Facial Expression Recognition Based On Attention Mechanism

As an important branch in the field of affective computing,facial expression recognition technology has good application prospects in the fields of safe driving and human–machine interaction.Research on facial expression recognition technology is of great significance to promote the further development of artificial intelligence.The existing facial expression recognition technologies are mainly suitable for facial expression images collected under laboratory conditions.However,most facial expression images collected in real life are subject to some environmental factors such as human posture,light changes,and face occlusion,which greatly increase the difficulty of facial expression recognition.This paper studies facial expression recognition based on attention mechanism,and the main research contents are as follows:(1)In a facial expression image,the expression features are often concentrated in some local areas of image,while convolutional neural network is used to extract expression features uniformly.To solve the problem of uneven distribution of expression features,a facial expression recognition method based on the spatial domain attention mechanism is proposed,which can adaptively select the salient features that are more important for facial expression recognition in facial expression images,and give these salient features relatively high weights.The experimental results on RAF-DB and FER2013 show that the spatial domain attention mechanism can increase the average recognition accuracy of facial expressions by approximately 1.1% and 1.0%,respectively.(2)Using convolution layers to extract the features from a facial expression image will get the feature map tensor composed of a group of feature maps.To solve the problem of redundant feature maps in feature map tensor,a facial expression recognition method based on the channel domain attention mechanism is proposed,which can strengthen the role of features on important channels in the feature map tensor in facial expression recognition tasks.The experimental results on RAF-DB and FER2013 show that the channel domain attention mechanism can increase the average recognition accuracy of facial expressions by approximately 1.3% and 1.1%,respectively.(3)By combining the spatial domain attention mechanism and the channel domain attention mechanism,a facial expression recognition method based on the hybrid domain attention mechanism is proposed,which can extract more important features of a facial expression image in both the spatial domain and the channel domain.The experimental results on RAF-DB and FER2013 show that the the hybrid domain attention mechanism can increase the average recognition accuracy of facial expressions by approximately 2.1% and 1.5%,respectively.(4)In order to make use of the abundant expression features in facial expression images,a multi-scale attention network MANet is proposed.This network uses multiple convolution kernels of different sizes to extract multi-scale features under different sizes of receptive fields.After the scale features are fused,the attention mechanism is added so that the network model can learn more important expression features.The experimental results on RAF-DB show that the accuracy of expression recognition is as high as 84.3%.(5)A face expression recognition system is designed includes face detection module and expression recognition module.Through the reasonable layout and design of the user interface,the results of facial expression recognition are displayed in a visual way.