Authors

Abstract

A chunking strategy based on YOLOv12n has been proposed in this study to address challenges of young fruit detection for entire trees in dwarf and densely planted apple orchards, such as small young fruit volume, high similarity between fruit skin color and surrounding young leaves. This strategy divides an image into sub-images before feeding them into the network for training. Experimental results show that, compared with using only the YOLOv12n model for whole-image recognition, the chunking strategy achieved a precision of 85.57% and a recall of 62.54% in young fruit detection. In contrast, the YOLOv12n model alone, due to ineffective small-target capture and missing edge features, achieved a detection precision of less than 1% for young fruits, making effective detection nearly impossible. The introduction of the chunking strategy achieved a qualitative leap in young fruit detection performance. This strategy provides technical support for precise young fruit counting, subsequent accurate thinning, and growth dynamic monitoring in dwarf and densely planted orchards, and also offers a new implementation approach for small-target detection scenarios in orchards.

Abstract in Chinese

本研究提出了一种基于YOLOv12n的分块策略，以解决矮化密植苹果园整树幼果检测的难题，如幼果体积小、果皮颜色与周围幼叶高度相似等。该策略通过将图像分割为子图像后输入网络进行训练。试验结果表明，与仅使用YOLOv12n模型进行整图识别，分块策略的使用在幼果识别的准确率上达到85.57%，召回率达到62.54%，而单独使用YOLOv12n模型则因小目标捕捉失效、边缘特征遗漏，对幼果的检测准确率不足1%，几乎无法完成有效检测，分块策略的引入实现了幼果识别性能的质的飞跃。该策略为矮砧密植果园幼果精准计数、后续精准疏果及生长动态监测等环节提供了技术支撑，也为果园小目标检测场景提供了新的实现思路。