YOLO-DMNet: A Dual-Module Enhanced Detector With Multi-Scale Attention Fusion

Accurate weed detection in precision agriculture remains a critical challenge due to small target sizes, dense distributions, and complex backgrounds in open-field environments. Existing object detection models often suffer from degraded performance under such conditions. This paper presents YOLO-DMNet, a lightweight and high-performance object detection model optimized for field weed scenarios. The proposed model is built upon the YOLOv8n architecture by incorporating two novel module types to enhance feature extraction and multi-scale perception. The C2f with Multi-Directional Perception Attention (C2f_MDPA) module integrates multi-scale dilated convolutions with channel and spatial attention mechanisms, thereby improving the representation of small and occluded targets. The Multi-Scale Enhanced Aggregation (MSEA) module, embedded in the detection head of the proposed model, introduces scale-adaptive convolutions and channel-guided attention to reinforce the perception of dense and morphologically diverse weed instances. Experimental evaluations on three publicly available weed datasets—CWC, CropsOrWeed9, and CottonWeedDet12—demonstrate that YOLO-DMNet achieves substantial improvements according to various indicators, compared to state-of-the-art models, while maintaining a compact architecture suitable for resource-constrained deployment. These results highlight the model’s robustness and applicability in complex agricultural settings, offering promising potential for intelligent machinery and real-time precision weed management in the field.