A boundary-statistical context fusion network for deposited bead region segmentation toward teaching-free repair in arc-based directed energy deposition

Segmentation of deposited bead region (DBR) boundaries remains a major challenge in teaching-free repair of arc-based directed energy deposition (DED-Arc) parts due to blurred geometric transitions and low feature contrast near the boundary regions. These characteristics limit the conventional segmentation approaches and hinder intelligent repair control. To address these issues, this paper presents a boundary-statistical context fusion network (BSCFNet) tailored for DBR boundary segmentation in teaching-free DED-Arc repair. A Boundary-Enhanced Hybrid Sampling method is introduced to retain the global structure and boundary-local features during sampling. A Multi-branch Statistical Channel Attention mechanism is designed to capture detailed channel-wise variations and integrate them with global contextual features from the Global Context Extractor, thereby enhancing feature discrimination. Experimental results on a real scanned dataset (DBRSet) show that the mean Intersection over Union (mIoU) and Overall Accuracy of BSCFNet improve by 14.17% and 6.04% compared with the baseline, respectively. The IoU of DBR and substrate are improved to 89.33% and 91.35%, outperforming the baseline by 18.96% and 9.84%, respectively. Additional validation on diverse geometries and non-uniform deposition morphologies indicates the robustness and industrial applicability of BSCFNet. A repair-oriented experiment on a composite workpiece demonstrates that BSCFNet can be effectively used for teaching-free DED-Arc repair.