Ancient murals often suffer from degradation, compromising their cultural significance. To address this, we propose AFMNet, a lightweight network based on a multi-head attention mechanism for high-quality mural completion. AFMNet effectively integrates multi-scale features and captures long-range dependencies, enhancing semantic consistency. Extensive experiments on a custom mural dataset and the Places2 benchmark demonstrate its superiority over competitive methods. Quantitatively, our method achieves a Peak Signal-to-Noise Ratio (PSNR) of 34.42 dB on murals with low degradation (1-10% missing area) and reduces the parameter count by 34.4% compared to the baseline, showcasing its high efficiency. The results exhibit excellent visual fidelity with fewer artifacts, providing a valuable tool for the digital preservation of ancient murals.
Hu et al. (Mon,) studied this question.