A handwritten Chinese character image evaluation method based on CLIP pre-training alignment

This paper proposed a method for evaluating handwritten Chinese character images using CLIP pre-training alignment. Traditional evaluation methods face challenges due to the lack of a unified standard and the need to consider both structure and style. Therefore, we introduced a CLIP model that aligns images with text descriptions through contrastive learning and designed an SSF-Attn module to enhance Chinese Handwritten Hanzi style perception. Additionally, we constructed a multimodal dataset and proposed the SSHID metric to measure style differences. Finally, the experiments shown that our HS-CLIP network outperforms traditional methods in image-text alignment, and SSHID outperforms FID in Chinese Handwritten Hanzi style perception. Future work will focus on improving the model's interpretability of Chinese Handwritten Hanzi image features.