【摘要】 为实现对复杂背景下多尺度水稻害虫的实时精准检测，构建了一个涵盖多种水稻害虫图像的数据集，提出了一种水稻害虫实时检测模型YOLO v8-FDI。该模型基于YOLO v8n架构，采用FasterNet作为主干网络，在保持对害虫特征敏感度的同时，优化了网络结构；通过引入Dynamic Head技术，模型能够动态调整输出层检测头，以更好地适应不同类型和尺寸的害虫特征，进而提升模型精度和泛化能力；通过采用Inner-IoU损失函数，模型在损失计算过程中能够自动调节比例因子，不仅加快了模型训练收敛速度，还进一步提高了模型性能。实验结果表明，YOLO v8-FDI模型处理单幅害虫图像平均时间为12.43 ms,具备每秒处理超过80帧图像能力，满足了实际应用中的实时性要求。在测试集上，YOLO v8-FDI模型的精确率、平均精度均值(mAP@0.5:0.95)和F1分数分别为97.7%、94.0%和97.2%,相较于YOLO v3-tiny、YOLO v5n、YOLO v7-tiny、YOLO v8n、YOLO v9t和YOLO v10n模型，精确率分别提升5.2、2.7、6.7、3.4、2.2、3.2个百分点；mAP@0.5:0.95分别提升10.8、5.4、18.1、2.3、1.0、6.4个百分点；F1分数分别提升2.6、2.0、4.9、1.2、1.3、2.9个百分点。所构建的YOLO v8-FDI模型实现了复杂背景下多尺度水稻害虫实时精准检测，可为农业害虫实时监测提供技术支撑。更多还原

【Abstract】 Aiming to achieve real-time and accurate detection of multi-scale rice pests in complex backgrounds, a dataset containing images of various rice pests was constructed and a pest detection model called YOLO v8-FDI was proposed. The model was based on the YOLO v8n architecture, utilizing the more efficient FasterNet as its backbone network. This design optimized the network structure while maintaining sensitivity to pest features. Dynamic Head technology was incorporated, allowing the model to dynamically adjust the detection heads in the output layer. This improved the model’s accuracy and generalization for pests of different types and sizes. Furthermore, the Inner-IoU loss function was adopted to automatically adjust scaling factors during loss calculation process. This accelerated training convergence and further improved model performance. Experimental results showed that the YOLO v8-FDI model processed a single pest image in an average time of 12.43 m, achieving a processing speed of 80 frames per second(FPS), meeting real-time requirements for practical applications. On the test set, the model’s detection precision, mean average precision mAP@0.5:0.95 and F1 score were 97.7%, 94.0%, and 97.2%, respectively. Compared with YOLO v3-tiny, YOLO v5n, YOLO v7-tiny, YOLO v8n, YOLO v9t, and YOLO v10n, precision was improved by 5.2, 2.7, 6.7, 3.4, 2.2, and 3.2 percentage points, mAP@0.5:0.95 was increased by 10.8, 5.4, 18.1, 2.3, 1.0, and 6.4 percentage points, and F1 score was raised by 2.6, 2.0, 4.9, 1.2, 1.3, and 2.9 percentage points. The novelty lied in the improvements made to the YOLO v8n architecture by integrating FasterNet, Dynamic Head, and the Inner-IoU loss function. These enhancements significantly improved the model’s accuracy and generalization, offering strong technical support for real-time and accurate pest monitoring in complex backgrounds.更多还原