YOLO-CWD: A novel model for crop and weed detection based on improved YOLOv8

Accurate and efficient crop and weed detection is pivotal for advancing precision agriculture. In this context, the development of lightweight and high-performance models is crucial for real-time applications. This study introduces YOLO-CWD, an improved version of the You Only Look Once version 8 n (YOLOv8n) model, designed to achieve high detection accuracy while maintaining a compact scale in real-time crop and weed detection tasks. A novel hybrid attention mechanism was proposed, enhancing the model's capacity to differentiate between crops and weeds, and outperforming other commonly used attention mechanisms. To address the limitations of the Complete Intersection over Union (CIoU) loss function in accurately locating predicted bounding boxes, a novel loss function, Point Intersection over Union (PIoU), was proposed, accelerating convergence during training. The resulting YOLO-CWD model features 3.49M parameters and 9.6 GFLOPS, balancing compactness and performance. Ablation experiments demonstrate significant improvements in detecting maize and weeds, with mAP@50 reaching 0.751 and mAP@50:95 reaching 0.506, increasing by 0.008 and 0.012, respectively, compared to the baseline. The model outperforms other state-of-the-art models, with its robustness validated across multiple datasets. Further evaluation under varying lighting and soil moisture conditions highlights the model's strong generalization capabilities. These findings confirm YOLO-CWD's superiority for crop and weed detection tasks, addressing key challenges in precision agriculture and paving the way for sustainable practices.