Using Deep Learning for Defect Classification on a Small Weld X-ray Image Dataset

This document provides a comparative evaluation of the performance of a deep learning network for different combinations of parameters and hyper-parameters. Although there are numerous studies that report on performance in deep learning networks for ordinary data sets, their performance on small data sets is much less evaluated. The objective of this work is to demonstrate that such a challenging small data set, such as a welding X-ray image data set, can be trained and evaluated obtaining high precision and that it is possible thanks to data augmentation. In fact, this article shows that data augmentation, also a typical technique in any learning process on a large data set, plus that two image channels, such as channels B (blue) and G (green), both are replaced by the Canny edge map and a binary image provided by an adaptive Gaussian threshold, respectively, gives to the network a 3% increase in accuracy, approximately. In summary, the objective of this work is to present the methodology used and the results obtained to estimate the classification accuracy of three main classes of welding defects obtained on a small set of welding X-ray image data.