CDGL–YOLOv11Lite: A Lightweight Field Maize–Seedling Weed Detector Based on Cross-Domain Shared Attention and a Conditional Gated Linear Unit

In maize-seedling fields, weeds and crop seedlings are small, visually similar, and often occluded amid drastic illumination and soil-background variation. We present CDGL–YOLOv11Lite, a lightweight YOLOv11-based detector that couples Cross-Domain Shared Attention (CDSA)—dual spatial/channel axial branches with shared cross-domain projections and post-alignment reweighting—with a Conditional Gated Linear Unit (CGLU) using depthwise 3×3 context-aware gating to suppress background misactivations and strengthen small-object cues. To satisfy edge constraints, we perform LAMP-based joint channel/layer pruning with light fine-tuning, then export to ONNX and optimize in TensorRT via operator fusion and mixed-precision inference. On a custom maize–weed dataset, CDGL–YOLOv11Lite achieves 90.25% AP, 88.37% recall, and [email protected] of 90.45%; the deployable model is 1.9MB (0.8M parameters, 3.1GFLOPs) and runs at 67FPS on NVIDIA Jetson Xavier NX, offering a practical balance of accuracy, efficiency, and edge-side deployability.

Ҳали таржима қилинмаган