ATFS: A deep learning framework for angle transformation and face swapping of face de-identification

This paper proposes an effective framework for improving angle transformation and face replacement. This ATFS framework aims to solve the image distortion caused by angled face swapping in face de-identification. In the ATFS framework, TransNet and SWGAN were proposed based on generative adversarial networks. When TransNet uses neural networks to reconstruct images, it combines the predicted facial points. As a result, Transnet can locate the target’s facial features and generate multi-angle transformed images to achieve angle transformation. The SWGAN applies a complicated neural network that combines residual operations and self-attention modules to extract the input image features and replace the face region of the input image with the target image to achieve face de-identification. Both TransNet and SWGAN adopt discriminators and various loss functions to train the neural network and accurately accelerate the training process. The experimental results show that the proposed method can maintain high-quality images and avoid image distortion during face swapping for image angle transformation compared to previously proposed methods.