Background: Image-based apple fruit disease classification can support consistent inspection, but its performance may depend on preprocessing and class imbalance. Objective: This study examined whether cropping or region of interest (ROI) segmentation improves MobileNetV2 classification of Blotch, Normal, Rot, and Scab apple images. Method: A comparative experiment evaluated five preprocessing scenarios (original images, center cropping, HSV ROI segmentation, GrabCut ROI segmentation, and hybrid ROI plus cropping) under identical training settings. The 5,246 cleaned images were split into 3,672 training, 787 validation, and 787 test images and evaluated using accuracy, macro-precision, macro-recall, macro-F1, per-class F1-scores, and a confusion matrix. Results: In a single run, center cropping produced the highest scores (85.01% accuracy; 84.91% macro-F1), followed by the original images (83.86%; 83.64%), with overlapping 95% confidence intervals. HSV, GrabCut, and hybrid preprocessing did not exceed the baseline. Scab was the most difficult class, and Rot and Scab were most often misclassified as Blotch. Conclusion: Complex preprocessing did not necessarily improve classification; center cropping is an efficient option for lightweight apple inspection systems, pending confirmation through repeated runs.
Copyrights © 2025