Photorealistic image style transfer based on attention mechanism and histogram loss

Abstract

At present, the image stylization method is commonly based on the artistic style transfer method, so the generated results usually have the following problems. Comparing the result of the generated image with the content image, there are obvious content distortion and deformation, accompanied by loss of details, distorted light-dark relationship, and unbalanced pixel distribution. In solve above problems, this paper suggests a photorealistic image style based on attention mechanism and histogram loss. The method based on the channel attention mechanism proposed in this paper can preserve the semantic information of content image fully. At the same time, the proposed histogram loss module can better measure the high-level perceptual and semantic differences among images and optimize the high-dimensional feature distribution of images. In addition, this paper also proposes an evaluation method based on the color gamut space histogram to analyze the image generation results more effectively. Experimental results show that, compared with other mainstream methods, the proposed network model can achieve more realistic style transfer while maintaining the details of image content.