Research On Facial Expression Recognition Method Based On Deep Learning

In recent years,facial expression recognition is a hot research direction in the field of computer vision.Facial expression is an important way for human beings to communicate and express emotions.By analyzing facial expressions,people’s emotional state can be inferred.Facial expression recognition has been widely used in various fields,but there are still many problems to be solved urgently.Facial expression data has obvious inter-class similarities(there are strong common features between different categories of facial expressions)and intra-class differences(due to the huge differences between different nations,cultures,races,and regions,for the same category There are obvious differences between expressions and images).In addition,most of the previous research on expression recognition was established in a controlled environment in the laboratory,which cannot reflect the complex scenes in real life and has certain limitations.In the real world,people’s expression changes are often spontaneous and unconstrained,and expression recognition is affected by external factors,such as different angles,occlusions,lighting changes,etc.Therefore,expression recognition in real-world scenarios can make the model more generalizable and have better practical significance.To solve this problem,this thesis proposes a facial expression recognition method based on feature enhancement and multi-head attention fusion,which includes three components: feature enhancement extraction network,multi-head attention network and attention fusion network.First,in the feature extraction part,considering the specificity of expression classification,the feature enhancement loss function is designed to make each type of feature gather near the feature center point as much as possible in the feature space,and impose edge restrictions on the classification boundary to minimize the class Maximize the inter-class distance while maximizing the intra-class distance,realize the separability and distinguishability of facial features,increase the inter-class difference,reduce the intraclass difference,and achieve the effect of feature enhancement.Secondly,the change of facial expression is not determined by a single location area,but the result of the combination of multiple key parts.Therefore,a multi-head attention mechanism is proposed to learn the regional correlation in expressions,and a fusion loss function is proposed to avoid overlapping attention areas..Finally,the expression category is output using an attention fusion network.In order to study a more realistic facial expression recognition model,this paper conducts experimental verification on the RAF-DB and Affect Net data sets based on realworld scenes,using accuracy as the evaluation index,and compares the performance of single attention and multi-head attention through ablation experiments.The experimental results under different forces,and the influence of different module compositions on the experimental results.The analysis shows that all the proposed network modules can effectively improve the recognition performance of the model.In the end,the methods achieved accuracy rates of 89.37% and 65.31%,respectively,which effectively improved expression recognition accuracy compared with several existing advanced models.