A modular weighted sparse representation based on Fisher discriminant and sparse residual for face recognition with occlusion

• We partition samples into quite small modules to capture the discriminative information.
• We identify a new weighting function based on the modular Fisher rate and residual.
• The method is based on sparse representation which makes the method robust.

Face recognition with occlusion is one of the main problems countered in face recognition in practical application. The occlusion in the image will decline the performance of global-based methods, so most of existing methods for this problem are block-based. Our method also divides image into modules. Considering that different modules have different discriminative information, we identify a new criterion to compute modular weight. The modular weight can not only depress the effect of low discriminant module but also can detect the occlusion module to some extent. The weighting function is based on the modular Fisher rate and the modular residual. The successful application of sparse representation-based classification (SRC) in image recognition inspires us to use SRC on the weighted dictionary and test image to perform the final identification. Experiments on the AR and extended Yale B database verify the effectiveness and robustness of the method.