Vision-based welding quality detection of steel bridge components in complex construction environments

Currently, welding quality detection remains dependent on manual operation, while the increase in the span and intricacy of steel bridges has rendered the conventional method of detection insufficient to fulfill the engineering requirements. This paper presents a systematic study of welding quality detection of steel bridges based on fusion of point clouds and images in complex construction environments. (1) A welding detection system is developed that could filter out stray light and capture weld images. (2) This paper enhances the centerline extraction method in 3D reconstruction, which could effectively filter out noise interference and precisely reconstruct weld contours. The contour dimensions of both filler and cover welds are identified through feature point extraction, with an estimated detection error under 0.6%. (3) This paper optimizes the feature extraction of the Faster R-CNN network based on the appearance feature and detection need of welding defects, resulting in an improvement of 28.3 in mAP. Experimental results demonstrate that the proposed welding quality detection is both efficient and accurate, and is capable of meeting the requirements of actual steel bridge construction.