IROS 2025

This paper presents a wireless collaborative inference framework optimized for deep learning-based weed instance segmentation on resource-limited weeding robots. Traditional Mask R-CNN struggles with detecting small weeds, suffers from low recall rates, and exhibits the checkerboard effect in segmentation results. To address these challenges, we introduce three key improvements: a feature fusion strategy in the backbone network to enhance small object detection, an improved Region Proposal Network (RPN) with Soft-NMS to reduce false positives and missed detections in complex environments, and a refined mask branch incorporating fully connected upsampling to mitigate checkerboard effects. Additionally, knowledge distillation is employed to compress the model, significantly improving inference speed while maintaining segmentation accuracy. To further enhance inference efficiency, we propose a two-stage approach for determining the optimal partition point and develop a resource-aware optimization algorithm that dynamically adjusts to fluctuating network bandwidth and computational constraints. Experimental evaluations confirm that the proposed approach surpasses existing methods and remains stable across varying resource conditions. A real-world implementation of a drone-server system validates the feasibility of the framework, showcasing its potential for robust and scalable weed detection and segmentation in precision agriculture applications.