Hyperspectral/Multispectral Reflectance Imaging Combining with Watershed Segmentation Algorithm for Detection of Early Bruises on Apples with Different Peel Colors

Abstract

Bruise damage on apples is one of the most key internal quality factors, which needs to be detected in postharvest quality sorting processes. However, detection of bruises is always a challenging issue. This study proposes a useful strategy for detection of early bruises on apples with different peel colors based on Vis-NIR hyperspectral/multispectral reflectance imaging combining with watershed segmentation algorithm, which has never been tried in the past. Three spectral regions, namely visible and near-infrared (Vis-NIR) (450–685 nm and 750–1000 nm), visible (Vis) (450–685 nm), and near-infrared (NIR) (750–1000 nm), were selected for principal component analysis (PCA) to identify the optimal region and principal component (PC) score image, respectively. The third PC score image (PC3) obtained by using PCA of the NIR spectra was found to be the most useful for detection of bruise damage on apples. Three important wavelength images at 786, 915, and 995 nm were further identified from data dimension reduction by weighting coefficient analysis of all sub-images of PC3 score images. Finally, bruise detection based on both the second PC score image obtained from multispectral PCA and the proposed improved watershed segmentation algorithm was performed to classify all 210 samples with three kinds of peel color including green peel, middle-color peel (green-red), and red peel (dark red); a 99.5% overall detection accuracy (99.2% for 120 bruised samples and 100% for 90 sound samples) was obtained, indicating feasibility of this study. The finding is significant because the study of apples with different surface colors was closer to the actual production of fruit sorting.