Authors

Abstract

Cucurbitaceae is widely planted and its fruits have great economic value. Object detection is one of the key aspects of cucurbit harvesting. In this paper, four models, YOLOv3, YOLOv4, YOLOv5s and improved Resnet_YOLO, were used to detect mixed bitter melon, cucumber, white melon, and "Boyang 9" melon fruits. Fruit images of bitter melon, cucumber, white melon and "Boyang 9" melon were collected under different natural conditions for model training. The results showed that "Boyang 9" melon had the best overall detection results among the four cucurbit species, with the highest AP and F1, 0.99 and 0.94 respectively. The YOLOv5s model performed best among the four models: the best weights size was the smallest at 14MB; the better mAP value of 0.971 for all classes of cucurbits; and the fastest detection speed with fps of 90.9 . This paper shows that four types of cucurbit fruit images, bitter melon, cucumber, white melon, and "Boyang 9" melon, can be detected based on deep learning methods for hybrid detection.

Abstract in Chinese

葫芦科种植面积广泛，其果实具有巨大的经济价值。目标检测是葫芦科采摘的关键环节之一。本人采用YOLOv3、YOLOv4、YOLOv5s和改进的Resnet_YOLO 四种模型对苦瓜、黄瓜、白皮甜瓜、 “博洋9”甜瓜果实进行混合检测。采集不同自然情况下的苦瓜、黄瓜、白皮甜瓜、“博洋9”甜瓜的果实图像进行模型训练。结果表明，在YOLOv3、YOLOv4、YOLO5、Resnet_YOLO四种模型检测苦瓜、黄瓜、白皮香瓜和“博洋9号”甜瓜时，”博洋9号”甜瓜检测结果整体最优，AP和F1最高，分别为0.99、0.94；YOLOv5模型表现最优：最佳权重内存最小，为14MB； 具有很好的mAP值，达到了0.971；检测速度最快，fps为90.9 。此方法表明苦瓜、黄瓜、白皮香瓜、“博洋9”甜瓜四类葫芦科果实图像可进行混合检测。