Disentanglement Learning Generative Adversarial Network for Facial Expression Recognition

Facial expression representation is often entangled with other facial attributes unrelated to expression. Therefore, the traditional convolutional neural network is difficult to identify different expressions. In order to solve the problem of face information entanglement, a new end-to-end expression disentanglement learning generation GAN network (DLGAN) is proposed in this paper. We adopt a novel Gan structure, which includes encoder, decoder and discriminator network to learn the expression representation separated from other facial attributes in the process of image reconstruction. We use the feature vector learned by Gan instead of the original image as the input of classification convolution neural network to recognize facial expressions. In addition, this paper also introduces the self attention module to improve the learning ability of DLgan. The experimental results show that our proposed method is comparable with the most advanced methods.