Line-guided gated fourier cascade and multi-head attention network for mural restoration
摘要
Murals, as invaluable cultural treasures of China, face inevitable damage over time, necessitating advanced restoration techniques. Addressing critical challenges such as color bias and texture loss, this study proposes the Line-Guided Gated Fourier Cascade and Multi-Head Attention Network (LGF-MANet). By integrating gated convolutions with a UnifiedCoordGate (UCG) for both spatial and channel dimensions, and employing Multi-Frequency Unbiased (MFU) for precise feature extraction, LGF-MANet enhances image understanding in the frequency domain, effectively restoring texture structure. The Dynamic Attention Feature Aggregator (DAFA) further refines detail and complex color recovery. Experimental results show an average PSNR of 28.7dB and SSIM of 0.95, significantly outperforming existing methods. This research advances digital preservation by offering novel solutions for cultural heritage restoration.