Paper cultural relics are important for heritage transmission as they record human history and humanities in different periods. However, they are highly susceptible to microorganisms such as mold during preservation. Mold can accelerate the degradation of cellulose, generating mold on the surface of paper. Scattered spores can spread widely with airflow, increasing the risk of mold on other paper cultural relics. Regular mold spot detection is crucial for understanding paper artifacts' status and restoration. Hyperspectral imaging technology is a non-contact and non-destructive detection method that simultaneously obtains spatial and spectral data. This technology can be combined with computer technology to enable large batches of real-time, non-destructive testing of paper cultural relics. This paper proposes a method for reducing the dimensionality of hyperspectral data for Aspergillus niger, a commonly occurring mold. The method is based on the attention mechanism and allows for adaptive preprocessing of hyperspectral redundant data. This paper reports on the collection of 20 samples of Aspergillus niger, mold spots on paper artifacts provided by the Chongqing China Three Gorges Museum. The average spectral curves of the infected and healthy areas are analyzed using ENVI software in the 413~855 and 855~1 021nm bands. The results showed a significant difference in average reflectance between the two areas. The paper compares the proposed method with traditional principal component analysis and independent component analysis preprocessing methods for processing original hyperspectral data. The results are then experimented on four semantic segmentation networks: classical U-Net, SegNet, DeepLabV3+, and PSPNet. The experimental results demonstrate that the preprocessed data produced by the algorithm presented in this paper exhibit significant advantages over the classical U-Net and SegNet networks. Furthermore, compared to the principal component analysis method and independent component analysis method, the accuracy of mold spot identification has improved significantly by 89.49% and 88.46%, respectively. These results confirm the effectiveness of the proposed algorithm and provide valuable support and new ideas for the field of cultural relics protection.