We propose an expression recognition algorithm based on asymmetric region Weber local descriptor (AR-WLD) and block similarity weighting, which can reduce the interference of occlusion area to facial expression recognition and the impacts of the final discriminants in unconstrained environment. In the feature description, compared with the traditional WLD, the AR-WLD extends the original square neighborhood into an asymmetric neighborhood, and enhances the feature analysis in a multiscale. In order to distinguish the contribution of different facial regions to expression recognition, the non-overlapping expression regions are classified in classification discrimination. Information entropy is introduced to measure the uncertain information contained in different sub-blocks, and the weight of similarity distance is defined according to the information amount. The facial expression discrimination is achieved by the block similarity weighted summation. The experimental results on the databases of JAFFE and CK show that the AR-WLD can effectively improve the classification performance and robustness of the WLD when the expression image is partially occluded, and the classification algorithm based on block similarity weighting can further reduce the interference of the occlusion area to facial expression recognition.