Automatic defect recognition in x-ray imaging by machine learning

Abstract

Digital radiography is widely used in medicine (e.g.: to evaluate and diagnose bones) and is increasingly used in industrial nondestructive testing (NDT) where it replaces photographic films. The latter is the focus of this work. Typically, a very efficient expert needs at least five minutes to process a single image (depending of its size), hence performing the quality assessment of a single object is very time consuming.

The purpose of this project is to decrease the work load of the human operator: specifically, the goal is to identify suspicious areas and not to classify defects with their type nor automate the whole NDT process, the human agent will always deliver the final decision according to its own expertise on those suspicious zones. Because defect recognition is usually performed on sensitive objects (e.g.: parts of aircrafts), the mandatory constraint imposed is that the detection algorithm must be tuned to never miss a real defect.

The protocol is explained. Tree-based ensemble methods, especially extremely randomized trees, have been used to learn defect representation from small patches sampled in the image. Defect patches are sampled using the annotation polygons provided in the datasets. No-defect patches are sampled randomly, with the possibility to use a mask to restrict the possible areas of extraction. To test a never seen image, a sliding window is used to extract consecutive patches that are then given a defectiveness probability by the model. Finally, these probabilities are aggregated to form a score heat map for each pixel.

To assess the quality of the protocol, specifically the amount of false positive achieved when tuning it to not miss any defect, four datasets are used. Two are composed of synthetic defects (of the same type) that we have generated and two are real ones, provided by \textit{X-RIS}.

The best achieved result is for the synthetic dataset of high intensity defects where only 6 additional non-defective zones must be analyzed by the human operator in addition to the 17 generated one. For the two real datasets, the amount of no-defective zones detected is high. For the best one, there are 2108 misclassified zones for 130 correctly classified zones representing 124 defects.

Finally, the entire protocol seems to be relevant but the patch classifier must be improved. On one hand, there are multiple of other image descriptors that can be evaluated, either engineered ones or learned ones (e.g.: with neural networks). On the other hand, there exist other classifiers than random forest and extremely randomized trees: a softmax layer in neural networks or support vector machines.