Research And Implementation Of Face Expression Recognition Algorithm Based On Video Image

Facial expression refers to a state formed by the combination of human facial features and muscles,which can accurately reflect human inner emotions.It is a subconscious natural physiological behavior,and also an important part of facial features.Expression recognition has a wide range of application scenarios,and plays an important role in human-computer interaction,fatigue driving recognition,clinical medicine and other fields.The complete expression recognition includes image-based static expression recognition and video-based dynamic expression recognition.The real and objective expression is a continuous dynamic change process,and a single picture can express less feature information,so it is difficult to accurately reflect the emotions expressed by human expressions.Therefore,in order to understand human expressions more accurately,it is necessary to understand them through information containing temporal relationship and video with context between contexts.The research focus of this thesis is the video frame sequence.In order to extract the spatiotemporal feature information of the video frame sequence effectively,a video expression recognition algorithm based on Capsule-LSTM is proposed.The research of this thesis includes:1.An optimized Capsule network is proposed.On the basis of the original Capsule network,several small convolution kernels are used in the Capsule encoder network to replace the large-size convolution kernels.The large convolution kernels can extract wider information in the region.In the case of the same receptive field,the parameters of the small convolution kernels in series are greatly reduced,and the nonlinear transformation has more possibilities.At the same time,the deconvolution layer is used to replace the original full-connection layer to optimize the network during image reconstruction,and the improved Capsule network has better performance than the original Capsule network through experiments.2.A Capsule-LSTM network model is built.In order to better extract the time sequence information between video sequences,the optimized Capsule network is combined with LSTM to recognize video facial expressions.Firstly,CAPSULE extracts the spatial information of facial expressions in video frames,and the output of CAPSULE encoder is used as the input of LSTM,which is used to extract the temporal sequence information between video frames and analyze the difference of expression changes between frames.The experimental results show that the Capsule-LSTM model proposed in this thesis can effectively improve the accuracy of video expression recognition on data sets MMI and AFEW.The accuracy is 72.11% on the MMI data set and 40.16% on the complex data set AFEW.