Enhanced detection of early bruises in apples using near-infrared hyperspectral imaging with geometrical influence correction for universal size adaptation

Near-infrared (NIR) imaging is effective in monitoring the optical property changes of fruit arising from mechanical damage. However, differences in fruit geometry and size severely limit the application of bruise detection solutions. By integrating NIR hyperspectral imaging (NIR-HSI) with geometrical influence correction (GIC), this paper presents a universal bruise enhancement and detection method for early-stage bruises inspection across apple cultivars with large size variations. HSI and shape data were collected via 360° rotational scanning of Sun Fuji, Shinano Sweet, and Esopus Spitzenburg apples before and during the first 24 h post-bruising. GIC was applied as a pretreatment method. For comparison, we applied whiteboard reflectance calibration (WRC) and WRC combined with the standard normal variate (SNV) approach. Using principal component analysis (PCA), a set of effective wavelength-loading coefficients for bruise enhancement was extracted across pooled datasets of average sound and bruise spectra from different samples. The optimal coefficients, determined using logistic regression, were applied uniformly across all HSI datasets for bruise enhancement. Finally, the local Otsu method combined with connected-domain screening was applied for bruise identification. Based on spectral analysis, PCA successfully extracted bruise-related wavelength coefficients with consistent trends across cultivars, facilitating universal bruise enhancement. GIC reduced shape-related interference, improving the use of the light scattering-related PC for bruise identification. GIC coupled with the universal enhancement emerged as the most effective method, consistently achieving the highest classification accuracy, superior identification accuracies for both central and edge bruises, and the earliest peak accuracy.