Research On Expression Recognition Of Unconstrained Environment Based On Deep Semantics

Facial expression contains rich emotional information,which is an important form of human emotional expression.With the rapid development of image processing and machine vision(CV)technology,more and more attention has been paid to the research of computer vision based on affective computing and psychoanalysis.Facial expression recognition has a broad application prospect in online education,intelligent business,public security and other fields.The current facial expression recognition technology under the laboratory constraint environment has been relatively mature,while in practical application,the facial expression will be affected by light,occlusion and other factors which will reduce the recognition accuracy.Therefore,facial expression recognition under unconstrained environment is becoming a research hotspot of CV due to the growing demand for expression recognition accuracy in natural scene.At present,the research of facial expression recognition under unconstrained environment mainly focuses on three kinds of problems: posture,occlusion and illumination.There is no unified expression recognition model that can solve the above-mentioned problems at the same time.In this paper,based on deep learning,two general models of facial expression recognition under unconstrained environment are proposed.By enhancing the representation learning and generalization ability of the model,the problem of low recognition accuracy caused by posture,occlusion and illumination is solved.The specific research work and innovation of this paper are as follows:(1)A Feature Fusion Attention Bilinear Pooling Model(FFABP)is proposed.By introducing attention mechanism to the bilinear pooling model,a novel attention enhanced bilinear pooling model is put forward in this paper,which uses bilinear pooling operation in one hand to calculate the second-order statistics of feature space to capture the subtle local differences between expressions and uses attention mechanism in the other hand to highlight the important spatial positions in high-dimensional feature spatial to enhance the representational learning ability of the model.At the same time,the self-attention model is utilized to learn the dependence between the features of different regions and obtain the first-order global geometric features.The first-order global geometry feature and the second-order local feature are fused together to further enhance the model's representation learning ability.(2)An Adaptive Cascade Count Sketch Model(ACCS)is presented.By improving the Count Sketch algorithm,this paper presents a Multi-order Count Sketch algorithm,and uses it to build a Cascade Count Sketch model to capture more high-order features,and effectively fuses the features of each level step by step.In order to adapt to the differences of different data sets and improve the generalization ability of the model,an adaptive attention model is designed to automatically track the weight of each level of features and adaptively retain important high-order features.In the meantime,a filter module is designed and added to the model for the purpose of selecting some Cascade Count Sketch branches to calculate the corresponding high-order features according to the important feature indexes retained by the adaptive attention model during the test phase,which effectively reduces the amount of calculation required by the model.