingerprint orientation. The block diagram of our proposed algorithm for latent fingerprint enhancement for banknotes is as shown below:
It mainly consists of two stages:
1. Monogenic wavelet based multi resolution orientation estimation along with related coherency map and
2. Coarse orientation estimation by combining different layers of the multi resolution pyramid.
Since there was no public database of latent fingerprints on banknotes for research, we have also created a database of 200 images of 25 subjects with 8 images per subject We have compared the result of our proposed enhancement algorithm with that of the Short Time Fourier Transform (STFT) method. The marked areas of the latent fingerprint inside the blue box obtained by manual segmentation clearly show that our algorithm outperforms the STFT method.
The reason behind the failure of2STFT method is that it is a block FFT based method, which finds orientation based on the blockenergy with the underlying assumption that the block contains fingerprint ridges. But unfortunately,the background of the banknote contains texture strongly similar to a fingerprint. Therefore, it mayextract orientation of the background instead of the fingerprint. But, in our case, multiresolutionorientation estimation helps us to extract orientation information correctly which is of a globalnature. The Fingerprint Orientation Model based on Fourier Expansion (FOMFE) method is a global model which will not fit well in different local areas of banknote background.
Publication :
P. Birajadar, V. Patidar, P. Shirvalkar, S. Gupta and V. Gadre, “Enhancement of latent fingerprints on banknotes using monogenic wavelets,” 2016 International Conference on Signal Processing and Communications (SPCOM), IISc ,Bangalore, 2016. doi: 10.1109/SPCOM.2016.7746603
3. IRIS Recognition using Iris phase encoding :
On account of the extremely unique structural features possessed by the human iris, this biometric has received substantial attention from the signal processing research community. The widely accepted and industry implemented algorithm for the purposes of identification and verification of an individual on the basis of his/her iris is due to John Daugman,a Professor of Information theory and Computer vision from the Cambridge University of England. It is outlined in a block diagram as below:
Daugman has used the Gabor wavelets in his algorithm to encode the features of iris, successfully. In general, the structural features of an image are described by the phase of the image. Gabor wavelets, being complex functions, are able to give such a phase description. In fact, any transform being used at the place of Gabor wavelets will try to define its own phase description suchthat it extracts as much structural and geometric information from the image as possible. In the 1D scenario, the analytic signal
representation of a real signal is able to capture the instantaneous phase of the signal which proves to be very useful in describing the signal behavior. Among the multiple extensions of the analytic signal concept to2D, proposed by different researchers, the monogenic signal representation using Riesz transform due to M. Felsberg from Kiel University, Germany promises to be the closestto the most faithful extension of the analytic signal concept to 2D. The monogenic signal also gives the local orientation of the signal in addition to the phase. Daugman used multi scale Gabor phase with multiple orientations to characterize the iris features. The monogenic wavelet framework is a new approach for iris encoding using the local phase vector. The monogenic wavelet analysis yields a local phase vector which contains not only the local phase but also the orientation information of the iris. This enables the evaluation of the structure and geometric information simultaneously. In the case of Gabor phase, in a specific scale, adapting the wavelet to different orientations involves the use of multiple filters. On the other hand, in the case of the monogenicwavelet, the two Riesz components readily give access to all the orientations
Publication :
P. Birajadar , P. Shirvalkar, S. Gupta , V. Patidar , U. Sharma, A. Naik and V. Gadre “A novel iris recognition technique using monogenic wavelet phase encoding,” 2016 International Conference on Signal and Information Processing (IConSIP), SGGS, Nanded , 2016. doi: 10.1109/ICONSIP.2016.7857494
4. Unconstrained Ear Recognition :
Ear as a biometric modality has various application possibilities in areas such as forensics, security, and surveillance. The advantage of using an ear is that the ear image acquisition process is contactless and non-intrusive. It is one of the stable biometric traits which has a unique shape, structure, and appearance that differs from person to person. Ear can also be used alongside other biometric modalities (such as face) in multimodal biometric systems and provide identity cues even when the other information is unreliable or unavailable. One direct application being person identification from CCTV footages, where even when the face is not properly captured, the ear can serve as a source of information to identify the person.
Many automated ear recognition approaches have been proposed over the last two decades in the literature. Most of these approaches use hand-crafted features and work best only for ear images taken under constrained settings. The performance of these techniques degrades when the ear images are captured under varied illumination conditions and contain various pose and occlusions. But, for practical applications, such as video surveillance and forensic applications, the ear images are captured in an unconstrained environment and are very challenging for feature extraction and identification due to large intra-class variability.
With the advancement of AI and Deep Learning, various CNN based approaches have been proposed recently which outperformed the previous techniques which uses only handcrafted features. But it still needs further improvement. We explored various deep learning techniques and a Scattering Wavelet Network (ScatNet)based local feature extraction technique and use a fusion of learned and handcrafted features for the task of unconstrained ear recognition. Additionally, we use center loss and domain adaptation techniques while training the deep CNN networks which help in boosting the accuracy. The experiments evaluated on AWE and UERC ear databases, show the effectiveness and robustness of the proposed feature extraction technique in terms of Rank-1 (R-1) & Rank-5 (R-5) accuracy.
Publications :
Birajdar P., Haria M., Sangodkar S., Gadre, “Unconstrained Ear Recognition Using Deep Scattering Wavelet Network”, IEEE IBSSC (2019) (Accepted)
5. Deformation Invariant Fingerprint Classification :
In a large-scale Automated Fingerprint Identification System (AFIS), fingerprint classification is an essential indexing step to reduce the search time in a large database for accurate matching.
Fingerprint classification is still a challenging machine learning problem due to a large intra-class and a small inter-class variability. Non linear elastic deformation is one of the main sources of intra-class variability which occurs due to the non-uniform pressure applied during fingerprint acquisition and the elastic nature of fingerprint itself.
We have developed a novel algorithm for fingerprint classification based on scattering wavelet network to extract translation and small deformation invariant local features. In texture-based classification problems, Gabor filter based features are widely used, but invariance to deformation is not considered in such approaches. With the spirit of deep convolutional networks, Mallat introduced Scattering Wavelet Network (ScatNet) by using a cascade of the wavelet transform and a modulus operator to build representations of signals which are stable to deformations, besides being translation invariant. It has also been shown that the features extracted using ScatNets give promising results in the context of texture classification.
The main properties of the ScatNet are local translation invariance, small deformation invariance, preservation of signal energy, and retention of higher order information. The most important advantage is that unlike convolutional neural networks, ScatNets are deep networks with prefixed weights and structure. Hence, no training is required in the setting of the hyper-parameters (such as no. of layers, size of hidden layers, filters, etc.) which takes a lot of experimentation and expertise.
Even though the Scatnet has been successful for classification of other textural databases, in the case of fingerprint classification, building an invariant representation using Scatnet is a challenging task because the useful classification information (oriented texture)is located in different regions of the fingerprint with small intra-class variability. Hence, we preferred the block-based approach for local feature extraction. Dividing the region into 16 blocks linearizes the deformation to some extent, but in order to build a complete local deformation invariant model, ScatNet features are extracted for each of these blocks.
Thus, the block wise Scatnet approach is useful in two ways. First, the oriental ridge-furrow structural information gets divided into local blocks, and second, a small deformation present in respective blocks can easily be linearized by means of scattering features.
The approach consists of following steps :
• Determine the registration point using R92 algorithm and perform fingerprint enhancement.
• Crop the Region Of Interest (ROI) around registration point and perform square tessellation of ROI.
• Apply second-order scattering wavelet transform using oriented Morlet wavelet filter bank on each local block.
• Calculate the mean of the scattering coefficients for each block of the ROI and obtain the final feature vector by concatenating the mean feature for all the blocks.
• Feed the feature vector to a trained PCA a ne classifier to perform fingerprint classification.