Research On Facial Expression Recognition Method Based On Two-stage Self-curing Network

The human face can convey rich emotional information,and in daily life,the emotional state conveyed by facial expressions is more straightforward than words.At present,facial expression recognition technology has been widely used in many aspects of society,such as companion robots,intelligent security,and assisted treatment robots.In particular,expression recognition can be applied in rehabilitating autistic children.The human-machine interaction of the rehabilitation robot can guide the child’s emotions,record the child’s various expressions during the human-machine interaction,analyze the child’s emotional state and inner activity,and assist the doctor in exploring the relevant pathological mechanisms.The rehabilitation robot based on expression recognition should be accurate for negative emotions to assist medical treatment better.However,the complexity of facial expressions and the similarity between expressions make studying expression recognition algorithms challenging.This topic focuses on the problem of the low recognition rate of negative emotions in facial expression recognition.Three main challenges are summarized by analyzing the research connotation behind the problem.(1)The facial expression dataset labels are manually labeled,so the labels in the dataset originate from the subjective judgment of the labelers.However,the subjective differences of the labelers and the ambiguity of the expressions themselves,the problem of mislabeled expressions have seriously impacted the training of deep learning.(2)Different individuals express the same expression differently,while the same individuals show similar expressions for different expressions.Such complex expressions pose challenges for expression classification and cause overfitting problems for deep learning models.(3)The existing large field expression datasets are collected from social networks,and most of the images shared in the network are positive emotions.It is challenging to collect negative expressions,which leads to the problem of unbalanced category distribution in the training of deep learning models.This paper proposes a recognition algorithm for static face expressions to address the above problems based on a deep learning framework.The main innovation points and research contents of the paper are as follows.(1)To address the problem that large expression recognition datasets are mislabeled and thus lead to poor recognition accuracy and robustness of expression recognition models,this paper proposes a self-curing network that can correct the mislabeling.The self-curing network algorithm aims to find the suspected incorrectly labeled samples and change them to have more accurate sample labels in the following training round.(2)To address the difficulty of classification caused by too similar expressions among existing expression recognition algorithms.In this paper,we propose a fusion algorithm of expression features and gender features.The fusion algorithm enables the expression recognition network to learn the relationship between gender and expressions adaptively and strengthen the influence of gender features on expression recognition.(3)To address the problem of uneven distribution of each expression category in the training dataset,which leads to a low overall expression recognition rate.In this paper,we propose a distribution-irrelevant loss function,which optimizes the expression recognition model with the traditional loss function to improve the attention to the minority class by considering the difference in the learned features between the minority and majority classes.