Research And Implementation Of Micro-expression Recognition Algorithm

Micro-expression recognition theory breaks through the limitations of common face recognition theory.The main trend is to use dynamic methods to capture local information.Micro-expressions are not only short-lived,but also have low intensity,reflecting the emotions people try to hide.Well-trained human experts can only achieve low detection rates,while automated micro-expression recognition systems can achieve higher detection rates.Due to their miniaturization and localization,most methods in recent years have extracted local and dynamic information from high frequency video for detection and classification.Micro-expression analysis consists of four steps: face detection,face positioning,feature extraction and expression recognition.The focus is on face positioning and feature extraction,and once the face is detected,the reference point is found.In many facial-expression recognitions,this step is necessary.The feature extraction process consists of two aspects,pre-designed and learned.Pre-design is to manually extract relevant information,and learning is automatically learned from the training data.The problems existing in some micro-expression recognition algorithms are studied.Based on the traditional micro-expression recognition algorithm,an improved micro-expression recognition algorithm is proposed.The purpose is to improve the shortcomings of the existing algorithms.The specific research work has the following aspects:(1)SRC algorithm is a more effective algorithm in micro-expression recognition algorithm.It uses sparse expression and dictionary learning technology to perform micro-expression recognition.It directly uses the whole set of training samples as a sparse-coded dictionary,learning the dictionary from the training samples instead of using pre-define a dictionary to learn,which produces the most effective results.However,the algorithm has a premise that the losses caused by all misclassifications are the same.However,in some micro-expressions,different error classifications may cause different losses,so the algorithm has certain limitations.In order to solve this problem,this theme proposes a cost-sensitive sparse expression based on the SRC algorithm,namely CS-SRC algorithm.The dictionary designed by this algorithm can generate cost-sensitive sparse coding,thus improving classification performance in this case.It mainly introduces a new "cost" penalty matrix,and enforces cost sensitivity throughout the learning process,using alternative optimization methods to effectively obtain the optimal solution.(2)LBP algorithm can extract micro-expression texture features,but it has the problem of image temporal,spatial variation and redundant difference.In order to solve this problem,an improved algorithm is proposed to generate intersection lines by intersectingthree orthogonal planes.The point of intersection is a point in the space-time domain,removing redundant intersections,providing a more compact and significant expression,resulting in less computational complexity.In addition,the algorithm achieves significant improvement in recognition accuracy and computational complexity by performing experiments on micro-expression images with different resolutions.(3)A robust principal component analysis(RPCA)is proposed to extract subtle micro-expression motion information.The improved EOH algorithm and BGC algorithm are used to extract local texture features,which can solve the problem about spatio-temporal domain of micro-expression sequences and obtain higher recognition accuracy.Principal component analysis is a method of data analysis that largely reduces high-dimensional data to low dimensions.