Image Style Transfer of Smooth Feature Affinity Network: SFAN
Ying Chen, Zhihao He, Xiaohong Qin · 2025
The traditional attention mechanism style transfer stylizes images by manipulating the pointwise similarity between content and style features. This method often results in content blurring due to uneven feature distribution, specifically manifesting as distorted content shapes and dirty textures. To address this issue, we proposed a smoother style transfer network, SFAN, which leverages the properties of the Pearson correlation coefficient to recompute the similarity between content and style features at multiple scales. Additionally, a Ushaped decoder structure is employed to achieve multi-scale feature fusion, ensuring the content remains more intact. To alleviate the issue of excessive content feature representation caused by this decoder structure, a content feature expression coefficient is designed. Moreover, a novel progressive feature loss was derived based on SFA Module to enhance overall visual quality. Experimental results indicated that the images generated by this method exhibited more natural styles and more complete content compared to the traditional attention mechanism. The conclusion is that this approach results in images with a more natural style and more complete content.