Learning features from covariance matrix of gabor wavelet for face recognition under adverse conditions

Face recognition under adverse conditions, such as low-resolution, difficult illumination, blur and noise remains a challenging task. Among existing face recognition methods, Gabor wavelet plays a significant role and has robust performance under adverse conditions since it models the visual cortices of mammalian brain. It has been demonstrated the subbands of Gabor Wavelet (GW) can be efficiently represented by a covariance matrix. However, because covariance matrix does not belong to Euclidean space, Euclidean-based measure such as 2-norm cannot be directly applied to covariance matrix, and more importantly, it is difficult to incorporate learning techniques for the covariance matrix to promote the performance of face recognition. To address this issue, we propose two promising methods by learning the Covariance Matrix of Gabor Wavelet (LCMoG). The first method, called LCMoG-CNN, uses a shallow Convolutional Neural Network (CNN) to project the covariance matrices of GW into a feature vector of Euclidean space; the second method, called LCMoG-LWPZ, uses matrix-logarithm to embed the covariance matrix in the linear space and then uses Whitening Principal Component Analysis (WPCA) to learn the face features from the embedded covariance matrix. The proposed methods are effective to extract the fine features from the face image and have better performance than Deep CNN (DCNN) for small-varying face pose . For the large-varying face pose , LCMoG features combining with DCNN feature can enhance the performance of face recognition. In the experiments, the proposed methods yield promising recognition & verification accuracies under adverse conditions. (c) 2021 Elsevier Ltd. All rights reserved.

Automated summary

This study proposes two promising methods (LCMoG) for improving face recognition performance by learning the covariance matrix of Gabor wavelet. One method uses a shallow Convolutional Neural Network to project covariance matrices into Euclidean space, while the other method embeds the covariance matrix in linear space using matrix logarithm and learns face features through Whitening Principal Component Analysis.