Peach fruit thinning image detection based on improved YOLOv8 and data enhancement techniques

Abstract

Fruit thinning is a crucial technical aspect of fruit tree cultivation; it can enhance the fruiting rate and quality of fruit trees while also addressing seasonal, labor-intensive, and other characteristics. Currently, peach fruit thinning is predominantly conducted manually. However, robotic fruit thinning is an inevitable trend in the modern orchard industry. The initial step in robotic fruit thinning is the detection of young fruit. The detection of young peach fruit currently faces significant challenges due to their dense overlapping growth, frequent obscuration by leaf shade, and considerable variation in size and color at the early growth stage. This study proposes using an enhanced YOLO(You Only Look Once) model and image enhancement techniques to address these challenges and facilitate the detection of young peaches during the fruit thinning stage. First, the FasterNet network is designed to reconstruct the backbone network of YOLOv8, which makes the network more lightweight and helps to improve the target feature extraction ability in a cluttered environment. Then, a weighted BiFPN(bidirectional feature pyramid mechanism)is introduced in the feature fusion process for the presence of multi-scale targets. By deleting the less efficient feature transmission nodes, more efficient feature fusion is realized, which improves the fusion efficiency for different scale features. Finally, the Wise-IoU(Intersection over Union)v3+MPDIoU(Minimum Point Distance Intersection over Union) loss function is applied to replace the original loss function. This mitigates the harmful gradient caused by low-quality anchor frames and improves the model’s localization ability for densely occluded targets. The Young Peach datasets were collected under various natural conditions, and data enhancement was performed by deflation and cutting. The ablation and comparison experiments conducted validated the research hypothesis, and the average mAP(mean Average Precision) reached 98.4% at an IoU(Intersection over Union) threshold of 0.5. The F1 score, mAP at 0.5, and mAP at 0.95 all showed significant improvements, which were 11.8%, 14.9%, and 33% higher than the results of the directly collected annotated dataset and the standard YOLOv8 model, respectively. The method is an effective means of meeting the accuracy and robustness requirements of vision systems in the detection of young peaches.