Projects & Publications

     In this study a face detection method using correlation filter is proposed where temporal information in the video is fully utilized. That is, instead of detecting each frame, the proposed approach exploits temporal relationships between the frames to detect human face in a video sequence. A modified version  of UMACE filter is generalized to form a video or a 3D spatio-temporal volume, termed as unconstrained video filter (UVF). After correlating this UVF with the target video, probable location of face is detected according to the position of high correlation peak in a three dimensional plane. This location is the region of interest (ROI) in the target scene.Publication : Pradipta K. Banerjee, Jayanta K. Chandra, Asit K. Datta, A frequency domain face recognition technique based on correlation plane features as input to a regression neural network, Procedia Computer Science, Elsevier, 2,75-82,2010

Class specific subspace based nonlinear correlation filter

The linear filter formulation considers images having nonuniform dynamic range and hence in the testing stage it is hard to discriminate authentic and impostor images that lie below a span of low grey level. To overcome this situation this paper proposes nonlinear correlation filter by exploiting the point nonlinearites  of image pixels so that the designed correlation filter achieves a uniform dynamic range. This type of nonlinear mapping stretches pixel distribution of face images in a wide range and consequently high frequency components are amplified. In this study three approaches are judicially combined to improve face recognition results under illumination variation viz, i) projection based method of designing correlation filter is used to improve upon the capability of recognition at all possible illumination variations ii) phase correlation method is used to enhance peak sharpness at the correlation plane for authentic face image as phase contains more information than the magnitude of spectrum and iii) point nonlinearities are considered to extend uniform dynamic range. To achieve these two correlation filters are designed (a) nonlinear optimum projecting image correlation filter Hp and (b) nonlinear optimum reconstructed image correlation filter Hr . The nature of design process of these two filters is same with only difference in image used for synthesis. Design of Hp uses projecting image and the design of Hr includes reconstructed image. The phase correlation between these two filters produces a response surface, the nature of which totally depends on the face class involved. Ideally a delta type peak at the correlation plane is obtained if these two filters are generated from the same face class.

Publication : Pradipta K. Banerjee, Asit K. Datta, Class specific subspace dependent nonlinear correlation filtering for illumination tolerant face recognition, Pattern Recognition Letters, Elsevier, 36, 177-185, 15 January 2014

Optimized preferential correlation filter

1) A preferential filter is designed for multiclass face recognition, where FRR and FAR are improved by incorporating the information preferentially of both intra-class and other class face images for a given database during the synthesis of the correlator filter. To achieve this task, during the design of kth correlation filter, information of false clients are included. To achieve minimum FRR, minimization of average correlation energy and maximization of correlation peak intensity of Ck class face images are made simultaneously. However to achieve minimum FAR for a given database, minimization of average correlation energy with no maximization of peak intensity for face images of all other classes of a given database except the Ck class is considered. 2) To reduce both FAR and FRR the overall class compactness is reduced. Class compactness of both types of classes is made which makes the system more robust to distortion tolerance as well as misclassification. This is achieved through Euclidean distance measures which improves the performance. The class compactness with minimizing the class boundary of both true and false clients reduces the misclassificatin in terms of FAR. 3) Finally, the optimization of tradeoff parameters considered to design the preferential filter of both constrained and unconstrained type are carried out by PSO. A relationship between constrained and unconstrained optimum preferential filter is given. Results on standard face databases have established that the performance of proposed filter is better and more robust than the other existing classes of correlation filters.

Publication : Pradipta K. Banerjee, Asit K. Datta, A preferential digital-optical correlator optimized by Particle Swarm technique for multi class face recognition, Optics and Laser Technology, Elsevier, 50, 33-42, 2013

Regression network trained correlation filter

In most of the cases, it is noted that to improve the performance of correlation filters with respect to high verification accuracy during illumination invariant face recognition, some preprocessing are made before the synthesis of filters. A trained preprocessing method is evolved before synthesizing a correlation filter to show improved performance. The preprocessing aims to achieve the pattern association between same class images while the discrimination is amplified for other class images. The preprocessing stage involves the use of two convolution kernels. The first one is a contour kernel for emphasizing the high frequency components of the facial images so as to extract the edges of the significant facial parts like nose, mouth, eye and eyebrow. The other one is a smoothing kernel for the removal of unwanted high frequencies which may produce noise due to the earlier operation of contour kernel. The elements of these two convolution kernels are determined by supervised learning of a generalized regression neural network (GRNN) . The training of GRNN is based on the empirical selection of kernel elements with respect to face features which are obtained by dimension reduction and feature extraction technique using principal component analysis (PCA). In addition to this preprocessing technique, a modification over a standard correlation filter is also carried out.

Publication : Pradipta K. Banerjee, Asit K. Datta, Generalized Regression Neural Network Trained Preprocessing of Frequency Domain Correlation Filter for Improved Face Recognition and its Optical Implementation, Optics and Laser Technology , Elsevier, 45, 217-227, 2013

Landmark localization

The correlation plane, in response to a test image, in frequency domain is obtained by element-wise multiplication of Fourier transformed test image and the synthesized filter. In other way round it may be considered as if the correlation plane and the test image is known before hand, the filter F can be generated. In case of designing a landmark filter, in the training phase the position of left-eye, right-eye, nose-tip and mid-mouth is known. It is expected that after correlation operation between test image and landmark filter, the generated correlation plane G must show a distinct peak at the position of landmark due to shift- invariant property. Hence the design aspect of the landmark filter consists of a training set {xi , gi }, where xi is the training image and gi is the desired correlation output plane. This gi is generated synthetically by assuming a 2D Gaussian curve centred at position (r, c) of one of the landmarks. To detect the facial landmarks multi-correlation approach is done. Due to the shift invariant property of Fourier transform time complexity is reduced.

Publication : Papia Banerjee, Pradipta K. Banerjee, Asit K. Datta, Face Detection and Landmark Localization using Parametric Approach and Landmark Filters, The Eighth International conference on Advances in Pattern Recognition, IEEE computational Intelligent Society, Jan 4-7, 2015.

Class specific optimum filter