A Semantic Segmentation Method for Weld Seam Detection and Localization in Robotic Welding

Abstract

Automated robotic welding requires precise detection and localization of weld seams to generate accurate motion trajectories. Existing approaches based on edge detection and threshold segmentation often lack robustness under challenging industrial environments—such as arc interference and surface contamination. Although recent deep learning techniques demonstrate promising improvements, they frequently struggle with multi-scale feature extraction and real-time processing demands. To address these limitations, this study presents an improved lightweight semantic segmentation framework based on dynamic squeeze networks for weld seam detection in X-ray images. Building upon the dynamic squeeze strategy, this study integrates a lightweight Atrous Spatial Pyramid Pooling (ASPP) module to enhance multi-scale contextual feature extraction through parallel depthwise separable convolutions with optimized dilation rates. Additionally, a Convolutional Block Attention Module (CBAM) replaces the original spatial attention mechanism, providing dual-channel and spatial attention to better highlight weld seam regions while suppressing background interference. Illumination-adaptive data augmentation techniques are employed to improve model robustness in complex X-ray imaging scenarios. Experimental validation on X-ray weld seam datasets demonstrates that the proposed method achieves superior segmentation accuracy while maintaining lightweight characteristics and real-time inference capability. The improved framework successfully balances segmentation precision and computational efficiency, making it suitable for industrial welding applications.