An Automatic Pigment Recognition Method For Faded Mural Hyperspectral Images Based On Attention-CNN

Murals are one of the most significant category of cultural heritage, serving as important materials for the study of ancient society, history, and religion. Currently, due to various environmental and human factors, many murals are experiencing different types and degrees of deterioration, such as fading issues that make important content on the murals invisible under visible light. While hyperspectral imaging (HSI) offers new opportunities for analyzing faded murals, the processing challenges of HSI data and lack of domain knowledge integration in mineral pigment analysis limit its effectiveness. Therefore, based on hyperspectral data, this study proposes a lightweight deep learning model — Attention-CNN— for the automatic identification of mineral pigments in faded murals. By introducing a spectral attention mechanism, the model is able to better focus on characteristic bands. The incorporation of band pyramid pooling allows the model to hierarchically pool band data, thereby extracting multi-scale spectral features. Through comparative validation with three traditional machine learning models, it is demonstrated that this algorithm can efficiently classify and identify different pigment regions in mural hyperspectral images with relatively high accuracy of 0.9970 (training set) and 0.9976 (test set), respectively.