Welding Defect Detection in Steel Reinforcing Bars using Deep Learning

Weld connection is an important process in on-site reinforcing bar engineering, which should finally be visually inspected by construction workers to ensure that the quality of the joint meets the code requirements. Manual inspection is mostly time-consuming and laborious. The development of deep learning has led to great advances in the field of industrial defect detection. However, there are some challenges to considering both accuracy and efficiency in the recognition of rebar welding defects. Hence, in this study, a defect detector has been proposed, based on YOLOv3. In the proposed network, we integrate the dilated convolution with different dilation rates to increase the receptive field of the backbone. Moreover, CIoU loss is utilized, which accelerates the bounding box regression and improves the accuracy of defect detection. Furthermore, Focal loss is indeed applied to solve the class imbalance problem. The experiments are run on a real-world rebar welding dataset that mAP improves by 6.83%, compared with YOLOv3. Meanwhile, the proposed detector also meets the real-time requirements of the rebar welding defect detection task.