Research

Pattern Recognition Research

Fingerprint Image Enhancement
The performance of any fingerprint recognizer depends upon the quality of the fingerprint image processed. Effective methodologies for cleaning the valleys between the ridge contours are relatively unexplored. We use a new composite filtering method by integrating an anisotropic filter and a directional median filter (DMF). The topology of DMFs and the length of the filtering windows are designed to follow the fingerprint ridge flow pattern for best results in an iterative algorithm. The new technique also allows mending of the broken ridges, the most common cause of artifacts and spurious minutiae.

Minutiae Detection
Most feature extraction algorithms described in the literature extract minutiae from a thinned skeleton image that is generated from a binarized fingerprint image. Thinning is a lossy and computationally expensive operation and the accuracy of the output skeletal representation varies for different algorithms. The thinning process also introduces artifacts such as ridge break and bridges that lead to spurious minutiae. To overcome these disadvantages we use a chaincode representation for detection of minutiae. The bifurcation points and endpoints of ridge contours and their orientation are readily extracted using a ridge contour following procedure.

Fingerprint Feature matching
Forensic applications of automatic fingerprint identification systems have to deal with partial and latent fingerprints. Previous methods have relied upon alignment of partial and whole prints by anchoring on singular ridge structures such as core and delta. Our research is about deriving secondary features based on relative positioning of the minutiae extracted from any partial print. These secondary features are orientation invariant and purely localized.

Signature Verification
We have developed a method of identifying the salient feature points in an off-line signature. The method can be readily adapted to on-line signatures as well. The minima and the maxima points are extracted and used to superimpose a grid on the signature image. Thus the grid enables automatic scaling of the signature and provides a reference structure for locating the features. We are also studying Regression Time Warping as a similarity measure of sequences in online signatures. This method outperforms the Dynamic Time Warping measure that is commonly used.

Cancelable Biometrics
Large-scale deployment of biometric systems raises concerns that go beyond ensuring security of transaction; it involves privacy of the original biometric data collected from the users. People have legitimate concerns about the use of their biometric data without their permission. This is exacerbated by the fact that biometric data (fingerprint or face) unlike passwords, if compromised cannot be changed. We propose to combine the fingerprint and signature data to construct a new cancelable biometric in a unique way to mitigate these issues.

Neural Networks for Object Classification
Artificial neural networks (ANNs) have their historical roots in biological systems. These highly interconnected feedforward and feedback networks can be used for many object recognition and classification tasks. We have demonstrated how a particular class, namely the analog version of the Adaptive Resonance Theory networks (ART2) can be used for hierarchical pattern classification. We are developing schemes for using this ART2 network for better identifying and subclassifying data.

Fingerprint scan is the most widely deployed biometric.