Early Bruise Detection in Apples Based on <scp>NIR</scp> Hyperspectral Characteristic Wavelength Images

ABSTRACT Apples are susceptible to damage during harvesting, packaging, and transportation. Early bruises are difficult to distinguish from normal tissue, making manual sorting difficult. To address this challenge, this study proposed an early bruise detection method for apples based on hyperspectral characteristic wavelength images, an adaptive threshold segmentation algorithm, and a deep learning model. NIR hyperspectral images of normal apples and apples with bruises (single/multiple bruises on fruit, stem, calyx surfaces) were captured. By combining the weight coefficients of the best principal component image with spectral features, three characteristic wavelengths were selected, and secondary PCA obtained the optimal principal component image. Using the designed adaptive threshold segmentation based on PCA image combined with Hough circle detection, effective detection of early bruises with varying severity, positions, and numbers was achieved. Comparing the Otsu threshold, fixed threshold, and the proposed adaptive threshold segmentation algorithm, the results showed that the proposed adaptive threshold segmentation algorithm significantly outperformed the Otsu and fixed threshold methods in terms of F1‐score (96.18%). Compared to full‐wavelength images, the detection of early‐stage bruising on apples using images based on three characteristic wavelengths demonstrated a significant speed advantage, validating the superiority of the adaptive threshold segmentation algorithm. Additionally, to further validate the effectiveness of characteristic wavelength images and improve detection accuracy, apple bruising detection based on the deep learning‐based YOLOv11 algorithm was also implemented, achieving an F1 score of 99.94% for bruise detection. This study provides technical support for the rapid and non‐destructive detection of early apple bruises and has potential application value.