Comparative Evaluation of YOLOv11 and U-Net for Automated Fruit Detection, Quality Classification, and Defect Segmentation

Abstract

Automated fruit monitoring can support postharvest grading, quality control, and precision horticulture, but most computer-vision studies address detection, classification, or segmentation as separate tasks. This study compared pretrained YOLOv11 and U-Net-based frameworks under a common experimental setting for fruit-type detection, three-class quality classification (fresh, mildly defective, and rotten), and pixel-level segmentation of defective regions. The FruQ-Multi and FruQ-DB datasets were used, comprising 9,421 and 5,647 images, respectively, across 11 fruit types. Images were resized, normalized, augmented within the training subset, and divided into training, validation, and test subsets at the sample level. Performance was assessed using accuracy, precision, recall, F1-score, mean average precision (mAP), and mean intersection over union (mIoU). For fruit detection, YOLOv11 achieved an average accuracy of 97.4%, precision of 98.0%, recall of 97.9%, mAP@0.50 of 95.5%, and mAP@0.50:0.95 of 87.4%, whereas the U-Net-based model achieved 99% accuracy, 98% precision, and 99% recall. YOLOv11 provided the strongest quality-classification performance, with 100% reported accuracy and approximately 98% precision and recall. For defect segmentation, U-Net achieved 97% precision, 99% recall, an F1-score of 98%, mIoU of 97.2%, and mAP of 91.6%. The results indicate that YOLOv11 is particularly effective for detection and quality classification, whereas U-Net is more suitable for precise defect delineation. Task-specific model selection may therefore improve automated fruit-quality monitoring and postharvest decision support.