Research And Implementation Of Facial Expression Recognition Based On Convolutional Neural Network

With the progress of the information age and the development of computer technology,facial expression recognition technology has begun to attract the attention of many experts and scholars.Due to its unique technical characteristics,it is widely used in the field of intelligence such as human-computer interaction,hospital care,game interaction,and other applications.Due to the fact that there are some individual differences in human facial expression features,traditional expression recognition technology is often accompanied by defects such as inadequate feature extraction and largely affected by the external environment,resulting in a low average recognition accuracy rate and the problem of poor generalization ability of different people.Using deep learning technology can better extract the individual feature details in facial expression images,so that its final recognition accuracy is high and has a strong group generalization ability.This paper studies and implements an asymmetric Batch Normalization-Convolutional Neural Networks(BN-CNN)facial expression recognition system based on Principal Component Extraction(PCE).The main work completed in this article are:(1)Put forward a new Convolutional Neural Networks(CNN)network structure suitable for expression recognition applications.The Pooling layer in the traditional CNN was replaced with a batch standardization layer(BN layer),and a Dropout layer and a Softmax layer were also added to complete the construction of the BN-CNN network structure.At the same time,the loss function in network training is optimized,and the gradient optimization function in network back propagation is selected and updated.(2)If the image is directly input to the BN-CNN network,it will cause the problem of imperfect feature extraction.In this paper,a preprocessing module is added on the basis of the BN-CNN network model.The preprocessing module improves the traditional principal component analysis method and proposes a principal component extraction method based on facial structure point detection.This method locates facial structural points,extracts the main facial organ regions,and then inputs the processed pictures into the BN-CNN network model for training.(3)Aiming at the problem of reducing the number of real-time dynamic recognition frames after adding the preprocessing module,asymmetric convolution is added to the BN-CNN network structure to obtain a good recognition accuracy and realtime dynamic recognition frame rate.The asymmetric BN-CNN facial expression recognition system proposed in this paper is very robust.