【摘要】 针对真实场景中虫害体积小、分布密集导致的错检和漏检问题，提出了一种基于改进YOLOv9 (You only look once version 9)的轻量化虫害检测模型。通过在YOLOv9的骨干网络中引入轻量化特征提取模块，显著降低模型参数量的同时，保留了关键特征信息，提升了模型特征提取的速度和精度。在YOLOv9的颈部网络中引入了改进的位置卷积模块，该模块能够有效增强模型对目标位置信息的感知能力，从而提高了检测精度。采用基于最小点距离的边界相似性比较度量损失函数(Minimum point distance based border similarity comparison metric loss function, MPDIoU)作为模型的边界框回归损失函数，通过优化边界框的回归性能，进一步提升了模型的定位精度，并加速了模型的收敛速度。在大规模多目标的标准化农作物害虫数据集(Large-scale multi-target standardized dataset of agricultural pests, Pest24)验证了本模型的改进效果，相比于YOLOv9,此方法的模型参数量降低了20.0%,并且平均精确率(Mean average precision,mAP)提升了0.4%,达到了75.0%。更多还原

【Abstract】 Aiming at the problems of misdetection and omission caused by the small size and dense distribution of pests in real scenes, a lightweight pest detection model is proposed based on improved YOLOv9(You only look once version 9). The lightweight feature extraction module into the backbone network of YOLOv9 is introduced. This approach significantly reduces the model’s parameter count while preserving key feature information, thereby enhancing both the speed and accuracy of feature extraction. An improved positional convolution module is introduced into the neck network of YOLOv9, which can effectively enhance the model’s ability to perceive the target position information, thus improving the detection accuracy. The minimum point distance is adopted based border similarity comparison metric loss function(MPDIoU) as the bounding box regression loss function of the model, which further improves the localization accuracy of the model and accelerates the convergence speed of the model by optimizing the regression performance of the bounding box. The proposed model’s improvements are validated on a large-scale multi-target standardized dataset of agricultural pests(Pest24). Compared to YOLOv9, this approach reduced model parameters by 20.0% while increasing the mean average precision(mAP) by 0.4%, achieving 75.0%.更多还原