<p>Existing weed detection methods for rapeseed fields still struggle to balance fine-grained feature representation, cross-scale semantic interaction, and real-time inference. In seedling-stage rapeseed fields, weeds often show irregular shapes, large scale variations, severe occlusion, and high visual similarity to rapeseed seedlings. Current lightweight detectors improve efficiency but remain limited in adaptively representing irregular weed morphology, modeling long-range cross-scale relationships, and maintaining accuracy under deployment-oriented computational constraints. To address these gaps, this paper proposes YOLO11-DHA, a lightweight and high-precision real-time weed detection model based on YOLO11n. Specifically, a Deformable Large Kernel Attention (D-LKA) module is introduced into the backbone to enhance adaptive feature extraction for irregular and multi-scale weed targets. A Hypergraph Cross-scale Relational Fusion Neck (HCRF-Neck) is designed to model high-order relationships among multi-level features, and Adaptive Downsampling (ADown) is adopted to reduce redundant computation while preserving local texture and edge information. Experimental results show that YOLO11-DHA achieves Precision, Recall, mAP@0.5, and mAP@0.5:0.95 of 96.6%, 92.5%, 96.5%, and 87.4%, respectively, with an inference speed of 120.3 FPS. Compared with YOLO11n, the proposed model reduces parameters, computational cost, and model size by 14.7%, 7.9%, and 12.7%, respectively, demonstrating a favorable balance among accuracy, complexity, and real-time performance.</p>