CWDet-YOLO: a lightweight network for corn field weed detection based on YOLOv11

The handling of weeds plays a crucial role in corn yield, but traditional herbicide spraying often leads to environmental pollution. Precise weed control in corn cultivation can address this issue. Despite advancements in deep learning-based object detection, challenges persist in agricultural field settings due to varying weed morphologies, complex backgrounds, leaf occlusions, and real-time requirements. The proposed CWDet-YOLO network improves detection performance significantly, addressing issues like misjudgments and missed detections caused by occlusions. The introduced simAM attention mechanism is a lightweight and efficient attention mechanism with no trainable parameters, low computational cost, and enhanced weed feature extraction capabilities. The improved AC_C3k2 module, designed by integrating Additiveblock and Convolutional GLU modules, reduces model complexity, facilitating better global context acquisition. The use of dysample dynamic upsampling as an upsampling operator enables adaptive upsampling by learning pixel-level sampling positions, thereby achieving more precise detail capture, especially beneficial for reconstructing complex textures or edge regions. Across its n, s, m, l, x versions, CWDet-YOLO networks demonstrate improved performance metrics compared to the five versions of YOLOv11, with reduced model complexity.