Image Translation Method Based on a New Self-Attention Cycle Generative Adversarial Network

The translation of images is expected to resolve the issue of acquiring a substantial number of images. Aiming at the problems of unstable training and weak feature learning ability of image translation network model, a new Self-Attention Cycle Generative Adversarial Network (SA-CycleGAN) is proposed. The model is based on the CycleGAN framework, but with several differences. It introduces a self-attention module in the generator and a residual network structure to guide it to focus on the most discriminative region between the source and target domains. Concurrently, LPIPS loss is incorporated into the loss function to facilitate the training of the generator to produce higher-quality images. The experiment demonstrated the efficacy of this method in preserving the image background while accurately translating the colour and texture nuances present in the standard data set. Concurrently, the translated images exhibit significant advantages in terms of the KID index in comparison with alternative translation methodologies.