Digital reproduction of ICH images should consider both the style fidelity and cultural semantic expression. Given that the existing methods are prone to losing the unique visual grammar of ICH in reconstruction, this paper proposes a joint model for deep reconstruction of ICH artistic image style and adaptive enhancement of visual features to address this problem. Gram matrix is used to extract the embedding of ICH style from the model, and then combined with the improved ICH-AdaIN mechanism to achieve decoupled reconstruction of content and style. An adaptive vision-enhancement module that is guided by cultural attention will be added to strengthen the foundation of culture. Experiments have been conducted on 1,280 image datasets of the four types of ICH, such as New Year pictures and paper-cuts; compared with the U-Net baseline, the proposed method reduced LPIPS-style by 27% and FID by 15.3, increased local SSIM to 0.91 in key areas, and reduced the standard deviation of background noise by 12%. The expert’s score reached 4.3/5 and had risen from the original 2.8/5. According to the above results, the model can reconstruct high-quality images that meet the visual requirements of ICH, and is thus suitable for use in digital protection and activation applications.
