期刊
APPLIED SOFT COMPUTING
卷 114, 期 -, 页码 -出版社
ELSEVIER
DOI: 10.1016/j.asoc.2021.108130
关键词
Generative adversarial network; Tree structure; Style transfer; Sample augmentation; Gear pitting detection
资金
- National Key R&D Program of China [2018YFB2001300]
- National Natural Science Foundation of China [52175075]
- graduate scientific research and innovation foundation of Chongqing, China [CYB19007]
This study introduces a novel Cycle Generative Adversarial Network based on a symmetric tree structure (Tree-CycleGAN) for augmenting gear pitting samples and improving gear pitting detection accuracy. Results demonstrate that Tree-CycleGAN outperforms other methods in terms of image quality and diversity, and its combination with U-net for gear pitting detection shows significant improvements in performance.
Visual detection is an available approach for measuring gear pitting. Unfortunately, the number of gear pitting images is limited, resulting in that the detection accuracy of gear pitting is unsatisfactory. In order to augment gear pitting samples with different styles, a novel Cycle Generative Adversarial Network based on a symmetric tree structure (Tree-CycleGAN) is proposed. In Tree-CycleGAN, a new type of generator with tree structure named tree generator is designed to produce various types of high quality target samples from the source-domain samples, and a maximum diversity loss is constructed to enlarge the difference between two arbitrary branches; then a similar tree reconstructor is designed for translating target samples into source samples. Two discriminators are designed for making the generated images approximate to the target images in two cyclic processes. Via inception score, structural similarity indexes and peak-signal-to-noise ratio, the quality and diversity of images obtained by Tree-CycleGAN are evaluated. Comparative results show the superiority of Tree-CycleGAN over other domain adaptation GANs. The proposed Tree-CycleGAN combined with U-net has been successfully applied to gear pitting detection. Experimental results prove that the proposed methodology precedes the basic U-net method without sample augmentation and the method based on CycleGAN and U-net. (C) 2021 Elsevier B.V. All rights reserved.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据