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

Automated robotic welding in construction is frequently hindered by harsh illumination and specular reflections on metallic surfaces, which degrade the performance of semantic segmentation models. The authors investigate whether these failures can be mitigated through a lightweight post-training optimization strategy rather than by increasing the architectural complexity of the segmentation network.

The study proposes a post-training framework that refines existing segmentation models (such as BiSeNetV2, U-Net, and DeepLabV3+) using a hybrid Cross-Entropy–Lovász loss objective. Starting from an Online Hard Example Mining (OHEM)-pretrained checkpoint, the framework performs controlled fine-tuning to improve pixel-level classification and region-level seam continuity. The authors evaluate this approach across multiple architectures and quantify performance using a new metric, the Recovery Rate, which measures the restoration of catastrophic zero-IoU failure cases caused by reflections.

The proposed framework substantially improves segmentation performance for lightweight architectures. For BiSeNetV2, the method increased Joint IoU from 59.40% to 81.76% and successfully recovered 96.33% of severe zero-IoU failure cases. Crucially, these improvements were achieved without increasing parameter counts, FLOPs, or inference latency, making the approach highly suitable for real-time robotic deployment. In practical robotic welding experiments, the optimized BiSeNetV2 model reduced the average joint-center offset by approximately 9.8 times compared to the baseline, enabling reliable trajectory generation where other models failed.

This research demonstrates that reflection robustness in industrial vision is often a problem of learning stability rather than model capacity. By focusing on optimization strategies, researchers can achieve high-performance, real-time perception on resource-constrained robotic platforms without the computational overhead of larger, more complex neural network architectures.