Research And Application Of Facial Expression Recognition Based On Lightweight Convolutional Network

Facial Expression Recognition(FER)technology is ascertained on the basis of human emotional face.Due to deep learning technology's growth,it has invited attention from researchers in recent years and relevant companies.Two key issues that affect the practical application of FER tasks are: the balance between speed and accuracy,and the quality of expression images.During the deep learning environment,the network model is too complex to be deployed in portable devices(mobile phones,embedded,etc.).Secondly,the face image trained by the network has complex challenges such as multi-pose,occlusion and rotation,which often make the features learned by the network redundant and cause misrecognition.Therefore,the research work of the thesis is as follows:1.To solve the balance's problem between precision and velocity in the FER task,the thesis puts forward a facial expression recognition approach that is based on a deep binary convolutional network.A lightweight convolution network(BRNet)composed of parallel working of conventional convolution and binary convolution is peculiarly devised to reduce network model's complexity.Secondly,a local binary pattern extraction technique that is based on dynamic radius taking a sample of strategy is propounded,and then an expression attention mechanism(EAM)is projected to rejoin network channels' weight,so that expression region's local feature is effectively fusible with face's global feature.Finally,L2 loss and cross-entropy are planned to reach the accurate classification.Experiments reveal that the suggested approach attains comparable accuracy to the state-of-arts when the recognition velocity gets to 29 milliseconds single-frame image,which goes beyond former CNN-based facial expression recognition approach marginally.2.Aiming at the problem of unbalanced number of facial expression samples and uneven quality of facial expression images,the thesis proposes a fusion equalization loss based on lightweight network for facial expression recognition.First off,the sample category's loss function is suggested to resolve the imbalance issue of expression database category samples,which is integrated into the network model by using a weighting factor.Secondly,in view of the variable quality of samples in the expression database,the loss function of sample quality is proposed.The image samples with good or bad expression quality are screened by the method of locating the key points in the expression region,and the loss function is incorporated by the way of weight influence.Then,making information representation of the network attention mechanism for use of the ample feature,a multidimensional attention loss function is devised,and the features that are formed by the two attention mechanisms are used as the cross entropy loss function's metric to identify the predicted value and the label value.Finally,the three losses are combined in the Kerasbased network model to form a new loss function---the fused equilibrium loss.The loss is integrated into the designed lightweight network model(XDNet)to perform the FER task,and the analysis approach and the experimental comparison proves to be the excellence of the algorithm.