Research On Facial Expression Recognition Based On Improved Loss Function And Attention Network Algorithm

Facial expression is the most important non-verbal emotional expression of human,so facial expression recognition has always been a research hotspot.At present,the research on facial expression recognition has made great breakthroughs,but there is still a certain gap between real-time and accurate detection and recognition in real environments.Therefore,the research focus is shifted from the controlled facial expression in the laboratory environment to the uncontrolled facial expression in the real environment.Uncontrolled facial expressions have the problems of non-uniform illumination,diverse head poses,local occlusion,and individual differences,which make it difficult for convolutional neural networks to extract deep features with facial expression discrimination.This paper studies uncontrolled facial expression recognition for the purpose of improving the expression feature learning ability of convolutional neural network and improving the recognition effect of occluded expressions.1.In view of problems of large differences in the same type of uncontrolled facial expressions,small differences between different types and unbalanced distribution of data sets,a spatial distribution loss(SD loss)is proposed to improve the feature learning ability of convolutional neural networks.SD loss draws on the algorithm idea of Softmax loss to separate feature points to separate feature clusters,and introduces the k-nearest neighbor algorithm to use the average of the k nearest neighbor features of a feature point as the cluster center to achieve compactness within the cluster.The backbone network of facial expression recognition uses the attention mechanism ECA to make the model pay more attention to the feature areas related to expression recognition.Experiments were carried out on the facial expression dataset RAF-DB and Affect Net,and the recognition rates were 86.52% and 60.36%,respectively,which were better than most facial expression recognition methods.2.In view of the problem of partial occlusion of facial expressions in the real environment,thesis proposes a facial expression recognition method based on multi-region joint,that is,jointly extracts facial expression features by jointly using facial local regions and global regions.Drawing on the algorithm idea of the attention mechanism,the self-attention network is used to realize the weighted fusion of the depth features of each local region and the depth features of the global region,and a weight limit loss is proposed to balance the weight distribution of the depth features of each local region and the global expression features.In this thesis,the occluded facial expression images are collected from the uncontrolled facial expression dataset to establish the occluded facial expression dataset,and the ablation experiments are performed on the occluded facial expression dataset to prove the effectiveness of the proposed method.The experiments are carried out on RAF-DB and Affect Net.Prove the advanced nature of the proposed method.