Neural Style Transfer with Content Feature Segmentation
Xiaotian Fang · Highlights in Science Engineering and Technology · 2023
In our work we propose a new model by segmenting content features (i.e., segregating the feature map into several parts) of the content image in style transfer. We demonstrate that our novel way for style transfer with segmentation ties together the style of both global effect and local appearance that are used in majority of parametric and non-parametric techniques respectively. Our new model can successfully figure out the difficult problems with the current transfer techniques. For one thing, the new model in our research is able to avoid the distortion of local style maps, at the same time it also allows transfer in the level of semanteme compared to parametric means. For another thing, in contrast to other nonparametric approaches, it keeps the style image's overall appearance consistent and avoids washing-out images. We also present a feature optimization strategy based on this. Our trials have proven that our approach is suitable to a wide range of style genes and simultaneously yields higher-quality results over earlier papers.