An end-to-end intelligent welding defect detection system for low-quality X-ray images with adaptive progressive learning

Abstract

Intelligent weld defect detection based on X-ray images is a paramount research topic in the field of industrial automation. Many excellent object detectors have been developed into industrial vision frameworks. Despite significant efforts, these methods still fail to effectively detect weld defects in low-quality X-ray images (low-brightness, foggy). We have carefully considered the evaluation process of human experts, and experienced engineers usually carefully adjust the details of the image, and then focus all their attention on defect evaluation after briefly browsing the image. Inspired by the above process, we develop a novel framework to detect different types of defects from low-quality X-ray images. Firstly, an adaptive image preprocessing method is designed to simulate the fine adjustment of image details (brightness, contrast) during manual evaluation. The parameters of image processing can be obtained through joint training of a differentiable preprocessing module and a back-end detector, thereby flexibly adjusting the details of the image. Secondly, a progressive feature extractor is proposed, which includes three inherently connected components: multi-view feature extraction, large-kernel perceptron, and feature recombination to simulate the rough to centralized process of manual evaluation, thereby improving the model’s ability to capture defects and overall performance. In practical application, we used the weld defect dataset based on X-ray in North China. We achieved a 8.5% AP improvement (23.4 M model training parameters), with inference time of 15.9 ms per image and resolution (640 × 640).