Design And Implementation Of Facial Expression Recognition Model Based On Deep Learning

In interpersonal communication,facial expressions convey a large proportion of information.As one of the important technologies of artificial intelligence,facial expression recognition plays an important role in many fields,such as assisting medical personnel in the treatment of depression in the medical field and providing safety warnings to drivers in the field of safe driving.This article designs a facial expression recognition model by studying deep learning methods,aiming at completing accurate and efficient facial expression recognition tasks.The main research contents of this article are as follows:(1)Aiming at the problems of CNN pooling layer loss features and insufficient extraction of the capsule network’s convolutional layer features,a facial expression recognition model using Res Net18 to improve the capsule network was proposed.This model only retained the convolution layer of Res Net18,adjusted the four residual blocks to varying degrees,and incorporated the CBAM attention mechanism.Then,it replaced the single convolution layer of the capsule network to extract facial expression features.Finally,the extracted features were fed into the capsule network,where the primary capsule layer learned the features and the digital capsule layer classified them.The model designed in this article had been tested on three datasets: CK+,RAF-db,and FER+.Experimental results showed that the model was feasible and effective in facial expression recognition.(2)Aiming at the problem of large differences in recognition accuracy caused by the uneven amount of various facial expression data in the facial expression dataset,a data enhancement method called CycleGAN was proposed to equalize various facial expressions in the dataset,and experiments were conducted on CK+ and RAF-db datasets.Given that both datasets contained the largest number of samples of happy expressions,six models were obtained using happy expressions as source domain data and the other six types of expression data,respectively,as target domain data,and then various types of expressions were generated.Finally,the balanced RAF-db dataset was tested on the proposed network.Experimental results showed that data equalization could effectively alleviate the problem of large differences in the accuracy of various facial expression recognition.(3)In response to the problem of a large number of parameters in the expression recognition model designed earlier,a traditional channel pruning method was used to compress the model.This method was to prune the trained model.First,use the L2 norm to score all channels,and then set the pruning proportion.Based on the score of each channel,prune the model multiple times.After each pruning,the model was retrained.Finally,fine-tune the pruning model to obtain the final model.Experimental results showed that this method could effectively reduce the size of the model at the cost of less accuracy loss.(4)In response to the problem of slow single iteration training speed of the expression recognition model designed earlier,an initialization channel pruning method based attention mechanism was proposed.This method was based on the SENet channel attention mechanism module for pruning.During the training process,the channel attention mechanism could assign different weights to each channel based on the importance of each channel.Therefore,the channel could be effectively pruned according to the size of the weight in conjunction with the pruning module.Experimental results showed that this method can significantly shorten the training time of the model.