Research On Face Expression Detection Based On Improved Faster R-CNN

With the development of science,the academic research on artificial intelligence is becoming more and more popular.As one of the important branches of computer vision,more and more researchers are engaged in it.In order to truly realize artificial intelligence in computers,it is essential to understand human emotions,and facial expressions can often directly express people's rich emotions,so the recognition and detection of facial expressions is very important.Nowadays,deep learning has achieved excellent results in computer vision fields.In this paper,a popular deep learning algorithm called Faster R-CNN model was studied and applied to the detection of facial expressions.Because facial expression itself is easy to be confused,and is easily affected by environment,Angle and other factors,there is still room for improvement in average detection accuracy of facial expression based on Faster R-CNN.This paper proposes an improved Faster R-CNN based facial expression detection method.In this method,histogram equalization and adaptive histogram equalization are preprocessed for SFEW 2.0 of the facial expression data set,and the facial expression data is enhanced and expanded.Then the repetitive experimental optimization of the hyperparameters is carried out to improve the training and learning effect of the model and improve the detection accuracy.The network structure of vgg-16 and resnet-101 was analyzed and compared.Through experiments,the network weight of vgg-16 was determined to transfer learning and improve the generalization ability of the model.In the end,based on the regularization model structure optimization,L1 Smooth regression loss function with parameter constraint term were proposed.The regularization method was used to optimize parameter weight,improve detection accuracy,and an improved Faster RCNN model adapted to face expression characteristics was obtained.Experiments show that the improved Faster R-CNN algorithm for face expression detection proposed in this paper can successfully detect the location of facial expressions and identify the expressions.The m AP reaches 81.33%,which increases the average accuracy by about 4% compared with the face recognition and detection using the unimproved Faster R-CNN.