AHG-YOLO: multi-category detection for occluded pear fruits in complex orchard scenes

To achieve fast detection of pear fruits in natural pear orchards and optimize path planning for harvesting robots, this study proposes the AHG-YOLO model for multi-category detection of pear fruit occlusion in complex orchard environments.

Using the Red Delicious pear as the research object, the pears are classified into three categories based on different occlusion statuses: non-occluded fruits (NO), fruits occluded by leaves/branches (OBL), and fruits in close contact with other fruits but not obstructed by leaves/branches (FCC). The YOLOv11n model is used as the base model for a lightweight design. First, the sampling method in the backbone and neck networks is replaced with ADown downsampling to capture higher-level image features, reducing floating-point operations and computational complexity. Next, shared weight parameters are introduced in the head network, and group convolution is applied to achieve a lightweight detection head. Finally, the boundary box loss function is changed to Generalized Intersection over Union (GIoU), improving the model's convergence speed and further enhancing detection performance.

Experimental results show that the AHG-YOLO model achieves 93.5% (FCC), 95.3% (NO), and 93.4% (OBL) in AP, with an mAP@0.5 of 94.1% across all categories. Compared to the base YOLOv11n network, precision, recall, mAP@0.5, and mAP@0.5:0.95 are improved by 2.5%, 3.6%, 2.3%, and 2.6%, respectively. The model size is only 5.1MB, with a 16.9% reduction in the number of parameters.

The improved model demonstrates enhanced suitability for deployment on pear-harvesting embedded devices, providing technical support for the path planning of fruit-picking robotic arms.