Abstract:

In this study, Yolo detection models were established based on the image production dataset of citrus fruits with unilateral complete canopy in natural state to provide theoretical and practical basis for the identification and localization of citrus fruits. After the citrus entered the fruit turning color period, the tree photos of unilateral canopy were collected randomly in Citrus unshiu Marc.cv. Miyagawa-wase orchard using camera, with a total of 800 images collected. Mosaic data augmentation techniques were used to expand to 2 000 images. The Yolov5, Yolov7, and Yolov8 algorithms were used to build target detection models for citrus fruits, respectively. The training loss values (loss) of different Yolo models all decreased rapidly in the first 15 rounds, followed by rapid convergence of the loss values of Yolov5 and Yolov7, while the loss value of Yolov8 showed a slow decreasing trend, the mean accuracy (mAP) of Yolov5 and Yolov8 reached the highest value of 91.9% and 92.5% in 49 and 29 iterations, respectively. But the mAP of Yolov7 fluctuated and increased to 300 rounds, reaching a maximum of 91.2%. The model prediction performance was evaluated on the plots of the unlabeled canopy group and fruit-bearing shoots group, respectively, with the Yolov7 model exhibiting the best comprehensive predictive performance. The fruit target detection model established with citrus canopy images in natural state has excellent recognition ability, and this study provides an important reference for the real-time detection of fruits in complex environments with target detection technology.

Key words: target detection, Citrus unshiu Marc.cv. Miyagawa-wase, Yolo model