A Welding Defect Identification Approach in X-ray Images Based on Deep Convolutional Neural Networks

摘要

Welding defect detection and identification is a crucial task in the industry that has a significant economic impact and can pose safety risks if left unnoticed. Currently, an inspector visually evaluates the condition of welds to guarantee reliability in industrial processes. This way is time-consuming and subjective. This paper proposes a deep learning-based approach to identify automatically multiple welding defect types and locations in X-ray images by adopting a pre-trained RetinaNet-based convolutional neural network. To realize this, a dataset including 6714 images labeled for three types of welding defect— blowhole, under fill or incomplete penetration, and tungsten inclusion—is developed. Then, the RetinaNet-based model is designed and trained using this database. The proposed approach can not only directly detect welding defect on the original input images without any pro-processing, but also identify the types of defect. Moreover, the experimental tests show that the proposed approach works well on the images even in low resolution and mean average precision (mAP) ratings are 0.76, 0.79, and 0.92 respectively for three defect types.