Pixel-level occlusion detection based on sparse representation for face recognition

Abstract Occlusion is an obstacle for face recognition in practical application. Most of the existing methods solve the problem by performing representation on patches and giving small weights for the occlusion patches and high weights for the un-occlusion patches to decline the occlusion influence for the final recognition. However, the recognition rates of those methods will drop dramatically when some patches contain un-occlusion part as well as occlusion part, because the discriminant information of the un-occlusion part in those patches can not be extracted. To address this issue, a pixel-level occlusion detection based on sparse representation is proposed. Since sparse representation based classification (SRC) is applied for image recognition successfully, we use SRC to represent the query image and get the residuals of each class. And each pixel is estimated as occlusion or un-occlusion in each class’s residual. According to the estimation of each pixel in all the class’s residual, an integrated estimation of the occlusion is obtained. Considering that the occlusion is contiguous, we perform dilation to eliminate the isolated occlusion estimation pixels. The final recognition is performed on the un-occlusion part by SRC. Experiments on AR, Yale B and PIE database verify the effectiveness and robustness of the method.