Defect Image Generation Algorithm of Commutator Based on Generative Adversarial Networks

Abstract

At present, there have been many studies on the application of deep learning algorithms in defect detection work. Based on the object detection algorithms introduced earlier, a defect image recognition system is constructed to detect various types of defects in the converter, which can save manpower and material resources and improve the safety of this work. However, deep learning algorithms require a certain amount of valid samples to be effective. However, in actual operation, it is difficult to collect defect images of the commutator, which means that there are situations where the requirements of deep learning detection algorithms cannot be met. This article aims to explore a feasible data augmentation scheme, which is to generate effective samples through generative adversarial networks. This article will investigate two classic image generation methods based on Generative Adversarial Networks (GANs). And apply it to the field of generating defect images for commutators, propose a method for generating defect images in a directional manner, and compare it with the above two methods. The experimental results show that our method has a higher FID value compared to the above two methods, and has a more similar data distribution compared to real images.