Image Animation Based on Improved CycleGAN

Aiming at texture blur and color distortion in the image generated by the original CycleGAN in the image animation task, this paper inserts self-attention mechanism into the skip connection of the generator to strengthen the acquisition of global information in the decoding part of the generator, and trains the model with the perceptual loss function Lpips to improve the sensitivity of content changes before and after image animation. In view of the network degradation, the feature converter is designed as the parallel residual block structure to improve the efficiency of feature conversion while reducing network depth. The experimental results show that compared with the original CycleGAN, CartoonGAN, and AnimeGAN, the improved CycleGAN not only has a better animation visual effect, but also outperforms the other three methods in the evaluation index FID.