Weld defect segmentation and detection using semi-supervised transfer learning-based Generative Adversarial Networks

Traditional Convolutional Neural Networks (CNNs) face challenges in limited labelled datasets and overfitting issues for weld image semantic segmentation. This paper proposes a semi-supervised GAN-based transfer learning network (SGT-Net) for weld defect segmentation and detection effectively. The SGT-Net utilises a Deep Convolutional Generative Adversarial Network (DCGAN) for unsupervised feature extraction from unlabelled augmented data. The discriminator weights of DCGAN are transferred to a secondary network entitled UNet-G for supervised training with limited labelled data. The principles of Generative Adversarial Networks (GANs) are smartly designed for weld defect segmentation, and the adversarial loss further enhances the accuracy of the segmentation network. This strategy effectively integrates unlabelled data features to enhance segmentation accuracy and generalisation. Experimental results demonstrate that the SGT-Net have a Dice coefficient of 91.7% and an IoU score of 93.3%, which is higher than traditional models. In additional, the weld defect detection and quantification are in agreement with American Welding Society standards, which further demonstrates the industrial applicability and effectiveness.