ResFormer diffusion: latent diffusion model with residual transformer for Qin bamboo character image expansion

The scarcity of Qin bamboo slips character image samples leads to poor performance in archaeological research tasks in practical scenarios such as recognition and restoration. This paper proposes a diffusion model based on residual attention for Qin bamboo slips character image expansion, which effectively solves the problem of small sample size. In order to reduce the impact of image background on the quality of expanded images, Laplace transform and median filter methods are used for preprocessing. In addition, this article integrate the Vision Transformer architecture into the potential diffusion model to enhance the global context modeling ability and network scalability, and also open up a direct path through the network for propagating attention scores to stabilize model training. Guided by the classifier, the category information of the character is added to the generation condition to achieve the generation of images of specified categories. Experimental results show that the FID values of the proposed method are reduced by 25.9% for WGAN-GP, and 22.3% for DDIM, respectively. The Qin bamboo slips character images generated by the proposed method are of high quality and diversity, which effectively solves the problem of insufficient training samples for Qin bamboo slips character images.