Algorithm for crack detection of terracotta warriors based on fusion of multi-network hyperspectral images

Terracotta warriors possess significant historical research value. Affected by environmental and human factors, they often exhibit varying degrees of defects, with cracks being the most common and having a high rate of missed detection, which leads to oversight in restoration processes. Hyperspectral images can simultaneously reflect material composition information and penetrate the shallow surface layer, enabling crack detection in this layer. This used hyperspectral methods to obtain information on terracotta warriors as follows: 1) Constructed a RGB and spectral database of terracotta warriors; 2) Simulated visual perception, focused on RGB spectral information, proposed a Visual Geometry Group (VGG) algorithm based on attention mechanisms to locate regions of cracks; 3) Focused on hyperspectral information, proposed an algorithm based on improved DeepLabv3+ to acquire shallow surface information. And VGG algorithm based on attention mechanisms and the improved DeepLabv3+ algorithm were fused to achieve crack detection. To validate the effectiveness of the algorithm, the research constructed a dataset of 300 groups containing cracks using hyperspectral data, and archaeologists annotated the cracks using visual and reverse confirmation methods as the ground truth. Experiments show that Mean Pixel Accuracy (MPA) of the proposed algorithm reaches 90.1 %, which is 19.5 % higher than that of the traditional Canny algorithm. This approach can assist conservation personnel in providing a basis for the protection and restoration of terracotta warriors.