Facial Expression Recognition Based On Squeeze And Excitation Block

Facial expressions contain rich semantic information,which is one of the important ways of communication between people.If computers can obtain the information conveyed by facial expressions,it will have a huge impact on our lives.At present,facial expression recognition has gradually gained widespread use in the fields of human-computer interaction,driving,communications,medical care,and security.Application and facial expression recognition have also become one of the research hotspots in industry and academia.In recent years,expression recognition algorithms based on convolutional neural networks have gradually gained a place among various expression recognition algorithms due to their strong adaptability.At present,many improvements to this algorithm focus on improving the depth and breadth of the network to better fit complex models,or improving the performance of specific classification tasks by improving the loss function,but the accuracy of recognition depends on the data.The limitations of the collection and network structure have not achieved obvious advantages over other methods.This article focuses on three aspects: the preprocessing of facial expression recognition samples,the improvement of each step of the SE module and the promotion of the structure,and the network model embedded in the SE improvement module.The main contents are as follows:First of all,due to the limited data sets available in the field of facial expression recognition,the number of samples collected in the laboratory environment is often relatively insufficient,which makes it difficult to support the training of complex networks.However,the data set obtained in the environment through big data and other means is due to the different frequency of the appearance of each expression,which causes the direct distribution of each sample to be unbalanced,which also causes difficulties in the training of the network.In this paper,corresponding data set preprocessing methods are adopted for the above problems.In addition to face recognition and cropping and sample adjustment and standardization,it also adopts an oversampling strategy for the uneven distribution of samples,and for the insufficient number of samples.Coping methods such as data augmentation and migration learning have been adopted,and certain results have been achieved.Secondly,in view of the limitations of the current network structure,which is different from the improvements in depth and width of networks such as VGG,Inception,and Res Net in recent years,this article provides a new entry point for improving the expression recognition network.In this paper,on the basis of each typical network,the Squeeze-and-Excitation module is introduced to improve the network structure.It uses the information of the entire convolution kernel threedimensionally to calculate the weight between each channel and adjust the original feature map.,It has an effect similar to the attention mechanism,and realizes the enhancement of important channel characteristics,and realizes the improvement of network performance while retaining the original network characteristics.This article also improves the structure of the Squeeze-and-Excitation module,scrutinizes the details of each step and extends the overall structure by analogy,and experimentally verifies the effectiveness of each improvement.Finally,in addition,this article also discusses the combination of Squeeze-andExcitation module and each typical network.Through experiments,the best combined module structure is obtained and the combined network model is given.The improved network has been researched on various commonly used facial expression recognition data sets,and has achieved 73.502% accuracy and 98.987% accuracy on the FER2013 and CK+ data sets,which is a certain improvement over the relevant research over the years.