A lightweight deep learning method for real-time weld feature extraction under strong noise

This paper proposes a lightweight deep learning (DL) framework for real-time accurate weld feature extraction from noisy images with light, smoke, or splash. Leveraging a two-dimensional human pose estimation paradigm, the framework follows a top-down architecture for accurate weld feature point localization. This study develops a semi-automatic annotation technique to dramatically reduce the annotation cost. Then, we design a lightweight yet faster You Only Look Once version 8 (YOLOv8) detector to rapidly detect the weld feature region in the presence of strong noise. To avoid reliance on high-resolution feature maps and achieve sub-pixel-level localization accuracy, a heatmap-free approach decomposes the feature point detection task into subtasks of horizontal and vertical coordinate classification. Comparison with mainstream DL-based weld recognition methods validates the superiority of the proposed method regarding real-time feature extraction accuracy and robustness.