Research On Face Expression Recognition Method Based On Residual Network

As one of the main external expressions of emotion,facial expressions also show diverse characteristics in the increasingly complex social relationships.Facial expression recognition technology is an important way to analyze human emotional states,understand individual emotions and intentions,and has shown broad application prospects in various fields such as medical assisted diagnosis,online teaching evaluation,and fatigue driving detection.Therefore,facial expression recognition technology,which involves cross-fertilization of multiple disciplines,has received extensive attention from scholars while the field of artificial intelligence is developing rapidly.In real scenarios,due to the influence of individualized head posture,wearing habits and external environment,the problem of insufficient information on the discriminative basis of facial expression recognition,which leads to the decrease of recognition accuracy,is a research difficulty that needs to be solved urgently.In recent years,a large number of efficient algorithms have emerged in the field of deep learning,and their advantages of automatic and fast learning of expression features make them the mainstream research methods in the field of expression recognition.After analyzing the difficulties and commonly used research methods in current facial expression recognition research,the residual convolutional neural network and image processing theory are used as the research foundation,and the facial expression recognition task is studied in the paper.In order to improve the feature extraction ability and feature utilization of the network,the residual network is optimized from two perspectives of feature extraction and loss function,the improvement methods are proposed and verified the superiority on the publicly available facial expression datasets.The main research completed is specified as follows:(1)A facial expression recognition method(AA-Net)combining anti-aliasing residual network and attention mechanism is studied to address the problems that it is difficult to extract facial expression features effectively in natural environment and the high similarity between expression categories leads to low accuracy of facial expression recognition.The method first improves the feature extraction ability of expression images by constructing an anti-aliasing residual network to minimize the loss of key contour feature information of facial expressions.Then a channel attention module with residual connectivity is introduced to give higher weights to the important information in the key regions of facial expressions and promote the network to focus on the expression features with higher differentiation.Finally,label-smoothing cross entropy loss is used to correct the overly absolute prediction probability by increasing the amount of information in decision expression categories.Through experimental validation,the proposed method effectively improves the facial expression recognition accuracy at the cost of increasing small number of parameters and computation,achieving 88.14% and 89.31% recognition accuracy on two datasets,RAF-DB and FERPlus,respectively,and partially occluded facial expression datasets are established through image preprocessing,which are named Occlusion-eyes-CK+,Occlusion-mouth-CK+,Occlusion-eyes-JAFFE,and Occlusion-mouth-JAFFE,respectively.It is verified that the AA-Net method has better robustness compared with the original network on the above occlusion expression datasets.(2)A facial expression recognition method(CRAP-Net)combining channel representation loss function and average prediction strategy is studied to address the problem of underutilization of features in single residual network leading to recognition performance bottleneck on facial expression recognition task.According to the characteristics of facial expression,the channel representation loss function adapted to the task of facial expression classification is redesigned to optimize network parameters in this method,i.e.a new constraint mutual-channel loss function is introduced based on label-smoothing cross entropy loss function.Channel representation loss function is designed to establish a mapping between high-level semantic feature channels and expression categories,to enhance the perception of strong discriminant information by convolution neural network,so as to increase the difference between expression categories and improve the accuracy of facial expression recognition.Average prediction strategy enlarges the fine-grained feature of the emoticon image,generates multiple local regions by clipping them at fixed positions,and aggregates them into a whole feature representation at the end of the network,which enables the network model to achieve better recognition performance in the end-to-end training process,thereby enhancing the robustness of the network model to the expression recognition when the face is incomplete.The above method achieves 88.66%,89.60% and 70.75%recognition accuracy on the two datasets RAF-DB,FERPlus and the existing publicly available occluded expression dataset FED-RO,respectively,and achieves 85.17%,88.53%and 88.71% recognition accuracy on the real occluded dataset Occlusion＿RAF-DB,and the pose change datasets Pose＿RAF-DB(>30),Pose＿RAF-DB(>45),respectively.Finally,the feasibility of the method is verified by using the above method to study facial expression recognition on images and videos in real scenes.