Expression Recognition Based On Global And Local Feature Dictionary Reconstruction Residuals

With the development of computer technology,the application of facial expression recognition technology is also expanding.At present,the research of facial expression recognition algorithm has become an important branch of artificial intelligence.The research and application of this technology make human-computer interaction more intelligent.One of the most important problems in facial expression recognition is that the features extracted from facial images will inevitably contain facial identity information.Therefore,this paper aims to study how to highlight expression information in feature extraction.This paper will start from two aspects of global and local features,the specific contents are as follows:(1)To improve the local feature method.We use the word bag model as the basis to extract local information.Firstly,the improved LBP algorithm is used to extract the local features of the prominent expression,the original pixel gradient information is added to the feature,we can get both the expression features and the position information of the expression features.Then,the non-negative matrix decomposition algorithm under sparse correlation constraints is used to reduce the dimension.According to the sparse constraints,the features after dimension reduction are concise.According to the correlation constraints,the features after dimension reduction can retain more similarity,which is convenient for the subsequent construction of dictionaries and codes.(2)The global feature method is improved.Due to the interference of face identity information in the direct extraction of image features,on the one hand,DWT algorithm is used to obtain the expression image of the face,which is fused with the overall facial features to obtain the expression weighted global information;on the other hand,mutual information is used to measure the similarity between the expression images,which is used to construct the adjacency matrix in the manifold dimension reduction,so as to make DeWitt In the feature,the geometric structure of expression information is preserved,the identity information of face is weakened,and make the obtained features more targeted.(3)The expression can not be fully represented by a single feature,so the local feature algorithm and the global feature algorithm mentioned above are fused in this paper.In order to avoid the small sample crisis caused by the direct fusion of local and global features,the reconstruction residuals of the two are fused at the decision level.In addition,the verification set is used to determine the weights in the process of feature fusion adaptively,so as to reduce the uncertainty of manual setting.According to the comparison and analysis of the experimental results,the three improved methods proposed in this paper can effectively improve the recognition rate.Therefore,the method proposed in this paper is feasible.