With the advancement of digital technologies, accurate color extraction and restoration in ceramic image processing have become pivotal in preserving cultural heritage and enhancing digital design applications. This study presents a novel ceramic image denoising framework based on a deep learning algorithm tailored for digital color enhancement. The proposed architecture integrates a densely connected residual network, a contrast-based graph generator, and an improved adaptive light scattering model. The residual modules capture multi-scale texture features, while the graph generator enhances structural detail learning by incorporating spatial saliency. The light scattering model adaptively fuses texture and color information to restore high-fidelity ceramic images. Extensive experiments on the Chinese Neolithic painted pottery and ARCA328 datasets show that the proposed method consistently surpasses prior approaches. Compared with the strongest baseline (PFF-Net), it achieves a 51.7% improvement in PSNR, a 31% increase in SSIM, as well as a 27.8% reduction in MSE and a 40.6% reduction in MAE. Furthermore, applying the method as a preprocessing step for segmentation tasks notably increases segmentation accuracy, highlighting its practical value in downstream ceramic image analysis. This work provides a robust foundation for intelligent ceramic modeling and digital preservation through enhanced visual quality.
Zhi Fu (Sat,) studied this question.