Abstract: Aiming at the problems existing in the process of face image posture correction, such as ignoring detailed features, blurring the texture of the face image after correction, and excessive gap between the face features and the identity features of the original image, we proposed a face correction method based on the dual discriminator and dual attention mechanism of the generative adversarial network. Firstly, while using the dual discriminator to determine the face identity and face posture, we designed a face representation attention module to enhance the overall features of the face, prevented the loss of face-related information, and improved the model’s ability to perfect face images. Secondly, we designed an adaptive edge-enhanced attention module. By using the adaptive attention mechanism and Sobel filter, the edge detail features and key features of the face were enhanced to generate a frontal face with realistic facial features and contours. Finally, a new normalization layer CrossNorm was adopted to improve the robustness under distribution changes,  and the results of testing  experiments on the Multi-PIE dataset and the CFP dataset show that the proposed model generates frontal face images with better face correction effects compared to the comparison model.

Key words: generative adversarial network, attention mechanism, face correction, face posture