Authors

Abstract

Accurate and efficient rice-pest detection is essential for field scouting, early warning and precision control. This study presents a lightweight YOLOv11n-based framework for accurate edge deployment. It integrates a Multi-cognitive Visual Adapter (Mona), a Detail-Preserving Contextual Fusion module (DPCF) and Wise-IoU-based non-maximum suppression (Wise-IoU-NMS). Mona improves multi-scale feature representation with limited overhead; DPCF preserves fine-grained pest details during contextual fusion; and Wise-IoU-NMS stabilizes localization of overlapping and densely distributed targets. On a dataset assembled from 5,212 publicly available images representing six near-balanced rice-pest classes and divided 8:1:1 for training, validation and testing, the proposed model achieved 95.4% mAP@50, 94.8% precision and 94.7% recall with 2.65 M parameters and 6.5 GFLOPs. A confidence threshold of 0.476 was adopted for alarms. In a Jetson Nano proof-of-concept using TensorRT FP32 at 640 x 640 input resolution, the model reached 36 FPS, and the six displayed device-side detections had confidence scores of 0.88-0.95. These results demonstrate real-time, threshold-based logging for field monitoring; integration with selective spraying remains future work.

Abstract in English

水稻害虫的快速、准确检测对于田间巡检、虫情预警和精准防控具有重要意义。本文提出一种面向边缘端高精度部署的轻量化 YOLOv11n 检测框架，融合多认知视觉适配器（Multi-cognitive Visual Adapter, Mona）、细节保留型上下文融合模块（Detail-Preserving Contextual Fusion, DPCF）和基于 Wise-IoU 的非极大值抑制策略（Wise-IoU-NMS）。Mona 以较低开销增强多尺度特征表达，DPCF 在上下文融合中保留细粒度虫体信息，Wise-IoU-NMS 则提高重叠及密集目标的定位稳定性。在由 5,212 张公开可获取图像构建、包含 6 类近均衡水稻害虫并按 8:1:1 划分训练集、验证集和测试集的数据集上，所提出模型取得 95.4% 的 mAP@50、94.8% 的精确率和 94.7% 的召回率，参数量为 2.65 M，计算量为 6.5 GFLOPs。报警阈值设为 0.476；在采用 TensorRT FP32、输入分辨率为 640 x 640 的 Jetson Nano 概念验证中，模型达到 36 FPS，设备端展示的 6 次检测置信度为 0.88-0.95，验证了面向田间监测的实时阈值报警与结果记录流程，而与选择性施药设备的闭环联动仍需后续验证。