WDB-YOLO-World: Deblurring-Enhanced Open-Vocabulary Weed Detection for Low-Altitude Remote Sensing
Agricultural low-altitude remote sensing presents three coupled challenges for weed detection: open-set categories, image degradation, and small targets. This study proposes WDB-YOLO-World, an open-vocabulary detector that combines image deblurring, high-resolution feature enhancement, and large language model (LLM)-assisted prompt refinement. A Restormer-based deblurring frontend module restores motion- and defocus-blurred unmanned aerial vehicle (UAV) images, while a cross-layer high-resolution feature injection mechanism preserves spatial detail for small weeds. An LLM-based prompt refinement module constructs an LLM-refined prompt vocabulary to improve vision-language alignment. Using PhenoBench, we generate paired sharp and synthetically blurred images and evaluate the model under mild, moderate, and severe blur. The deblurring frontend module increases the average PSNR from 26.12 to 30.11 dB and SSIM from 0.7635 to 0.8739, while reducing LPIPS from 0.2819 to 0.1284. Under the synthetic-blur test protocol, WDB-YOLO-World achieves a precision of 85.26%, a recall of 81.90%, an F1-score of 83.55%, and an mAP@0.5 of 85.03%, outperforming YOLOv8, OV-R-CNN, YOLO-World, and grounding DINO. The model contains 61.33 M parameters and requires 232.54 GFLOPs, compared with 162.12 M parameters and 461.46 GFLOPs for grounding DINO. These results demonstrate that, in the synthetic blur test scenarios constructed in this study, WDB-YOLO-World improves open-vocabulary weed detection in UAV images degraded under simulated blur conditions, thereby providing an experimental basis for further validation in real agricultural environments.