Lightweight mural image color restoration based on multi-reference semantic fusion

In the field of mural heritage preservation and exhibition, virtual color restoration of faded mural images has increasingly become a focal research area. Addressing the challenges in color restoration, such as the inability of a single reference image to provide accurate color references and the inefficiency of existing methods in processing ultra-high-resolution mural images, this paper proposes a lightweight color restoration method for mural images based on multi-reference semantic fusion. The method first employs two micro-encoders to extract the content features of the faded mural image and the color features from multiple reference images, respectively. Then, a dual modulation module is introduced to convert the multi-reference features into color modulation information, thereby providing rich and effective color information for the restoration process. Subsequently, a multi-reference semantic fusion module based on spatial attention is constructed to integrate the multi-reference features. Finally, a visual enhancement module based on Robust Retinex is utilized to optimize the color and content details of the intermediate restored image while suppressing high-frequency noise. Experimental results demonstrate that the proposed method can quickly and accurately restore the colors of faded mural images while effectively preserving their content structure and texture details. Moreover, it performs well when processing ultra-high-resolution images. Compared to the baseline methods, the proposed method improves BRISQUE, PIQE, and LPIPS evaluation metrics by 5.4%, 14.2%, and 7.1%, respectively, increases processing speed by 4-8 times, and reduces the number of parameters by 5-130 times.