Improving CycleGAN for Image-to-Image Style Transfer by DenseNet

With the continuous progress and development of machine learning, deep learning has become a hot research field today. Among them, image-to-image style transfer is one of the research hotspots. Generally, the main goal of image style transfer is to extract the features of the original image and convert them into images with the target style through a deep neural network. Applying the cycle-consistent adversarial network known as CycleGAN to transfer the style of images is one of the existing image style transfer techniques, but CycleGAN has problems such as insufficient brightness enhancement, color distortion, and low authenticity of the generated images. To address this problem, in this paper, we construct an improved CycleGAN, use DenseNet to replace the original ResNet, and compare the image effects generated by the improved network using DenseNet and the original CycleGAN network based on ResNet. After calculating through the image evaluation indicators PSNR and SSIM, the results show that the images generated by the improved recurrent consistency network based on DenseNet have better visual effects and realism.