Detection of Nectarines on Trees with YOLO Models

In agricultural production, the introduction of artificial intelligence can speed up the detection of fruit and yield statistics. To address the problem that nectarines are not easily detected due to their uneven color, easy exposure, and leaf shading in natural picking environments, our research proposes a YOLO-based nectarine detection network. Based on the datasets with data augmentation, we use the YOLO family of deep learning models to train on the collected nectarine data, compare different pre-trained models for single species target detection of nectarines, compare the results of each model, and find the model that can accurately detect nectarines. In this experiment, Precision (accuracy), Recall (recall), F-Measure (F1) and mean Average Precision (mAP) were used as accuracy metrics for the models. The experimental results showed that the weight files obtained from training with the YOLOv3 network were relatively stable, with 98.1% accuracy, 95% recall, 96.5% F1 value, and 98.1% mAP. Compared with other networks, the precision, recall, and F1 value were on average 1.275% higher and the mAP was 4.85% higher. The model was found to maintain high accuracy and performance in the detection of nectarines in a natural picking environment, providing technical support for the rapid, accurate, and stable detection and counting of nectarines by artificial intelligence in complex natural environments.