Toward Trustworthy and Stroke-Consistent Chinese Ancient Character Image Restoration via Edge-guided and Patch Correlation Feature Completion Network

Chinese ancient character image restoration is a challenging interdisciplinary vision task that plays an important role in digital archaeology and intelligent cultural heritage analysis, where reliable restoration quality is essential for trustworthy downstream interpretation. Over time, these ancient character images suffer from various forms of degradation that compromise stroke clarity and legibility, increasing the risk of unreliable character interpretation. Most existing methods employ pixel-wise regression or structure-agnostic generative inpainting models and neglect the reconstruction of edge details, leading to unclear strokes in the restored images. To address these issues, we propose a novel Edge-guided and Patch Correlation Feature Completion Network (EPCF-Net) for trustworthy and high-fidelity restoration of Chinese ancient character images. Firstly, we construct an Edge Semantic Prior Network (ESPN) to predict multi-scale edge priors within a semantic encoder, where edge-aware supervision derived from a pre-trained edge detector guides the network toward structurally reliable stroke reconstruction. Secondly, we propose an Edge-aware Feature Transform Module (EFTM), which fuses predicted edge features with image features to mitigate stroke distortion during restoration. Thirdly, we introduce a Patch Correlation Feature Completion Module (PCFCM) that leverages similar patch features of intact strokes to infer the features of degraded strokes, while employing multi-scale feature aggregation to restore fine structural details of ancient character edges. Extensive experiments on the ARMCD dataset show that EPCF-Net achieves the highest PSNR (25.12) and SSIM (0.9983), alongside the lowest MAE (0.0139), FID (12.76), and LPIPS (0.0468), surpassing state-of-the-art methods and demonstrating its effectiveness in restoring trustworthy and stroke-consistent Chinese ancient characters.