Conditional Adversarial Domain Adaptation With Attention Mechanism For Facial Expression Recognition

Facial expression recognition is an important field in computer vision.With the continuous development and in-depth research of facial expression recognition,facial expression recognition has been widely used in psychology,education,digital entertainment and human-computer interaction.When the convolution neural network is used to extract facial expression features,a large amount of training data is often needed.Because of facial expression variability and the facial expression data collected in a controlled laboratory environment as well as the facial expression data collected from the outside world without control factors,there exist large differences between different data sets.Meanwhile facial expressions of posture change,complicated background,facial expressions,shelter and the expression of the same facial expression by sampling subject exist different scale changes are the factors expanding the distribution.How to make full use of the distribution discrepancy datasets makes the facial expression migration study becoming an important research subject.Most of the existing face recognition are based on transfer learning adopts shallow network and deep learning network.This paper proposes a face expression transfer learning method with embedded attention mechanism model for face expression recognition.The main research work is as follows:1.In the face expression feature learning part,this paper uses the convolutional neural network of Res Net50 for feature extraction,takes the features of the last full connection layer as the classification features,and improves the effectiveness and accuracy of feature extraction of face expression images by embedding the attention mechanism model to improve the convolutional neural network of Res Net50.2.In the transfer learning part,the feature distribution of the source domain and the target domain is aligned based on the conditional generative adversarial domain adaptation network and the domain classification loss is used to reduce the distribution difference between the two domain feature representations,so that the model trained by the source domain data can be applied to the target domain data.Thus,the unlabeled data can be used to train the convolutional neural network for facial expression recognition.At the same time,the entropy function is used to guarantee the portability of the uncertain facial expressions predicted by the classifier.3.In order to study how the Conditional Adversarial Domain Adaptation with Attention Mechanism for Cross-domain Facial Expression Recognition method proposed in this paper works,control of laboratory data collected by CK + database,JAFFE data sets and the facial expression in real life data SFEW database,FER2013 data sets are used as target data sets in the experiment.Through the contrast experiments and result analysis,we show the effectiveness of the proposed algorithm.