Optimization of image style transfer algorithm based on CycleGAN

Abstract

Abstract In response to the current problems of uneven quality of images generated by the CycleGAN algorithm, low learning efficiency, instability of training models, and insufficient diversity in generation quality, as well as the violation of perception and aesthetics between images and humans after transfer. This article proposes a CycleGAN model that can generate images with better quality and fuller colors to address the aforementioned issues. This model incorporates a globally connected residual network with self-attention mechanism and improves the network structure of the generator and discriminator to improve the quality of generated images and solve the training imbalance problem caused by the original model. This article uses deep learning frameworks based on Python 3.8, Pytorch 1.10.0, and cuda 1.11.0 for experimental training. The experimental results show that the improved CycleGAN model generates better image quality and higher color saturation. Adding self-attention mechanism can indeed better grasp the degree of restoration of original image details and produce more realistic effects. However, further exploration is needed in terms of training efficiency and generated image clarity.