An improved semantic and sketch biconditional guided image inpainting model for Chinese landscape painting

Traditional Chinese landscape paintings frequently suffer from degradation, such as wormholes and cracks, due to aging. While deep learning has advanced natural image inpainting, existing algorithms often overlook the unique artistic style and complex structure of these paintings, yielding results that lack stylistic consistency. To address this, we propose a novel inpainting model utilizing biconditional guidance from semantic and sketch priors, specifically tailored for Chinese landscape paintings. Our model fuses structural and semantic information into a Transformer backbone to enhance coherence within defective regions. We introduce a lightweight stylistic feature extraction (LSFE) module and an improved multi-branch feature fusion module to reconstruct traditional painting strokes. Furthermore, a composite loss function integrating structural, style, perceptual, and reconstruction losses ensures stylistic authenticity. Validated on a specialized dataset, objective (PSNR, SSIM, FID, LPIPS) and subjective evaluations indicate that our model outperforms state-of-the-art methods, substantiating its efficacy in Chinese landscape painting inpainting.