<p>Industrial X-ray inspection, part of the non-destructive quality control in manufacturing industry, is often resource intensive; and the interpretation of the X-ray images of critical parts is still done by humans within many quality critical segments, e.g. the aerospace industry. A higher level of automation of the X-ray image interpretation is desirable, and costs can be lowered as well as more quantitative information, e.g. for manufacturing optimization, can be extracted. We claim, that for a safe automatic interpretation of industrial X-ray images, the algorithms should react reasonable to out-of-distribution data by assigning a low confidence to these samples; and ideally there should be no unknown unknowns (i.e. surprising or undefined behavior of the algorithm). In this work we have explored the hypothesis that such a confidence estimation can be achieved by comparing input X-ray images with a high precision model of the accepted X-ray image intensity variation. Deviations in that residual image would indicate a low confidence, an anomaly. For the X-ray image model we have designed an autoencoder, a deep convolutional neural network, and for the residual analysis a kernel/patch-based one. The public dataset GDXray with digitized X-ray images of metal fusion-welds was utilized for unsupervised training and testing. We achieved a true positive rate at some 80-90 % at a false positive rate at around 4 %, as well as correctly detected an artificial (as in designed) out-of-distribution data example as an anomaly.</p>

Autoencoder-Based Anomaly Detection in Industrial X-Ray Images

Paper Proceedings: https://asmedigitalcollection.asme.org/qnde/QNDE2021/volume/85529

ASME 2021 48th Annual Review of Progress in Quantitative Nondestructive Evaluation (QNDE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.