Facial Expression Recognition And Application Based On Deep Learning Attention

In addition to verbal and physical actions,expressions are also an important form of communicating information in social interactions and everyday life.Other people can judge a person’s emotional state by their expressions,which are more accurate in reflecting the true emotions of the human heart than verbal behaviour or physical actions.As hardware technology continues to develop and artificial intelligence advances rapidly,making machine recognition of human expressions a hot topic at the moment,high accuracy recognition of human expressions through machines has been promoted in areas such as human-computer interaction,security,robot manufacturing,healthcare and driving.Recently,more and more scholars have been using deep learning for Facial Expression Recognition(FER),but the traditional Convolutional Neural Network(CNN)has the disadvantages of complex models and coarse extracted features,resulting in single scenarios and low recognition accuracy.However,the traditional Convolutional Neural Network(CNN)has the disadvantages of complex models and coarse feature extraction,resulting in the use of a single scenario and low recognition accuracy.In addition,the background of face expression recognition in real life is complex and there is a lot of useless information,which increases the difficulty of feature extraction and makes the recognition accuracy decrease.To address the above problems,this paper improves the face recognition algorithm with the following main research points:(1)The traditional deep residual network can solve the problems of gradient disappearance and overfitting,but when the depth of the network reaches hundreds or thousands of layers,it needs to double the number of layers if it wants to improve the accuracy rate by one percent again,making the training speed of the network very slow.To address this problem,this paper introduces Wide Residual Network(WRN)for expression recognition,which reduces the depth and increases the width of the traditional residual network,resulting in a 2.58% increase in recognition accuracy.(2)Facial expressions are composed of multiple regions such as eyebrows,mouth,eyes,chin,etc.,and there are different focuses when extracting features for different expressions,such as focusing more on the mouth region when smiling,and focusing more on the eyes region when angry.The commonly used single-headed attention mechanism cannot assign different weights to these regions,so this paper introduces a multi-headed cross-attention network and uses a partition loss function to assign each attention to a different facial region,capturing multiple local features.expressions are then predicted.(3)Facial expressions have high intra-class variability and inter-class similarity,which poses a greater challenge to the classification problem.In this paper,we introduce Affinity Angle Loss(AALoss),which uses the intrinsic relationship between class centres to increase the interclass distance,and uses the angle information contained between the feature vector and the weight vector to make the learned features have angular distribution properties,further reducing the intra-class spacing,making the same class tighter and the separation between different classes contours more clearly.