Enhanced seam segmentation for automated welding robot in construction through transfer learning: addressing limitations of bilateral segmentation network.

Reliable weld seam segmentation is essential for autonomous robotic welding in construction, where severe illumination changes, specular reflections, and thin weld geometries frequently degrade segmentation performance. This paper proposes a lightweight post-training optimization framework that systematically improves existing real-time semantic segmentation networks without modifying their architectures. Starting from an Online Hard Example Mining (OHEM)-pretrained checkpoint, the proposed framework combines controlled fine-tuning with a hybrid Cross-Entropy-Lovász objective to enhance pixel-level classification, region-level seam continuity, and recovery from reflection-induced segmentation failures while preserving real-time inference efficiency. Extensive experiments demonstrate that the proposed framework substantially improves segmentation performance. Using BiSeNetV2, Joint IoU increases from 59.40% to 81.76% (+22.36 percentage points), while mIoU reaches 90.73% without increasing parameter count, FLOPs, inference latency, or memory consumption. Furthermore, the proposed framework successfully recovers 96.33% of severe zero-IoU failure cases caused by strong reflections, substantially improving the operational reliability of downstream robotic perception. To evaluate the generality of the proposed optimization framework, additional experiments were conducted using U-Net, DeepLabV3+, SegFormer-B0, and PIDNet-S under multiple fine-tuning configurations. The results reveal that the effectiveness of post-training optimization is strongly architecture-dependent: lightweight real-time segmentation networks, particularly BiSeNetV2 and PIDNet-S, consistently benefit from the proposed framework, whereas larger semantic-oriented architectures exhibit comparatively smaller and less stable improvements. In robotic welding experiments, the proposed BiSeNetV2 model was the only evaluated approach capable of consistently generating a valid weld seam trajectory, achieving an average joint-center offset of [Formula: see text] mm during robotic path planning, corresponding to an approximately 9.8 × reduction compared with the previous BiSeNetV2-based robotic welding system. Overall, the results demonstrate that carefully designed post-training optimization provides a practical and computationally efficient alternative to architectural redesign, substantially improving failure recovery, seam continuity, and robotic trajectory generation for autonomous welding in reflective construction environments.