A mural image color restoration method based on multi-scale convolutional image enhancement

Dunhuang murals, a treasure of ancient Chinese art, have suffered color fading due to long-term degradation. This paper proposes a novel mural image color restoration method using multi-scale convolutional image enhancement. First, an improved Se-MobileNet encoder is employed to extract deep features while reducing computational cost. Then, a multi-scale perceptual convolution is introduced to capture both local and global features, thereby enhancing texture representation. Subsequently, a feature-statistics-matching color transfer module is applied to ensure semantic consistency between degraded and restored images. Finally, a hybrid structure module integrates multi-scale features to further improve color consistency with reference images. Experiments show the proposed method outperforms S2WAT and AesUST, achieving an average improvement of 0.25 in SSIM and 3.42 in PIQE. The approach produces high-quality restoration results while preserving texture details, offering strong potential for digital mural preservation and cultural heritage protection.