Unsupervised machine learning used for anomaly detection in x-ray images

Abstract

To have computers understand images has long been a hard task due to the largely variable composition of data, but advancements in the field of computer vision has given capable technologies of managing it. The primary technology is the usage of machine learning, where convolutional neural networks has shown to be really potent. This project has tackled the task of detecting minuscule cavities in massive x-ray images of welded metal. The challenge lies both within processing and providing data so that it favors the machine learning model, but also developing models that are complex enough so that it yields a general understanding of the underlying data structure. This project utilizes the benefits of unsupervised machine learning by using a convolutional autoencoder that reconstructs inputs through encoding and decoding. The reconstruction error of the autoencoder is here the primary measurement for anomaly detection. By having the autoencoder trained on millions of small, non-faulty fragments of the x-ray images, the reconstruction error should be larger for fragments with faulty anomalies than those without. Various statistical methodologies have been evaluated in order to find the most suitable alternative to use on the reconstruction error. While there still is work to be done, worthwhile achievements has been made.