A Scale-adaptive Color Preservation Neural Style Transfer Method

Abstract

Since the feature maps of deep neural networks were adopted to compute the representation of style and content information, neural style transfer (NST) methods have sprung up like mushrooms. But the existing methods ignore a fundamental fact that a style or an artistic image not only contains style information but also contains content information. And we find that there may be a conflict between style and content. Motivated by this idea, we propose a novel method, which only adopts the detail layer of the style image to compute the style loss. To avoid the potential conflicts between the style loss and the content loss, we just abandon the latter. The smooth base layer of the content image will be added to the intermediate results to keep the semantic content invariant. Our ablation studies show that this strategy can make the results scale-adaptive to the style image. Furthermore, we use an interpolation method so that the overall color of our results remains unchanged and our results have a colorful stroke. The qualitative and quantitative analyses show that our results have a better visual effect than the existing methods.