Research And Application Of SCN Expression Recognition Algorithm Based On Improved MobileNetV2 Feature Extraction

The development of machine learning and deep learning has brought unprecedented opportunities to the research of face expression recognition.In the application of deep neural network algorithm combined with face expression recognition has achieved remarkable results.At present,most face image and video recognition tasks are feature extracted based on convolutional neural network.However,deep neural network has problems in practical application,such as large computation amount and many model parameters.For embedded devices and mobile terminal devices,its application is greatly limited.Therefore,the study of lightweight convolutional neural network and the application of lightweight convolutional neural network in face expression recognition algorithm have great significance.This study addresses the lightweight and facial expression recognition techniques of convolutional neural networks.Using the advantages of self-repair network(SelfCure Network,SCN)expression recognition algorithm,the improved MobileNetV2 is used to make a feature extraction network for facial expression images.Adjust the network layer structure of MobileNetV2;embed CA(Coordinate Attention)attention mechanism in the model to enhance the learning of important features;adopt global depth-by-depth convolution(global depthwise convolution,GDConv)instead of the global average pooling layer;adjust the convolution kernel size of depth convolution(DW);and adjust the Mobile Net width factor to further reduce the parameters of the model.The MSCN model was finally designed and constructed with a width factor of0.5.Without pre-training,the identification accuracy of the model on RAF-DB and FERPlus was 79.46% and 81.65%,7.81% and 5.37% higher than the original SCN algorithm,respectively.The number of model parameters and the computed amount were reduced by 11511608 and 1731.21 MFlops,respectively,which are greatly reduced by 1.52% and 4.95% of the original model,respectively.At the same time,the improved MSCN model is compared with the classical convolutional neural network model.While the calculation amount and parameter number of MSCN are much lower than other comparison models on the premise of maintaining the highest accuracy.In order to realize the application of MSCN on mobile terminal devices,the trained model was used to build a face expression recognition system,and a We Chat small program was developed with We Chat small program as the system platform.The system test results show that the system has high expression recognition accuracy,convenient operation,smooth operation,complete function,and proves that MSCN lightweight face expression recognition system has a wide application prospect.In conclusion,this paper presents a lightweight face expression recognition model with high accuracy,small volume and low parameter volume computation,which is successfully applied to mobile devices.