Research And Application Of Facial Expression Recognition Algorithm Based On Attention

With the development of digitalization,facial expression recognition algorithms in big data environments have broad application prospects in intelligent control,safe driving,and assisted teaching.In recent years,facial expression recognition methods based on deep learning have become a popular research direction.In deep learning-based methods,neural networks can correctly extract the subtle features of expressions and understand the deep semantic content,which is suitable for expression recognition.However,in natural scenes,the captured facial expression images contain a large amount of noise,and the noise will hurt the recognition results.To explore expression recognition methods with robustness in natural scenes,many researchers have achieved good recognition results on facial expression recognition tasks,but there is less research on attention methods on face recognition tasks.Attention methods can improve the attention of neural networks to the correct features and improve the recognition accuracy of neural networks.In this thesis,we conduct a study on facial expression recognition based on neural networks and attention.The main studies done are as follows:(1)For facial expression recognition in natural scenes,Res Net-18 is selected for facial expression recognition research,and a network model based on the attention mechanism of convolutional channels is proposed.Firstly,The effect of different attention on improving model recognition accuracy was explored on the Res Net-18 network.Secondly,to address the problem that the compression ratio parameter of SE channel attention features is difficult to adjust,this thesis proposes the convolution-based channel attention module CSE,which replaces the fully connected layer with a convolutional layer with convolutional kernel size of 1×1.In addition,for the problem of uneven distribution of face expression dataset in natural scenes with large intra-class variance and small inter-class variance,the Focal Loss function and Is Land Loss function are combined as the model loss function.Finally,to address the complex features of expressions that are difficult to be understood by a single network,the channel attention CSE is inserted into the VGG-11 network and trained as a Res Net-18 auxiliary classification network to jointly recognize face expressions through multi-network decision fusion,and the recognition accuracy reaches 73.89% on the Fer2013 dataset.(2)The feasibility of Self-Attention on facial expression recognition task is investigated.First,the attention regions of Vision Transformer and Swin Transformer on face expression pictures are compared,and the experimental results show that the attention regions of Swin Transformer are more continuous and more helpful for face expression classification.After removing the self-attention module in the network,the accuracy of the network in recognizing facial expressions in natural scenes decreases significantly,which proves that the self-attention can improve the model’s ability to classify expressions in natural scenes.(3)Based on the PyQt toolkit,a face expression detection and recognition system in the classroom is designed.Firstly,to address the problem that MTCNN face expression detection algorithm is difficult to train and has the problem of missing and false detection,we trained the improved Yolo V4-Tiny for face detection to solve the problem of missing and false detection of MTCNN.Finally,the system is combined with an expression recognition network for expression recognition,and the system has a good recognition effect and application value.