Sparse Representation For Face Expression Recognition Based On Dictionary Learning

Face facial expression, containing a large amount of human behavior information, has being one of the most common nonverbal way of human to express feelings. Along with the development of science and society, the relationship between the study, work and life of people and computer has become more closer, people urgently demand for human-computer interaction, therefore, facial expression recognition technology has become a research hotspot at home and abroad step by step, the method of sparse representation have got more and more attention and have become used for face facial expression recognition in recent years. This paper summarizes and discusses some of the key problems in facial expression recognition, introduces sparse representation based classification, maximum scatter difference discriminant criterion and cluster based dictionary learning is put forward, test samples are classified by minimizing the weighted reconstruction error, this paper mainly analyzes from the following aspects:(1) Facial expression feature extraction. The process of HOG feature extraction and LBP feature extraction are introduced in detail, the reason of choosing this two methods is represented.(2) Sparse representation based classification. The principle of SRC, dictionary learning, solving sparse coefficient are introduced, the key part of the sparse representation classification is dictionary learning. In order to verify the robustness to occlusion of sparse representation based classification, the experiments on the shining images and on the occlusion images are conducted, the experimental results show that sparse representation based classification can obtain good effect on the occlusion images.(3) The dictionary learning of sparse representation based classification. Due to the problems and insufficiency on the dictionary learning of SRC, dictionary learning based on maximum scatter difference discriminant criterion and cluster is proposed, the dictionary not only has good ability of representation and discrimination on the samples, but also takes low time on sparse representation coding. The experiments on the Cohn-Kanade face expression database are conducted, the experimental results show that the dictionary learning method proposed has good performance.(4) The feature fusion based expression classification. The edge and shape information of face images extracted by the HOG feature and the texture information of face images extracted by the LBP feature are complementary, so, the two features can be fused to obtain a new feature by canonical correlation analysis algorithm, the dictionary learning for sparse coding and classification is conducted on this new feature, the experimental results show that this algorithm can have good performance than single feature based expression classification.