Research And Application Of Occluded Dynamic Expression Recognition Based On Multi-network

In recent years,facial expression recognition has broad research and application prospects in AI service equipment,virtual network communication,medical treatment,driver fatigue monitoring,psychology and many other fields.But in real life,the changes of light,shadow,posture and so on will produce certain occlusion to the facial expression,so the facial expression will lose some important information,which will lead to the facial expression recognition is not robust enough.In addition,we only recognize the expression of the image,ignoring the real dynamic expression.Therefore,this paper optimizes the generation of confrontation network structure and the best cascade of convolution network to improve the rate.Work as follows:1.A generative confrontation network structure of facial expression occlusion completion is proposed,and a parallel network structure P-IncepNet(para perception)is constructed for context feature extraction,the conditional countermeasure network is used to complete different degrees of occlusion area,which the G model width are increased,and the new partial occlusion completion network can carry out more perfect feature learning.The network has parallel inside and outside the inception structure,different convolution kernels are used to learn information of different space sizes,and 1 × 1convolution reduces the depth of feature map,and many different feature map outputs are fused on the channel.The output of different feature graphs are fused on the channels.Experiments on Celeb A and MMI datasets show that compared with the traditional generation confrontation network,the proposed network structure has better occlusion repair effect and more stable network model.2.The cascade network structure of P-IncepNet and LSTM is constructed for dynamic expression recognition,so that each module dealing with different tasks is superimposed in turn to form a deeper network.Through this structure,the factors unrelated to feature representation can be gradually filtered out.The feature representation of facial expression image is extracted by P-IncepNet model,and then these features are input into loop network LSTM to enhance temporal information coding.One hidden layer of LSTM is designed as128 embedded nodes.In addition,dropout layer is added at the top of LSTM hidden layer to prevent over fitting.Compared with c3d-lstm hierarchical network,the results show that the network structure not only has excellent performance in non occluded expression recognition,but also has an average contribution rate of 4.45% in occluded.3.Joint occlusion completion and expression recognition network a real-time recognition system of occluded expression is established,which is used to monitor the driver's expression in real time in safe driving,and prevent traffic accidents caused by road anger and other bad mood.By using the improved generation of the best model of confrontation expression for network training,the expression model is called in the expression recognition network to realize the dynamic facial expression recognition with occlusion,and the system visual interface is constructed.