Deep learning approaches for bruised mandarin orange classification by fluorescence hyperspectral imaging

초록

Citrus fruit is extensively consumed worldwide, and the bruising of these fruit significantly affects their quality, impacting consumers' purchasing decisions. Detecting such damage during post-harvest operations is crucial. However, the bruised area does not exhibit distinct color differences compared with normal regions, complicating visual inspection and rendering it time-consuming. Consequently, this study investigates the potential of fluorescence hyperspectral imaging to discern bruised mandarin oranges. Hyperspectral images were acquired following illumination with a pair of 365 nm UV lights. Three multivariate data analyses-decision tree, support vector machine, and partial least squares discriminant analysis-and three deep learning models-ResNet50, EfficientNetB0, and MobileNet-were employed for the classification of bruised mandarins. Preprocessing steps, including dark and white correction, spectra preprocessing, and region of interest (ROI) extraction, were conducted prior to model development. Classification accuracy was determined through model training. Among the models, ResNet50 with nine principal component images exhibited high classification accuracy: 99.65 % for the training group, 100 % for the validation group, and 100 % for the test group. GradCAM visualization further confirmed the successful formation of heatmap over bruised areas. Thus, the classification of bruised mandarins using fluorescence hyperspectral imaging and deep learning is feasible, laying the groundwork for automated sorting technologies.

키워드