DGEA-HRNet: A Fast and Accurate Branch Detection and Segmentation Network for Tree Pruning Robot

Landscaping trees play a crucial role in the city's ecosystem, economy, and appearance. In recent years, pruning robots have been developed for branch pruning, where fast and accurate detection of tree branches is essential. In this paper, a lightweight and high-precision semantic segmentation model, DGEA-HRNet, is proposed for real-time segmentation of tree branches. By designing the lightweight core component DGE-Block, computational complexity is significantly reduced, and segmentation efficiency is improved. Meanwhile, the Coordinate Attention (CA) module is introduced to enhance the spatial and location information extraction ability, and the multi-branch features are fused by CBAM to improve feature fusion and segmentation accuracy. Experimental results show that the model achieves the highest mIoU and branch_IoU on the self-built dataset, at 95.27% and 94.17%, respectively. The model also performs well in terms of lightweight efficiency, with FLOPs, Params, and Model Size of 49.02G, 1.73M, and 6.84M, respectively. In summary, DGEA-HRNet is able to quickly and accurately perform tree branches segmentation and provides strong technical support for tree pruning robots.