Sparse Representation for accurate classification of corrupted and occluded facial expressions

Abstract

Facial expression recognition remains a challenging problem especially when the face is partially corrupted or occluded. We propose using a new classification method, termed Sparse Representation based Classification (SRC), to accurately recognize expressions under these conditions. A test vector is representable as a linear combination of vectors from its own class and so its representation as a linear combination of all available training vectors is sparse. Efficient methods have been developed in the area of compressed sensing to recover this sparse representation. SRC gives state of the art performance on clean and noise corrupted images matching the recognition rate obtained using Gabor based features. When test images are occluded by square black blocks, SRC improves significantly on the performance obtained using Gabor features; SRC increases the recognition rate by 6.6% when the block occlusion length is 30 and by 11.2% when the block length is 40.