Restoration of Liye Qin slips characters using generative adversarial network with effective connected component constraint

Bamboo slips, once key carriers of culture and writing. The Liye Qin Slips, unearthed in Liye Town, document aspects of Qin society and county administration, making their restoration vital for cultural heritage studies. Yet, blurred handwriting, faded ink, and distorted character structures hinder interpretation, while manual transcription remains labor-intensive. To address these challenges, we propose a generative adversarial network (GAN) with an Effective Connected Domain (ECD) constraint. Built upon a U-Net backbone, the model replaces standard convolutional blocks with Local Residual Dense Blocks (LRDB) and employs a 3-layer residual convolutional bottleneck to enhance feature extraction and preservation. The ECD constraint serves as an auxiliary loss, enforcing character structural integrity and improving restoration quality. Experimental results show that our method achieves superior performance, with PSNR of 15.97, SSIM of 0.82, and LPIPS of 0.07, validating its effectiveness in bamboo slip restoration.