RIAWELC: A Novel Dataset of Radiographic Images for Automatic Weld Defects Classification
Benito Totino, Fanny Spagnolo, Stefania Perri
Abstract
In the last few years, extracting, analyzing and classifying welding defects in radiographic images received a great deal of attention in several industry manufacturing. Nowadays, computer vision affords considerable accuracy in many practical applications, but making automatic processes approachable also in this field is still a challenge. As an example, Convolutional Neural Networks (CNNs) are widely recognized as efficient and accurate classification structures, but, due to the limited availability of specific datasets, training a CNN to classify welding defects is not trivial. This paper presents a new dataset collecting 24,407 radiographic images related to several classes of welding defects: lack of penetration, cracks, porosity and no defect. The proposed dataset of welding defects in radiographic images is released freely to the research community. As an example of application, the dataset has been used to train a customized version of the SqueezeNet CNN obtaining a test accuracy higher than 93%.