Research On Facial Expression Recognition Of Image Sequence Based On Deep Residual Network

With the rapid development of computer vision technology,facial expression recognition technology has gradually risen.Whether in academia or applied in industry,how to realize the automatic recognition of facial expressions has become a hot topic of research.The current facial expression recognition technology has huge development potential and rich application scenarios,which can be used in scenarios such as intelligent human-computer interaction,safe driving of cars,auxiliary medical care,and online education.So far,most of the research objects of facial expression recognition algorithms mainly use static facial expression images,but facial expression changes are a dynamic process.Only using static facial expression images cannot take advantage of the temporal and spatial characteristics of facial expression changes.Compared with static images,facial expressions in image sequences can capture more motion features and texture features,which can improve the accuracy when performing facial expression classification.The main research work and innovations of this article are as follows:1.Analyze the preprocessing operation of the input image to improve the recognition rate of facial expressions.Different from the traditional convolutional neural network,which directly detects the original image of the human face as input,the algorithm in this paper performs a series of image preprocessing operations before entering the network.Firstly,the latest high-accuracy face detection algorithm Retina Face which was used to improve the recognition rate of complex and non-positive faces.Then,the face alignment method based on the four points of the face was used to align the faces,followed by image cropping and intensity Normalize.Finally,the traditional method is used to extract the rotation invariant local binary pattern(LBP)map as the final input.2.The algorithm of deep residual network as the backbone network for static facial expression recognition is studied.First,a series of image preprocessing operations are used to obtain the LBP map of the input image,and then the LBP map is used as the input of the deep residual network.Finally,the network model is trained to use the Softmax layer to classify facial expressions.The effects of different levels of residual networks,different forms of LBP operators,and other network structures on expression recognition are compared.Experiments were performed on the FER2013 database,and the results show that the algorithm in this paper has a high recognition rate.3.The algorithm of image sequence expression recognition based on deep residual network and long short term memory(LSTM)network is studied.First,the network trained on the FER2013 database is used as a feature extractor;next,the image sequence is used as a unit to extract each frame feature in turn,and then combined into a vector to form the expression time sequence feature;finally,facial expression classification is performed using the temporal features of expressions as input to long-term and short-term memorynetworks.At the same time,image sequence expression recognition was performed using support vector machine algorithm and random forest algorithm,and it was verified on Cohn-Kanade database and AFEW6.0 database.Compared with other image sequence expression recognition algorithms,the algorithm recognition rate has been improved in this paper.