Enhanced Seam Segmentation for Automated Welding Robot in Construction Through Transfer Learning: Addressing Limitations of Bilateral Segmentation Network

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Construction Technology & Building · Depth: Expert, extended

Summary

A new reflection-robust seam segmentation framework is proposed for automated robotic welding in construction, addressing challenges like harsh illumination and specular reflections. The framework enhances a BiSeNetV2 backbone through transfer learning and a hybrid Cross-Entropy–Lovász loss, focusing on learning-stability-oriented optimization rather than increased architectural complexity. It achieves 81.76% Joint IoU and 90.73% mIoU, marking a +22.36 percentage point improvement in Joint IoU over an OHEM-based baseline. Crucially, the method recovers 96.33% of severe zero-IoU failure cases under reflective conditions while maintaining identical FLOPs, parameter count, and inference speed. Comparative experiments across BiSeNetV2, DeepLabV3+, UNet, and SegFormer demonstrate its particular effectiveness for lightweight real-time segmentation architectures, improving seam continuity and robustness in challenging welding environments using the WJ3600 dataset.

Key takeaway

For Machine Learning Engineers developing perception systems for automated welding robots, you should prioritize learning-stability-oriented optimization over architectural complexity. Implementing transfer learning with a hybrid Cross-Entropy–Lovász loss on lightweight models like BiSeNetV2 can drastically improve seam continuity and recover critical zero-IoU failures under reflective conditions, ensuring more reliable robotic trajectory generation without increasing computational overhead. Consider fine-tuning with a Lovász-dominant loss for optimal performance.

Key insights

Optimizing learning stability with hybrid loss and transfer learning significantly improves weld seam segmentation robustness without increasing model complexity.

Principles

• Reflection robustness is a learning-stability problem, not just model capacity.
• IoU-oriented optimization is critical for thin, continuous structures.
• Recovery of severe zero-IoU failures is paramount for robotic reliability.

Method

The framework uses a BiSeNetV2 backbone, pretrained with OHEM, then fine-tuned for one epoch with a hybrid Cross-Entropy–Lovász loss (optimal λ=0.04) and reflection-aware augmentations.

In practice

• Apply transfer learning with hybrid CE–Lovász loss.
• Prioritize lightweight architectures for real-time tasks.
• Use reflection-aware data augmentation.

Topics

• Automated Welding
• Semantic Segmentation
• Transfer Learning
• BiSeNetV2
• Hybrid Loss Functions
• Robotic Perception

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• BIFROST: Bridging Invariant Feature Representation for Observation-space Sim2Real Transfer — BIFROST enables zero-shot sim2real transfer by learning domain-invariant latent states via a cross-domain bisimulation objective.
• Learning Structurally Consistent Representations for Multi-View Radar Semantic Segmentation — Higher-order relational modeling and cross-view consistency improve radar semantic segmentation in adverse conditions.
• Hybrid Deep Learning for Traceability and Classification of Industrial Slate Tiles — A hybrid deep learning model combining feature matching and classification significantly improves industrial slate tile traceability and quality control.
• STEP-Parts: Geometric Partitioning of Boundary Representations for Large-Scale CAD Processing — STEP-Parts extracts stable geometric instance partitions from B-Reps, preserving analytic structure for CAD learning.
• S23DR 2026 Winning Solution — A flow-matching DiT conditioned on scene tokens effectively reconstructs 3D wireframes from sparse multi-modal inputs.
• Geometry-Aware Infrastructure-Anchored Denoiser for UWB Sensing and Work-Zone Reconstruction — Geometry-aware range denoising significantly improves UWB-based work-zone reconstruction accuracy by coupling temporal modeling with spatial consistency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.