Deep learning technology for weld defects classification based on transfer learning and activation features

Abstract

Weld defects detection using X-ray images is an effective method of nondestructive testing. Conventionally, this work is based on qualified human experts, although it requires their personal intervention for the extraction and classification of heterogeneity. Many approaches have been done using machine learning (ML) and image processing tools to solve those tasks. Although the detection and classification have been enhanced with regard to the problems of low contrast and poor quality, their result is still unsatisfying. Unlike the previous research based on ML, this paper proposes a novel classification method based on deep learning network. In this work, an original approach based on the use of the pretrained network AlexNet architecture aims at the classification of the shortcomings of welds and the increase of the correct recognition in our dataset. Transfer learning is used as methodology with the pretrained AlexNet model. For deep learning applications, a large amount of X-ray images is required, but there are few datasets of pipeline welding defects. For this, we have enhanced our dataset focusing on two types of defects and augmented using data augmentation (random image transformations over data such as translation and reflection). Finally, a fine-tuning technique is applied to classify the welding images and is compared to the deep convolutional activation features (DCFA) and several pretrained DCNN models, namely, VGG-16, VGG-19, ResNet50, ResNet101, and GoogLeNet. The main objective of this work is to explore the capacity of AlexNet and different pretrained architecture with transfer learning for the classification of X-ray images. The accuracy achieved with our model is thoroughly presented. The experimental results obtained on the weld dataset with our proposed model are validated using GDXray database. The results obtained also in the validation test set are compared to the others offered by DCNN models, which show a best performance in less time. This can be seen as evidence of the strength of our proposed classification model.