A Multi-Strategy Framework for Enhancing Shatian Pomelo Detection in Real-World Orchards

Shatian pomelo detection in orchards is essential for yield estimation and lean production, but models tuned to ideal datasets often degrade in practice due to device-dependent tone shifts, illumination changes, large scale variation, and frequent occlusion. We introduce STP-AgriData, a multi-scenario dataset combining real-orchard imagery with curated web images, and apply contrast/brightness augmentations to emulate unstable lighting. To better address scale and occlusion, we propose REAS-Det, featuring Global-Selective Visibility Convolution (GSV-Conv) that expands the visible feature space under global semantic guidance while retaining efficient spatial aggregation, plus C3RFEM, MultiSEAM, and Soft-NMS for refined separation and localization. On STP-AgriData, REAS-Det achieves 86.5% precision, 77.2% recall, 84.3% mAP@0.50, and 53.6% mAP@0.50:0.95, outperforming recent detectors and improving robustness in real orchard environments. The source code is available at: https://github.com/Genk641/REAS-Det.

💡 Research Summary

This paper addresses the practical challenges of detecting Shatian pomelos in real‑world orchards, where variations in imaging devices, illumination, object scale, and occlusion severely degrade the performance of models trained on ideal datasets. To bridge this gap, the authors introduce a comprehensive multi‑strategy framework comprising (1) a novel multi‑scenario dataset (STP‑AgriData) that blends 150 high‑resolution orchard images captured with a Hikvision camera and 167 diverse web‑sourced images, thereby covering a wide range of device‑dependent tone shifts and environmental conditions; (2) extensive data augmentation (random flips, grayscale conversion, noise injection, contrast and brightness adjustments) that expands the training set to 1,330 images and simulates unstable lighting; and (3) a new detection architecture named REAS‑Det built on YOLOv8.

The core technical contribution is the Global‑Selective Visibility Convolution (GSV‑Conv). Unlike conventional convolutions that aggregate features only within a fixed local receptive field, GSV‑Conv first reorders the feature map based on global semantic similarity, making distant but semantically related locations “visible” before spatial aggregation. This non‑causal global modeling enables long‑range context integration without the computational overhead of full self‑attention, while preserving the inductive bias and efficiency of standard convolutions.

To complement GSV‑Conv, the authors add three auxiliary modules: (i) Composite Receptive Field Enhancement Module (C3RFEM), which combines dilated convolutions and multi‑path structures to enlarge the receptive field without sacrificing fine‑grained texture details; (ii) Multi‑Scale Multi‑Head Feature Selection (MultiSEAM), a dynamic selector that fuses features from multiple scales and heads, thereby strengthening representations of heavily occluded fruits; and (iii) Soft‑NMS, a softened non‑maximum suppression that reduces false suppression in densely clustered pomelo groups.

Experimental evaluation on STP‑AgriData shows that REAS‑Det achieves 86.5 % precision, 77.2 % recall, 84.3 % mAP@0.50, and 53.6 % mAP@0.50:0.95, outperforming recent state‑of‑the‑art detectors such as YOLOv5‑s, YOLOv8‑m, and Faster‑RCNN‑ResNet101 by 3–7 % absolute mAP. Ablation studies confirm the individual impact of each component: GSV‑Conv alone yields a 3.2 % mAP gain; adding C3RFEM contributes another 1.8 %; MultiSEAM adds 1.1 %; and Soft‑NMS provides the final boost to reach the reported performance. Qualitative results illustrate robust detection under low‑light, extreme scale variance, and heavy foliage occlusion.

The paper also discusses limitations: the dataset is currently limited to the Shatian pomelo variety grown in a specific Chinese region, so cross‑regional or cross‑species generalization remains to be validated; and the global reordering step of GSV‑Conv increases memory consumption, which may hinder deployment on ultra‑light edge devices. Future work is suggested to expand the dataset to other citrus varieties, explore more memory‑efficient global visibility mechanisms, and integrate lightweight backbones for real‑time robotic harvesting.

Overall, the study presents a well‑structured solution that combines data‑centric strategies with novel architectural innovations to substantially improve fruit detection robustness in complex, real‑world agricultural settings.