Lightweight pear detection in unstructured orchards via selective information propagation

Abstract

Accurate pear detection in unstructured orchards is important for robotic harvesting and orchard perception. However, pear detection poses compound challenges that differ from those in more chromatically distinctive fruits: mature pears share yellow-green hues with surrounding foliage, their near-spherical geometry offers limited contour priors, and they typically grow in tight spur clusters where mutual boundary occlusion occurs even without branch interference. Under these coupled degradations, lightweight detectors tend to lose accuracy and become difficult to deploy on embedded agricultural platforms. To address this issue, we propose a lightweight pear detection framework guided by the principle of selective information propagation—the idea that, under tight computational budgets, how information is routed at each stage matters more than overall network capacity. The framework instantiates this principle along four stages of the detection pipeline through dedicated modules for efficient global–local context modeling, input-adaptive feature transformation, detail-preserving multiscale fusion, and an adaptive IoU loss tailored for small and occluded fruits. On the self-built Orchard Pear dataset, the proposed method achieves 95.2% mAP@50 and 54.6% mAP@50:95 with only 2.56 M parameters and 5.60 GFLOPs. Consistent improvements are also observed on the public Minne Apple and Mango datasets. Deployment experiments on embedded platforms further show that the proposed method supports real-time inference for agricultural robotic applications. These results suggest that selective feature representation, fusion, and optimization benefit lightweight fruit detection in complex orchard scenes.