Abstract
Pears are one of the most widely consumed fruits, and their quality directly impacts consumer satisfaction. Surface defects, such as black spots and minor blemishes, are crucial indicators of pear quality, but it is still challenging to detect them due to the similarity in visual features. This study presents PearSurfaceDefects, a self-constructed dataset, containing 13,915 images across six categories, with 66,189 bounding box annotations. These images were captured using a custom-built image acquisition platform. A comprehensive novel benchmark of 27 state-of-the-art YOLO object detectors of seven versions Scaled-YOLOv4, YOLOR, YOLOv5, YOLOv6, YOLOv7, YOLOv8, and YOLOv9,has been established on the dataset. To further ensure the comprehensiveness of the evaluation, three advanced non YOLO object detection models, T-DETR, RT-DERTV2, and D-FINE, were also included. Through experiments, it was found that the detection accuracy of YOLOv4-P7 at mAP@0.5 reached 73.20%, and YOLOv5n and YOLOv6n also show great potential for real-time pear surface defect detection, and data augmentation can further improve the accuracy of pear surface defect detection. The pear surface defect detection dataset and software program code for model benchmarking in this study are both public, which will not only promote future research on pear surface defect detection and grading, but also provide valuable resources and reference for other fruit big data and similar research.