Micro-Expression Recognition With Spatiotemporal Feature Information Fusion

With the rapid development of computer technology and pattern recognition,facial expression recognition has been widely used in many fields.In recent years,researchers have made many achievements in the field of face recognition,and have begun to study a special facial expression which called micro-expression.Different from the general static facial expression recognition,the micro-expression recognition needs to extract the spatial feature information of the face table situation in each frame image,and also needs to consider the temporal motion information of the continuous image sequence.Currently,the research on micro-expression recognition approach is still at the stage of development,and there are still some unresolved or overlooked problems: 1)the features extracted by most of feature extraction algorithms have very high dimensions and contain much redundant information,which will affect the recognition performance;2)due to the weak behavior of the micro-expression itself,the feature extraction method of micro-expression recognition has poor robustness and poor recognition effect.In view of the above problems,this thesis studies and designs the feature extraction and optimization algorithm for micro-expression recognition.The main work is as follows:To handle information redundancy problems of the existing feature extraction methods,a microexpression recognition method based on color space information fusion(CSFL)is proposed.Firstly,the color information of the micro-expression color space channel is effectively merged,and the global and local information of the spatiotemporal features are combined to construct the close relationship of different color channel features,thus obtaining a new fusion feature.Compared with traditional features,the new fusion features have lower dimensions,less interference information,and more compact and effective features.Then,by comparing the features directly extracted on the gray space and the color space without information fusing,the support vector machine is used as the emotion recognition classifier to evaluate the recognition effect,verify and analyze the performance of the proposed micro-expression recognition method.The experimental results show that the new fusion feature can effectively utilize the color information and the micro-expression recognition information in the color space to enhance the performance of the space-time descriptor in processing the micro-expression recognition task,and the micro-expression recognition rate is significantly improved.To deal with the problem that the traditional feature extraction classification algorithm has poor robustness and recognition effect in the micro-expression recognition task,a micro-expression recognition method based on block sparse Bayesian learning(BSBL)is proposed.By exploiting the characteristics and correlation of the internal structure,the proposed method tries to reconstruct the sparse representation coefficient of the feature vectors.Compared with the traditional sparse representation classification methods,the coefficient of sparse representation solved by the proposed method can efficiently characterize the inherent difference information of the microexpression.Finally,we use a residual discriminant rule to test the category of the micro-expression samples.Experimental results show that our method can effectively improve the performance in terms of accuracy and robustness.