Edge-deployable cotton–weed pose estimation for growth-point localization in mechanical weeding

Precision mechanical weeding requires not only reliable crop–weed recognition but also an explicit image-plane target for each weed, because bounding boxes and segmentation masks do not directly indicate where a mechanical end effector should act. This study aimed to develop an edge-deployable method that jointly recognizes cotton and five weed species and directly predicts one visually defined growth-point keypoint for each plant instance. We propose WGPose, a lightweight pose-estimation network that represents each plant using its class, bounding box, and a single growth-point keypoint; only growth-point keypoints associated with weed instances are retained as candidate references for subsequent mechanical positioning. To reduce inference cost while preserving local plant structures, the explicit multiscale neck of YOLO11n-Pose was removed, backbone features were connected to a shared lightweight prediction head, and detail-preserving downsampling, high-level feature regulation, and multiscale channel-wise knowledge distillation were incorporated. The model was evaluated using 2,177 cotton-field images containing 2,830 annotated plant instances. WGPose + CWD achieved a Pose mAP50–95 of 94.30 ± 0.57% across three independent training runs using random seeds 42, 43, and 44, with 1.17 million parameters and 3.91 GFLOPs. Compared with YOLO11n-Pose (92.42 ± 0.65%), the mean Pose mAP50–95 increased by 1.88 percentage points, while the parameter count and GFLOPs decreased by 56.02% and 40.40%, respectively. Relative to non-distilled WGPose (92.96 ± 1.02%), CWD improved the mean Pose mAP50–95 by 1.34 percentage points without increasing inference complexity. In a separate single-checkpoint coordinate-level evaluation, the mean normalized distance error decreased from 0.0562 to 0.0468, and [email protected] increased from 59.46% to 70.30%. In a three-run offline RGB-image benchmark on the Jetson AGX Orin, the TensorRT prediction pipeline achieved a mean latency of 20.59 ± 0.08 ms per image, corresponding to 48.57 ± 0.18 FPS. These results demonstrate that directly predicting instance-level growth-point keypoints can provide compact and explicit visual references for weed-specific positioning while maintaining localization accuracy and edge-deployment efficiency.