The Khitan Small Script, a crucial cultural artifact for Liao Dynasty research, presents significant restoration challenges due to severe damage, incomplete decipherment, and scarce extant materials. Current methods rely on predefined mask regions, yet the unknown extent of damage complicates mask determination. Furthermore, its unique character-combination system, with an undetermined number of elements, adds to the complexity of restoration. This paper proposes a blind restoration method based on an improved CycleGAN model. The model establishes global pixel correlations between damaged and restored images, transferring incomplete text style features to a complete style for restoration. U-Net replaces ResNet to enhance detail and structural consistency, while dilated convolutions and self-attention mechanisms improve feature representation and text detail handling. WGAN-GP adversarial loss ensures the stability and precision of the restoration. Experimental results show that the improved CycleGAN model effectively restores missing regions in Khitan Small Script inscriptions, achieving superior peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) compared to traditional methods.