HSGAN-based near-infrared hyperspectral reconstruction from characteristic wavelengths images for apple bruise...

Hyperspectral images contain richer spectral and spatial information than multispectral images, yet traditional equipment suffers from limitations such as bulky size and complex data processing. This study constructs a task-oriented near-infrared (NIR) hyperspectral reconstruction and detection framework for the precise detection of early apple bruises. First, apple samples were collected using a hyperspectral imaging system. The Weight Extremum Method was employed to screen seven characteristic wavelengths (976.4 nm, 1064.8 nm, 1175.8 nm, 1192.5 nm, 1295.4 nm, 1449 nm, and 1631.6 nm), which were further reduced to three key wavelengths (1064.8 nm, 1175.8 nm, and 1449 nm). Based on the datasets constructed from these bands, the HSGAN framework was used as the reconstruction backbone to reconstruct hyperspectral images ranging from 866 nm to 1701 nm. Results demonstrated that reconstruction performance was optimal with seven input bands (PSNR = 37.81, SSIM = 0.973) and remained favorable with three bands (PSNR = 34.70, SSIM = 0.950). Finally, YOLOv11n was used to detect bruises on both original and reconstructed images. Detection accuracy using reconstructed spectra from the 7-band input approached that of the original images (mAP50 = 0.994), while the 3-band input also maintained high precision (mAP50 = 0.992, Recall = 0.993). These results demonstrate that reconstructing 254 NIR bands from just three characteristic wavelengths is feasible. This framework significantly reduces data acquisition costs while enabling high-precision early bruise detection, offering a practical solution for agricultural quality control.