Facial Attributes Recognition Combined with Feature Decoupling and Static-Dynamic Joint Graph Convolutional Network

The existing facial attributes recognition methods either use attributes grouping to extract features or calculate the attributes co-occurrence probability to construct static attributes relational graph.The former learns insufficient complementary information and it cannot determine the correlation of attributes between groups and the correlation degree between attributes.The latter may reduce the generality of the model due to the bias of occasional co-occurrence attributes pairs in face images.In order to solve the above problems,a facial attributes recognition method combined with feature decoupling and static-dynamic graph convolu-tional neural network is proposed.Firstly,deep convolutional neural network ResNet-50 is used to extract features containing attributes representation information.Then,a feature decoupling module is designed to learn the specific features corresponding to each attribute.Finally,the static graph and dynamic graph are combined to learn the correlation between attributes,and the one-dimensional convolution layer is used to recognize facial attributes.The average accuracy on CelebA and LFWA datasets reach 91.85%and 88.17%,respectively,which are superior to other state-of-the-art methods.