Weld Defect Detection Based on Deep Learning Method

Abstract

Welding is an important joining technology but the defects in welds wreck the quality of the product evidently. Due to the variety of weld defects' characteristics, weld defect detection is a complex task in industry. In this paper, we try to explore a possible solution for weld defect detection and a novel image-based approach is proposed using small X-ray image data sets. An image-processing based data augmentation approach and a WGAN based data augmentation approach are applied to deal with imbalanced image sets. Then we train two deep convolutional neural networks (CNNs) on the augmented image sets using feature-extraction based transfer learning techniques. The two trained CNNs are combined to classify defects through a multi-model ensemble framework, aiming at lower false detection rate. Both of the experiments on augmented images and real world defect images achieve satisfying accuracy, which substantiates the possibility that the proposed approach is promising for weld defect detection.