CycleGANs for Semi-Supervised Defects Segmentation

At present, many vision-based inspection methods are widely using for quality control in different fields. And the deep learning method has made a magnificent breakthrough in a variety of computer vision tasks, mainly through the use of largescale annotated datasets. Utilizing these progress is an option to improve defect segmentation performance. However, in the field of vision-based non-destructive testing (NDT), obtaining large scale annotated datasets is a great challenge. In this paper, a fully convolution neural network (FCN) is supervised trained using a small number of pixel-level annotated data for defect segmentation. Simultaneously, Cycle-Consistent Generative Adversarial Networks (CycleGANs) are used to learn the segmentation in an unsupervised way as a supplement. The requirement of annotated data is then reducing by utilizing many un-annotated data. Experiments on the published GDXray dataset show that the framework based on CycleGANs is effectiveness for defect image segmentation using only a few labelled samples.