Image-to-Image Translation Based on Pre-trained Multi-scale Discriminator

Image-to-image translation has become a crucial research area, aiming to enable one image to absorb the style of another while retaining its own content. In this research the CycleGAN has made significant strides but is marred by the problem of generating images with inadequate authenticity. This study presents an innovative solution by introducing a pre-trained multi-scale discriminator into the CycleGAN framework. The framework consists of a cycle-consistent network and the discriminator. It projects generated and real images into a pretrained feature space, leveraging advanced semantic features to direct the generator. Multi-scale feature discrimination and a cross-channel mixing strategy in the random projection are employed to comprehensively evaluate image quality and enhance the discriminator's focus on multi-channel correlation features, thereby enhancing the authenticity of the generated images. During the experimental phase, the Cityscapes and Foggy Cityscapes datasets were employed, yielding remarkable results. The generated images exhibit a high level of realism in aspects such as light and shadow, texture, and the overall scene atmosphere. Comparative experiments with CycleGAN evidently demonstrate the superiority of this method in generating more authentic and higher quality images, thereby effectively achieving the conversion between different image domains.