Improved CycleGAN for natural scenery images style transfer

Abstract

Natural scenery images style transfer is a technique using computer technology to change the stylization effects of images by processing the high-level features which are extracted by images in neural networks, and is used to improve the diversities and aesthetics of images. Since existing neural network models cannot achieve a good effect when dealing with the style transfer tasks of natural photos, this paper proposes an improved CycleGAN method that has the advantage of changing two unpaired image datasets in style. In order to save more image content and solve the model overfitting problem, we added a channel attention mechanism to the generator and optimized the cycle consistency loss. We defined the developed loss function as MS-SSIM+SmoothL1 in this paper. The method can alleviate the overfitting phenomenon of the model as the epoch increases. The images generated by our proposed method have better performance in detail. Experiments demonstrate that the images generated by our proposed improved network are more correspond with human perception in visual. In the FID score, our proposed method was 42.24% lower in the Summer2winter datasets and 23.76% lower in the Monet2photo datasets than CycleGAN.