Research on Image Style Transfer Algorithm Based on Structural Consistency
Chuxiong Yang, Zhuang Chen · 2023
The algorithm proposed in this paper aims to address the issue of structural content distortions in images that occur after applying image style transfer. It introduces a structural consistency-based approach called the PoolNet algorithm. This algorithm incorporates saliency regions to perform style transfer at the block level. By employing a saliency detection network, a saliency map is generated for both the synthetic and content images. During the training process, two losses are computed to ensure that the salient regions in the synthetic image remain consistent with those in the content image. Additionally, local affine constraints in the color space are employed to enhance the stylized images by constraining the image transformation. Experimental results demonstrate that the stylized images produced by this style transfer model exhibit improved visual effects and preserve the semantic information of the content images. This improvement is particularly noticeable when dealing with distorted migration images.