Deep Face-Swap Model Combining Attention Mechanism and CycleGAN

Abstract

Deep face swapping is a hot field in computer vision, which uses deep learning technology to replace the face in the original image while keeping other contents unchanged. To solve the problem of discontinuities across the blending boundaries in existing mainstream deepfake methods, we propose a face swapping model combining attention mechanism and CycleGAN. Firstly, a generator incorporating the attention module is used to obtain the generated graph. Then, the discriminator based on PatchGAN is used to discriminate if the graph is synthesized. Finally, the least square loss and SmoothL1 loss are used to improve the loss function. Experimental results manifest that our method reduces the difference by 0.43 in facial region and increases the similarity by 0.56 in background region. The model we proposed can improve the quality of face swapping images effectively, and enrich the existing face swapping methods, which has great significance for both generation and detection of deep face swapping research.