Abstract:
Sparse coding based face recognition makes use of a dictionary having training face images as its columns. A test face image is represented as a sparse linear combination of these training images. This thesis establishes the significance of dictionary for sparse coding
based face recognition, which is relatively less explored in the related literature. In this work, we propose a new dictionary which is named as weighted decomposition of face (WD Face) dictionary. This dictionary is derived based on the assumption that a face image can be decomposed into three components as, (i) a common component, (ii) a noise component, and (iii) a component which contains the subject specific unique information of the person. WD Face dictionary is generated by giving higher weightage to the subject specific
components, which play a crucial role in identifying one person from the other. It also addresses the requirement of large number of training face images. The effect of illumination in computation of WD Face image is reduced using edginess based representation of
image, which is derived using one-dimensional (1-D) processing of image. 1-D processing of image provides multiple partial edge evidences, which are combined to enhance the face
recognition performance. Moreover, we experimentally show that the proposed dictionary ensures sparse representation of test image as a linear combination of training images even when the number of training images is small. Further, WD Face dictionary is incorporated to address the problem of pose variation in face recognition with the help of a two stage approach. The first stage makes use of a pose detection technique based on sparse coding to approximate the pose of an incoming test image and classify it in to one of the target pose classes. In the second stage, a WD Face dictionary is chosen corresponding to the detected pose and the identity of the person is
obtained using sparse coding based face recognition with the help of the chosen dictionary. Moreover, the problem of distant face recognition is addressed with the help of image super resolution and the proposed WD Face dictionary. The missing facial features from a test
image are restored using image super resolution. Then, the identity of this image is obtained by sparse coding based face recognition, which incorporates WD Face dictionary derived from high resolution training images.