Research On Heuristic Facial Expression Recognition Based On Deep Learning

Facial expression recognition technology is the core of human-computer interaction,which is widely used in intelligent companion robot,education,driving,medical care and other fields,so that it is a research hotspot in the field of brain cognition and intelligent science.With the development of computer vision,deep learning methods have achieved remarkable results in many face-related tasks.However,in the practical application process,there are still many problems.Firstly,the data quantity is insufficient and the non-equilibrium is serious,easily leading to over-fitting.Secondly,there is interaction between expression categories,which can easily lead to directional misclassification.Thirdly,the difficulty for recognizing the same expression in different environments is different,such as in the case of both normal lighting and dim lighting as well as individual’s age,gender,race and so on.It needs to solve the problem of the personalized expression recognition.Therefore,this paper mainly focuses on the above problems,making full use of the textual knowledge of facial expression and the relativity of human visual cognition as heuristic knowledge,and studies the heuristic facial expression recognition method based on deep learning.The specific work includes the following two aspects.(1)To enhance the differences between expressions across different categories,this paper defines the specific relationship between facial expressions and facial action units as heuristic domain knowledge.Through the heuristic domain knowledge,analysis of the mixed relationship between different expression categories,this paper puts forward a kind of based on facial expression and facial action unit heuristic objective loss function(HO Loss),which used to widen the distance between easy mixed classification,in order to learn more effective features by guiding the deep neural network,so as to improve the accuracy of facial expression recognition.Through extensive experimental validation,the heuristic objective function achieves 89.03%,89.03% and 64.02% recognition accuracy on RAF-DB,FER2013 Plus and Affect Net7 datasets,respectively,demonstrating the effectiveness and superiority of this method in facial expression recognition tasks.(2)Inspired by the relativity of human visual cognition,a heuristic feature augment method is proposed for facial expression recognition.It designs the augmenting features by relative transformation module(AFRT)and the augmenting features by graph convolutional neural network module(AFGCN)for the cognitive loss problem of deep model.Among them,the AFRT module first combines the relative transformation theory into the deep learning model to complete the facial expression recognition task,and innovatively constructs the heuristic explicit cognitive features through the relative transformation of the expression data.The AFGCN module uses the heuristic knowledge of the interaction between expressions,takes the class center as the graph node and the confusion matrix as the initial edge weight,to update learning through the graph convolutional network and infer the heuristic implicit cognitive features between expressions.In addition,these two heuristic feature augment modules are universal and can show superior performance in different deep models,especially the AFRT module can improve the network performance independently of the facial expression recognition task.Finally,extensive experiments on three public facial expression datasets,the results prove the effectiveness and universality of heuristic feature enhancement methods.