基于YOLOV11-SMALL的轻量化脐橙病虫害检测研究

Research on Lightweight Detection of Pests and Diseases of Navel Oranges Based on YOLOV11-SMALL

摘要: 本文围绕脐橙果园病虫害智能识别的实际需求，开展基于深度学习的目标检测算法优化研究。通过构建包含复杂背景的脐橙病虫害图像数据集，提出一种轻量化改进模型YOLOv11-SMALL。该模型在YOLOv11n的基础上引入ADown下采样模块以降低参数与计算量，嵌入HGNetV2主干网络增强多尺度特征提取能力，并利用ASFF自适应空间特征融合机制提升小目标与复杂背景下的检测性能。实验表明，改进模型在准确率、mAP@0.5和mAP@0.5:0.95分别达到0.975、0.971和0.912，参数量仅1728K，模型大小3.6M，在嵌入式芯片上推理速度达25.1 FPS，综合性能优于YOLO系列多个版本及Faster R-CNN、RT-DETR等对比模型。本研究为轻量化病虫害检测算法在边缘设备中的实际应用提供了有效解决方案。

Abstract: This paper addresses the need for intelligent identification of pests and diseases in navel orange orchards by conducting research on the optimization of a deep learning-based object detection algorithm. A lightweight improved model named YOLOv11-SMALL is proposed, utilizing a constructed image dataset of navel orange pests and diseases under complex backgrounds. Based on YOLOv11n, the model incorporates the A Down downsampling module to reduce parameters and computational cost, embeds the HGNetV2 backbone network to enhance multi-scale feature extraction, and employs the ASFF (Adaptively Spatial Feature Fusion) mechanism to improve detection performance for small targets and in complex environments. Experimental results show that the improved model achieves an accuracy of 0.975, mAP@0.5 of 0.971, and mAP@0.5:0.95 of 0.912, with only 1.728K parameters and a model size of 3.6M. It achieves an inference speed of 25.1 FPS on an embedded chip, outperforming multiple versions of the YOLO series as well as comparative models such as Faster R-CNN and RT-DETR in overall performance. This study provides an effective solution for the practical application of lightweight pest and disease detection algorithms on edge devices.