Application of Multispectral Image Analysis Based on Transformer Model in the Restoration of Fine Fiber Materials in Ancient Murals

Accurate identification of micro-damage in fiber-based mural materials is essential for quantitative restoration and non-destructive cultural heritage conservation. Since multispectral imaging exploits the interaction between electromagnetic waves and material surfaces to reveal spectral responses that are invisible in conventional imaging, it provides an effective foundation for fine structural analysis. This study proposes a Transformer-based multispectral image analysis framework for the restoration of ancient mural fiber materials. The method constructs a spectral–spatial joint representation through channel-level fusion and positional encoding and employs multi-head self-attention to capture weak cross-band fracture features and long-range contextual dependencies. An end-to-end decoding strategy with cross-scale feature fusion is further designed to achieve pixel-level segmentation and quantitative characterization of fiber microcracks. Experiments conducted on 128 multispectral mural images demonstrate an Intersection over Union of 90.2% for longitudinal crack detection, a damage density estimation error of 0.05, and over 88.3% of crack width predictions with an absolute error below 0.15 mm. The proposed framework enables reliable nondestructive assessment of mural fiber degradation while providing a transferable spectral analysis strategy for electromagnetic imaging, intelligent material inspection, and related engineering applications involving multispectral sensing.