Song Jiaxin, Cheng Ke, Chen Fei, Hua Xuecheng
School of Computer, Jiangsu University of Science and Technology, Zhenjiang 212100, China.
Insects. 2025 May 21;16(5):545. doi: 10.3390/insects16050545.
Due to target diversity, life-cycle variations, and complex backgrounds, traditional pest detection methods often struggle with accuracy and efficiency. This study introduces RDW-YOLO, an improved pest detection algorithm based on YOLO11, featuring three key innovations. First, the Reparameterized Dilated Fusion Block (RDFBlock) enhances feature extraction via multi-branch dilated convolutions for fine-grained pest characteristics. Second, the DualPathDown (DPDown) module integrates hybrid pooling and convolution for better multi-scale adaptability. Third, an enhanced Wise-Wasserstein IoU (WWIoU) loss function optimizes the matching mechanism and improves bounding-box regression. Experiments on the enhanced IP102 dataset show that RDW-YOLO achieves an mAP@0.5 of 71.3% and an mAP@0.5:0.95 of 50.0%, surpassing YOLO11 by 3.1% and 2.0%, respectively. The model also adopts a lightweight design and has a computational complexity of 5.6 G, ensuring efficient deployment without sacrificing accuracy. These results highlight RDW-YOLO's potential for precise and efficient pest detection in sustainable agriculture.
由于目标的多样性、生命周期的变化以及复杂的背景，传统的害虫检测方法在准确性和效率方面常常面临困难。本研究介绍了RDW-YOLO，一种基于YOLO11改进的害虫检测算法，具有三项关键创新。首先，重参数化扩张融合模块（RDFBlock）通过多分支扩张卷积增强特征提取，以获取细粒度的害虫特征。其次，双路径下采样（DPDown）模块集成了混合池化和卷积，以实现更好的多尺度适应性。第三，增强的Wise-Wasserstein交并比（WWIoU）损失函数优化了匹配机制，改进了边界框回归。在增强的IP102数据集上的实验表明，RDW-YOLO的mAP@0.5达到71.3%，mAP@0.5:0.95达到50.0%，分别比YOLO11高出3.1%和2.0%。该模型还采用了轻量级设计，计算复杂度为5.6 G，确保在不牺牲准确性的前提下实现高效部署。这些结果凸显了RDW-YOLO在可持续农业中进行精确高效害虫检测的潜力。
