Facial Expression Recognition Based On Asymmetric Region Local Gradient Coding And Multi-feature Fusion

Facial expression recognition technology is one of the hot and difficult topics in the field of computer vision and pattern recognition. It has an important theoretical value and commercial significance. In recent years, facial expression recognition attracts a large number of scholars and research institutions engaged in the related research. This thesis analyzes and researches the feature extraction of facial expression recognition, improves the traditional texture feature extraction method, also presents an effective feature fusion method according to the characteristics of different feature. The main work is as follows:(1) In order to overcome the deficiency that Local Gradient Coding only extract texture feature in the neighborhood of a fixed size, this thesis proposes a novel multi-scale Asymmetric Region Local Gradient Coding fusion method for feature extraction of facial expression. Firstly, the normalized face image is preprocessed by Gaussian filter to reduce the impact of noise. Secondly, the preprocessed expression image is divided into several blocks. For each pixel of each sub-block image, Multiple Asymmetric Region Local Gradient Coding operators of different sizes are used to extract feature and obtain two binary sequences. These two binary sequences are fused into a new binary sequence according to the logical XOR. Thirdly, the new binary sequence is encoded, each sub-block histogram distribution is counted and all sub-block histograms are cascaded to get texture feature of facial expression. Finally, the process of expression classification is completed by using SVM. By fusing the intensity between neighborhood of different gradients and different scales, the multi-scale Asymmetric Region Local Gradient Coding fusion method combined with the block histograms can perform well in local feature description and global feature description. Experiments are performed on the JAFFE database and CK database. Experiment results show that the proposed method is better than usual typical feature extraction algorithms.(2) In order to overcome the one-sidedness and limitations of a single feature subspace for facial expression classification, this thesis proposes a feature fusion method. Firstly, the proposed method calculates the diversity between candidate feature and current feature subset, and computes the importance of candidate feature. Then by calculating the complementarity degree according to the diversity and the importance, features can be dynamically selected from candidate feature set to construct the optimal feature subset. Secondly, by using Particle Swarm Optimization, each feature of optimal feature subset is assigned with appropriate weight to construct weighted characteristics. Finally, the SVM is used for expression classification. Experiment results on JAFFE database and CK database show that the proposed method can improve the recognition rate and reliability.