基于改进YOLO v8n的柑橘苗害虫轻量化检测方法

Lightweight Detection Method for Citrus Seedling Pests Based on Improved YOLO v8n

摘要:

针对柑橘苗温室培育过程中常见的小目标害虫（如柑橘木虱、斜纹夜蛾、桔小实蝇）检测存在的特征提取困难、复杂环境下检测精度低,以及边缘化设备需高效实时检测等问题,本研究提出一种轻量化柑橘害虫识别模型YOLO v8n-CPestCSR。在主干网络通过引入CPNFocalNeXt模块替换C2f结构,在保证检测精度的同时提升了计算效率,显著降低了模型内存占用量与参数量,为边缘设备部署提供轻量化基础;在颈部结构中采用SimAM注意力机制与C2f进行融合,通过无参数化设计进一步降低模型内存占用量与参数量;检测网络特征融合阶段引入E-ReparamLKBlock模块,通过在检测头额外堆叠3个该模块,强化了对柑橘害虫的关键特征捕捉与区分能力,有效提高模型对小目标害虫检测性能。试验结果表明,改进YOLO v8n-CPestCSR模型精确率、召回率和平均精度均值分别达96.10%、93.33%、96.08%,相较于YOLO v8n分别提升3.75、0.94、2.8个百分点,模型内存占用量下降8.1%,浮点运算量下降9.9%;与Faster R-CNN、YOLO v5s、SSD模型相比,本模型平均精度均值分别提高19.71、1.71、11.83个百分点,并且模型参数量和复杂度较低。结果表明,YOLO v8n-CPestCSR模型在提高柑橘害虫检测精度方面取得显著成效,且有利于部署到移动端。将模型部署至Jetson Xavier NX移动端平台,并进行推理加速。检测结果表明,YOLO v8n-CPestCSR模型精确率、召回率以及检测速度均优于YOLO v8n。最后,通过吸虫装置结合检测模型进行实地试验,验证了模型在实际应用的可行性和有效性,可有效支撑温室柑橘苗精准虫害预警与绿色防控,符合现代设施农业对高效、低耗害虫防治的实际需求。

Abstract:

Aiming to address the challenges of difficult feature extraction, low detection accuracy in complex environments, and the need for efficient real-time detection on edge devices for small-target pests (e.g., Diaphorina citri, Spodoptera litura, and Bactrocera dorsalis) commonly encountered during greenhouse citrus seedling cultivation, a lightweight citrus pest recognition method, YOLO v8n-CPestCSR, was proposed. In the backbone network, the C2f structure was replaced with a CPNFocalNeXt module, which improved computational efficiency while maintaining detection accuracy, significantly reducing model weight and parameter count, thereby providing a lightweight foundation for edge device deployment. In the neck structure, the SimAM attention mechanism was integrated with C2f, further reducing model weight and parameter count through a parameter-free design. During the feature fusion stage of the detection network, an E-ReparamLKBlock module was introduced;by additionally stacking three such blocks in the detection head, the model's ability to capture and discriminate key features of citrus pests was enhanced, effectively improving small-target pest detection performance. Experimental results showed that the improved YOLO v8n-CPestCSR model achieved a precision of 96.10%, a recall of 93.33%, and a mean average precision (mAP) of 96.08%, representing increases of 3.75, 0.94, and 2.8 percentage points over that of the baseline YOLO v8n, respectively. Meanwhile, model weight was reduced by 8.1% and floating point operations (FLOPs) was reduced by 9.9%. Compared with models such as Faster R-CNN, YOLO v5s, and SSD, the proposed model improved mAP by 19.71, 1.71, and 11.83 percentage points, respectively, while maintaining a lower parameter count and model complexity. These results demonstrated that the improved YOLO v8n-CPestCSR model achieved significant success in enhancing citrus pest detection accuracy and was favorable for mobile deployment. The model was deployed on a Jetson Xavier NX mobile platform with inference acceleration. Detection results confirmed that the improved YOLO v8n-CPestCSR outperformed YOLO v8n in precision, recall, and inference speed. Finally, field experiments combining a pest suction device with the detection model verified the feasibility and effectiveness of the algorithm in real-world applications, demonstrating its capacity to effectively support precise pest early warning and green pest management for greenhouse citrus seedlings, in line with the practical demands of modern protected agriculture for efficient, low-consumption pest control.