A Multi-Label Hyperspectral Tongue Coating Classification Network Based on Cross-Image Collaborative Learning and Second-Order Spectral Interaction.

Tongue diagnosis is valuable for disease diagnosis and therapeutic evaluation in traditional Chinese medicine, and automatic tongue image recognition provides objective clinical support. However, most existing methods rely on RGB images with limited spectral information. Although hyperspectral imaging provides richer spectral information, multi-label tongue coating classification still faces challenges including feature coupling caused by attribute co-occurrence and insufficient modeling of cross-band relationships in hyperspectral data. To address these challenges, we propose a hyperspectral tongue coating classification network (HSTIC-Net) for multi-label classification of tongue coating color, tongue coating thickness, and greasy coating. HSTIC-Net integrates Cross-Image Spectral-Spatial Collaborative Attention (CSCA) and Cross-Group Second-Order Spectral Interaction (GSC). CSCA exploits cross-image collaborative information from samples with shared attributes to enhance attribute-related spectral-spatial representations and alleviate feature entanglement caused by attribute co-occurrence. GSC introduces cross-group channel interaction and second-order spectral modeling to strengthen the representation of cross-band dependencies in high-dimensional hyperspectral features. HSTIC-Net outperforms mainstream methods in multi-label tongue coating classification, achieving an mAP of 85.30%, an OF1 of 87.22%, and a CF1 of 84.53%. Moreover, t-SNE and attention map analyses demonstrate improved feature separability and discriminative region localization. The results demonstrate that HSTIC-Net effectively improves hyperspectral tongue coating classification and provides a feasible approach for objective and intelligent tongue image analysis.