Debasis Mazumdar

Determination of Feature Hierarchy from Gabor and SIFT Features for Face Recognition

This paper shows how the most important features can be selected from the face so that the performance of any face recognition engine can be improved by matching only the maximally distinguishable features. Creating an automated face recognition system that can duplicate human performance in recognizing a face is one of the key goal of computer vision researchers. So, it is necessary that computational researchers should know the key findings from a facial image. Here the feature hierarchy in accordance with importance to recognize a face is used in our Face Recognition system and it is observed that the performance have been improved drastically after selecting the mostly contributing feature set.

Evolutionary-rough feature selection for face recognition

Elastic Bunch Graph Matching is a feature-based face recognition algorithm which has been used to determine facial attributes from an image. However the dimension of the feature vectors, in case of EBGM, is quite high. Feature selection is a useful preprocessing step for reducing dimensionality, removing irrelevant data, improving learning accuracy and enhancing output comprehensibility. In rough set theory reducts are the minimal subsets of attributes that are necessary and sufficient to represent a correct decision about classification. The high complexity of the problem has motivated investigators to apply various approximation techniques like the multi-objective GAs to find near optimal solutions for reducts. We present here an application of the evolutionary-rough feature selection algorithm to the face recognition problem. The input corresponds to biometric features, modeled as Gabor jets at each node of the EBGM. Reducts correspond to feature subsets of reduced cardinality, for efficiently discriminating between the faces. The whole process is optimized using MOGA. The simulation is performed on large number of Caucasian and Indian faces, using the FERET and CDAC databases. The merit of clustering and their optimality is determined using cluster validity indices. Successful retrieval of faces is also performed.

A Novel Data Mining Approach for Performance Improvement of EBGM Based Face Recognition Engine to Handle Large Database

Computerized human face recognition is a complex task of deformable pattern recognition. The principal source of complexities lies in the significant inter-class overlapping of faces due to the variations caused by different poses, illuminations, and expressions (PIE). Elastic Bunch Graph Matching (EBGM) is a feature-based face recognition algorithm which has been used fairly reliably to determine facial attributes from an image. It extracts the texture using Gabor wavelets around a set of biometric landmark points on a face, and generates a level graph. One of the degrading factor of the performance of EBGM based face recognition system is the size of the database, particularly when the database size is in tuned to millions, the performance of FRE falls drastically.In the present paper data mining approach is presented to improve the performance of the EBGM based face recognition engine in case of large database. We have proposed entropy based decision tree for feature selection and feature hierarchy. The selected features are taken to form suitable feature vector for Fuzzy C-means clustering. The clustered set becomes the reduced search space for the query face. Improvement in the performance of the EBGM based FRE is presented with suitable experimental results.

MRF Based LSB Steganalysis: A New Measure of Steganography Capacity

A new steganalysis algorithm is described based on the MRF model of image LSB plane. In this framework the limitation of the Cachins definition of the steganography capacity is quantified and a new measure is proposed.

A DOG filter model of the occurrence of Mach bands on spatial contrast discontinuities

The present work proposes a unified model to explain two previously reported properties of the Mach band illusion. The first is the frequently referenced fact that Mach bands are prominently visible at ramps, but practically vanish at intensity steps. The second property, less studied, on the other hand may also be related to the first. It concerns the fact that the width of the illusory Mach bands appears to be a function of the slope of the ramp itself. The model proposed here combines the difference of Gaussians (DOG) model of lateral inhibition in receptive fields with the models of feature detection, based on a holistic approach. The sharpness of discontinuity (SOD) concept for Mach band stimulus has been defined and is related to the slope of the ramp. It is suggested that calculation of SOD leads to an adaptive change in inhibitory surround, a notion that has the support of physiological experiments too.

A Pose Recovery Algorithm to Minimize the Effects of Pose Variation on the Matching Performance of Elastic Bunch Graph Based Multiview Face Recognition System

Multiview face recognition is a challenging task of deformable pattern recognition. Machine recognition of human from his/her facial images available at any pose, intensity and expression is the ultimate goal of multiview face recognition system. In the present work the effect of variation of poses on the overall performance of face recognition engine based on elastic bunch graph matching technique is studied using K-means clustering algorithm. A geometry-based pose recovery algorithm is proposed to improve the performance of the system

A chosen plaintext steganalysis of Hide4PGP v 2.0

A chosen plaintext steganalysis algorithm is described to isolate the corrupted bits in an image tampered with Hide4PGP V 2.0. The method is developed from the notion of representation of two dimensional image data in terms of a linear bit stream consisting of a set of basic building blocks. Its performance for message extraction is demonstrated on different 24 bit BMP images.

An adaptive scale Gaussian filter to explain White’s illusion from the viewpoint of lightness assimilation for a large range of variation in spatial frequency of the grating and aspect ratio of the targets

The variation between the actual and perceived lightness of a stimulus has strong dependency on its background, a phenomena commonly known as lightness induction in the literature of visual neuroscience and psychology. For instance, a gray patch may perceptually appear to be darker in a background while it looks brighter when the background is reversed. In the literature it is further reported that such variation can take place in two possible ways. In case of stimulus like the Simultaneous Brightness Contrast (SBC), the apparent lightness changes in the direction opposite to that of the background lightness, a phenomenon often referred to as lightness contrast, while in the others like neon colour spreading or checkerboard illusion it occurs opposite to that, and known as lightness assimilation. The White’s illusion is a typical one which according to many, does not completely conform to any of these two processes. This paper presents the result of quantification of the perceptual strength of the White’s illusion as a function of the width of the background square grating as well as the length of the gray patch. A linear filter model is further proposed to simulate the possible neurophysiological phenomena responsible for this particular visual experience. The model assumes that for the White’s illusion, where the edges are strong and quite a few, i.e., the spectrum is rich in high frequency components, the inhibitory surround in the classical Difference-of-Gaussians (DoG) filter gets suppressed, and the filter essentially reduces to an adaptive scale Gaussian kernel that brings about lightness assimilation. The linear filter model with a Gaussian kernel is used to simulate the White’s illusion phenomena with wide variation of spatial frequency of the background grating as well as the length of the gray patch. The appropriateness of the model is presented through simulation results, which are highly tuned to the present as well as earlier psychometric results.

Improvement and Estimation of Intensity of Facial Expression Recognition for Human-Computer Interaction

With the ubiquity of new information technology and media, more effective and friendly methods for human computer interaction (HCI) are being developed. The first step for any intelligent HCI system is face detection and one of the most friendly HCI systems is facial expression recognition. Although Facial Expression Recognition for HCI introduces the frontiers of vision-based interfaces for intelligent human computer interaction, very little has been explored for capturing one or more expressions from mixed expressions which are a mixture of two closely related expressions. This paper presents the idea of improving the recognition accuracy of one or more of the six prototypic expressions namely happiness, surprise, fear, disgust, sadness and anger from the mixture of two facial expressions. For this purpose a motion gradient based optical flow for muscle movement is computed between frames of a given video sequence. The computed optical flow is further used to generate feature vector as the signature of six basic prototypic expressions. Decision Tree generated rule base is used for clustering the feature vectors obtained in the video sequence and the result of clustering is used for recognition of expressions. Manhattan distance metric is used which captures the relative intensity of expressions for a given face present in a frame. Based on the score of intensity of each expression, degree of presence of each of the six basic facial expressions has been determined. With the introduction of Component Based Analysis which is basically computing the feature vectors on the proposed regions of interest on a face, considerable improvement has been noticed regarding recognition of one or more expressions. The results have been validated against human judgement.

Assisting the autistic with improved facial expression recognition from mixed expressions

People suffering from autism have difficulty with recognizing other peoples emotions and are therefore unable to react to it. Although there have been attempts aimed at developing a system for analyzing facial expressions for persons suffering from autism, very little has been explored for capturing one or more expressions from mixed expressions which are a mixture of two closely related expressions. This is essential for psychotherapeutic tool for analysis during counseling. This paper presents the idea of improving the recognition accuracy of one or more of the six prototypic expressions namely happiness, surprise, fear, disgust, sadness and anger from the mixture of two facial expressions. For this purpose a motion gradient based optical flow for muscle movement is computed between frames of a given video sequence. The computed optical flow is further used to generate feature vector as the signature of six basic prototypic expressions. Decision Tree generated rule base is used for clustering the feature vectors obtained in the video sequence and the result of clustering is used for recognition of expressions. The relative intensity of expressions for a given face present in a frame is measured. With the introduction of Component Based Analysis which is basically computing the feature vectors on the proposed regions of interest on a face, considerable improvement has been noticed regarding recognition of one or more expressions. The results have been validated against human judgement.