Research On Fine-grained Expression Recognition Algorithm Based On Graph Representation And Label Relationship Mining

Expression is an important method to spread human emotional signals and coordinate interpersonal relationships.Expression recognition is a key research problem in the field of facial attribution recognition.Most of the existing expression recognition algorithms are based on six basic expressions,but the six basic expressions in the laboratory environment can not express the rich and diverse facial expressions in the natural environment.In recent years,the development of the Internet has brought a huge number of face image resources.Researchers began to turn to the research of fine-grained expression recognition.The application of fine-grained expression recognition in humancomputer interaction,security system,robot manufacturing,retrieval application,intelligent transportation and other fields shows a vigorous development trend.However,the difference of fine-grained expressions is much lower than that of coarse-grained expressions.For the fine-grained database with strict hierarchical relationship between coarse-grained labels and fine-grained labels,coarse-grained labels are not considered to guide fine-grained expression recognition.To solve the above problems,this thesis mainly carries out the following research work:(1)This thesis proposed a fine-grained expression recognition algorithm based on graph representation learning.Most of the existing algorithms focus on how to mine the global or local features of images.These methods ignore the influence of topological relationship at the image level,the semantic association of labels,and the mapping relationship between images and labels.The above relationship can explore more discriminative visual features from the perspective of image and label level,and alleviate the problem of low category discrimination between fine-grained expressions.Based on the graph representation learning method,this thesis constructs image subgraph,label subgraph and global graph respectively for the above relationship,and uses the local optimization method based on information entropy to update the global graph,so as to transform the fine-grained expression recognition problem into edge prediction problem.Furthermore,this thesis proposed a global graph data update model based on graph convolution network,which dynamically updates the node features and edge features of the above graph.The probability of the edge on the global graph is the classification probability.The experimental results on FG-Emotions data set show the effectiveness of the algorithm.(2)This thesis proposed a fine-grained expression recognition algorithm based on hierarchical label optimization.Most of the existing algorithms treat the fine-grained expression recognition task in isolation,and do not consider the possible constraints of the coarse-grained label that the fine-grained label belongs to improve the performance of the algorithm.In view of this,this thesis proposes a fine-grained expression recognition model based on hierarchical label optimization.The face image are input two branches at the same time: fine-grained expression classification branch and coarse-grained expression classification branch.In the fine-grained classification branch,the prior distribution of the hierarchical relationship between coarse-grained labels and finegrained labels is integrated into the construction of label subgraphs;In the coarse-grained classification branch,pretrain a classifier with excellent performance in coarse-grained expression,and output the confidence score of the coarse-grained class of the image.The loss function of the fine-grained classification branch is modified by the confidence score of the coarse-grained class,so as to adjust the output of the fine-grained expression recognition network based on graph representation learning,and guide the training direction of the network.Experiments show the effectiveness of the algorithm.