Abstract
<title>Abstract</title> <p>The digital restoration of traditional silk paintings requires a carefully controlled balance between damage removal, preservation of historical evidence, and continuity of material texture. Unlike generic natural-image inpainting, cracks, pigment loss, low-frequency stains, vertical aging stripes, and background oxidation in silk paintings are often intertwined, at the same spatial scale, with garment strokes, hairlines, inscriptions, seal boundaries, and silk-fiber textures. If a generic generative inpainting model is applied directly, the algorithm may pursue visual continuity at the cost of overwriting original brushwork and documentary information; if the process relies entirely on conservative morphological operations, it may miss low-contrast fine cracks and chronic sedimentary stains in figure regions. To address this tension, this study proposes a semantic-protection and quality-gated multi-stage digital restoration method for composite damage in traditional silk paintings. The framework first constructs explicit protection regions for figures, text, seals, fine line drawing, and high-risk pigment areas, placing cultural-heritage semantic safety before damage detection and image generation. It then integrates grayscale, HSV, LAB, local statistics, low-frequency illumination, gradient, and multi-scale morphological responses to form interpretable evidence for damage candidates. Candidate connected components are further routed to harmonic restoration, deep generative inpainting, or conservative skipping according to their semantic location, area, mean width, and local detail density. Finally, a restoration quality gate based on boundary luminance discontinuity, high-frequency energy ratio, and gradient residual ratio performs local soft rollback for connected components with risks of seams, residual structures, or over-generation. Process validation on twelve real silk figure-painting images shows that figure regions account for 36.77% of image area on average, and final candidate restoration regions account for 24.32% on average. In three no-reference quality-evaluation samples, the mean gradient preservation ratio is 0.9824, and changes outside the mask are substantially lower than those inside the mask, indicating that the framework is able to maintain relatively stable structural constraints under substantial damage coverage. The proposed method is not intended as one-step automatic beautification; rather, it offers an interpretable, auditable, rollback-enabled restoration paradigm suitable for expert-in-the-loop cultural-heritage image conservation.</p>