Abstract
Precisely segmenting multi-category farmland weeds is of great significance for achieving targeted weeding and sustainable agriculture. However, the similar morphology between field crops and weeds, complex occlusions, variable lighting conditions, and the diversity of target scales pose severe challenges to the accuracy and efficiency of existing methods on resource-constrained platforms. This study proposes a novel instance segmentation framework, CPD-WeedNet, specifically designed for fine-grained weed identification in complex field scenarios. CPD-WeedNet innovatively presents three core components: the CSP-MUIB backbone module, which enhances the discriminative ability of initial features at a low computational cost; the PFA neck module, which efficiently integrates shallow-layer details to improve the contour capture of small and medium-sized targets; and the DFS neck module, which utilizes the Transformer to enhance global context understanding and cope with large targets and complex occlusions. On a self-constructed soybean field weed dataset, CPD-WeedNet achieved 80.6% mAP50(Mask) and 85.3% mAP50(Box), with pixel-level mIoU and mAcc reaching 86.6% and 94.6% respectively, significantly outperforming mainstream YOLO baselines. On the public Fine24 dataset, CPD-WeedNet attained 75.4% mIoU, 81.7% mAcc, and 65.9% mAP50 (Mask), demonstrating an excellent balance between performance and efficiency. The proposed CPD-WeedNet achieves an excellent balance between performance and efficiency, demonstrating its significant potential as a key vision technology for the development of low-cost, real-time intelligent weeding systems. This research is of great significance for promoting precision agriculture.