Research On Cross-database Facial Expression Recognition Method Based On Transfer Learning

With the rapid development of artificial intelligence,facial expression recognition in recent years has important applications in security intelligent monitoring,clinical medical detection,commercial marketing recommendations,online education,etc.The research on accurate facial expression recognition technology is of great significance.Because facial expression images have subtle inter-class different information and intra-class public information,extracting discriminative local features becomes a key issue in fine-grained facial expression recognition.In the past few years,facial expression recognition based on convolutional neural network has achieved excellent performance,where training and test data are generally considered to have the same distribution.In practice,this assumption is usually broken,especially when the training and test data come from different databases,which is a cross-database facial expression recognition problem.Therefore,in response to the above problems,this article mainly does the following work:(1)Based on ResNet-50,a hierarchical bilinear pooling residual network based on channel attention is proposed to better locate the significant local area changes of facial expressions.The model uses an effective channel attention mechanism to explicitly model the importance of each channel.It assigns different weights to the output feature map,and locates the significant region according to the weight value.A new hierarchical bilinear pooling layer is added,which integrates multiple cross-layer bilinear features to capture the inter-layer part feature relations,and spatial pooling is carried out in the feature map in an end-to-end deep learning way,so that the proposed network model is more suitable for fine facial expression classification.(2)Based on the attention hierarchical bilinear pooling residual network,this paper proposes a adversarial domain adaptive expression recognition method based on attention hierarchical bilinear pooling to learn a better target domain task classifier.By integrating task classifier and domain classifier,the network overcomes the limitations of pattern collapse caused by the separate design of task classifier and domain classifier,and the joint distribution alignment of features and categories across domains caused by pattern collapse.The network has a new adversarial target,which establishes a mutual inhibition relationship between the category prediction and the domain prediction of the input instance.Learning domain invariant features to reduce domain shift is used for cross-database expression recognition,to deal with the problem of difficulty in obtaining a large amount of labeled data in practical applications.