关键词: 水稻害虫检测, 小目标检测, 多尺度特征融合, YOLOv11, 智慧农业

Abstract: In the context of rice pest detection in complex environments, the traditional YOLOv11 model faces challenges of insufficient small target recognition capability and inadequate multi-scale feature fusion. To address these issues, an improved detection model based on YOLOv11, named QMDF-YOLO11, is proposed. Firstly, a novel feature fusion network (QS-RepGFPN) is proposed to replace the neck structure of the original model, achieving efficient fusion of multi-level features and significantly enhancing the perception and detection capabilities for small targets. Secondly, the backbone network incorporates modulated deformable convolutions (MDConv), enhancing the network’s robustness and expressive power for deformed targets, while adopting the DySample dynamic sampling module to replace traditional upsampling methods, improving the interpolation accuracy in the multi-scale feature fusion process. Additionally, by combining the QS-RepGFPN structure, a novel four detection head, QASFFHead (quadruple adaptive spatial fusion head), is innovatively designed, further optimizing the utilization and fusion of multi-scale features and enhancing the precision and efficiency of the model in detecting targets of different scales. Experimental results show that the improved model achieves an mAP@0.5 of 94.57% on the RicePests dataset, an increase of 5.26 percentage points, with precision (P) and mAP@0.5：0.95 improved by 5.89 percentage points and 7.46 percentage points, respectively, particularly outstanding in complex backgrounds and scenarios with dense small targets. Meanwhile, the increase in model parameters and computational costs is kept within a reasonable range, ensuring efficient inference speed. The experimental results indicate that the improved model is suitable for resource-constrained agricultural monitoring platforms, real-time target detection potential, providing references for the further development of smart agriculture.

Key words: rice pest detection, small object detection, multi-scale feature fusion, YOLOv11, smart agriculture