Combining CycleGAN and perceptual loss for image style transfer: Double training process
Hanjun Li · Applied and Computational Engineering · 2024
Image style transfer is an important field within the broader area of image processing. While existing style transfer models have addressed certain challenges, there remains room for improvement in areas such as color restoration, edge detection, and the ability to effectively transform lighting conditions. Additionally, there is a lack of research exploring the features and properties of these models. To address these limitations, this paper proposes a double training method that combines CycleGAN with the minimization of perceptual loss for style transfer. The proposed method is tested and compared to existing models, resulting in impressive performance improvements, particularly when the style image background is white. These comparative results not only validate the effectiveness of the double training method with CycleGAN, but also provide valuable qualitative insights into CycleGAN and perceptual loss. Drawing from these conclusions, this work further proposes a new model hypothesis that builds upon the double training method and includes background-object segmentation and background whitening. This novel approach aims to enhance the effectiveness of style transfer by allowing objects to maintain their original forms while transferring the style of the source image to the target image.