Hyperspectral non-destructive detection based on geometrical influence correction and deep learning technology: A case study on early bruise detection in fruits

Hyperspectral imaging is widely used for non-destructive detection in fruit because it can capture spectral differences arising from changes in tissue structure and chemical composition. However, variations in the curved-surface geometry of fruit often introduce non-structural spectral distortions. These distortions reduce detection accuracy and limit model generalisation. To address the adverse effects of illumination–geometry–spectral coupling, this study proposes a fruit bruise detection framework integrating hyperspectral imaging with three-dimensional (3D) point cloud data through geometrical influence correction based on an intrinsic decomposition. In particular, pixel-level surface normals and lighting features are first extracted from 3D point clouds to jointly acquire spectral and geometric information. A dual-branch encoder–decoder network is then designed to incorporate geometric constraints from the point cloud. This network decouples the observed spectra into intrinsic reflectance and illumination components. Finally, a bruise detection model is constructed using intrinsic reflectance images, enabling pixel-level identification of bruised regions on the fruit surface. Experimental results demonstrate that the proposed method maintains spectral structural consistency while significantly enhancing the discriminability of bruised areas. Compared with models using raw spectral data, the proposed approach improves detection accuracy and increases the recall rate of bruised regions. The introduced framework offers a novel physics-based decoupling mechanism for hyperspectral detection and demonstrates clear advantages in suppressing non-structural spectral variations arising from varying fruit surface geometry.