Expression invariant face recognition using semidecimated DWT, Patch-LDSMT, feature and score level fusion

This paper addresses the issue of human face recognition in presence of expression variations, which pose a great challenge to face recognition systems. Typically, the discriminant features lie in both spatial as well as transform domain. In this paper, we propose combination of Discrete Wavelet Transform (DWT) and proposed Semi-decimated Discrete Wavelet Transform (SDWT) to develop an expression invariant face recognition algorithm followed by a novel wavelet coefficients enhancement function. The wavelet coefficients are boosted using the proposed coefficients enhancement function and extracted using the Weber Local Descriptors (WLD). This enhances weak skin edges based features, resulting in increased probability of recognition. The proposed algorithm also exploits spatial domain features using our customized version of Complete Local binary patterns (CLBP) named Patch Local Difference Sign Magnitude Transform (Patch-LDSMT) applied on complete images and physiologically meaningful overlapping regions of human facial images for the first time. Feature level fusion of the wavelet based features and Patch-LDSMT yields a robust feature vector whose dimensionality is reduced using Linear Discriminant Analysis (LDA). Comprehensive experimentation is carried out on the JAFFE, CMU-AMP, ORL, Yale, Cohn-Kanade (CK) and database collected by us. Benchmarking analysis illustrates that the proposed face recognition algorithm offers much better rank one recognition performance when compared with the current state-of-the-art expression invariant face recognition approaches.