基于Cycle-GAN的绝缘子图像生成方法

Image Generation Method for Insulators Based on Cycle-GAN

深度学习模型应用于输电线路绝缘子目标检测时,在训练样本方面存在公开样本集缺乏和优质样本不足的问题,为此提出一种基于循环一致性生成对抗网络(cycle-generative adversarial networks,Cycle-GAN)的绝缘子图像生成方法。首先分析绝缘子样本集,对绝缘子图像基于背景色彩特征进行风格域划分;之后在划分好的绝缘子风格域样本集基础上,采用Cycle-GAN生成绝缘子图像样本;最后,搭建分类网络验证生成图像用于扩充的有效性,并进一步探究了生成图像不同扩增比例对分类性能的影响。结果表明:绝缘子生成样本可一定程度上替代真实样本;生成图像不同扩充比例对网络性能影响不同,当扩充比例在40%~50%时,分类网络性能提升效果最佳。

In the application of deep learning in target detection of transmission line insulators,there are problems of lacking public sample set and high quality samples in sample training.Therefore,this paper proposes a method of insulator image generation based on cycle-generative adversarial networks(Cycle-GAN).It firstly analyzes the sample set and divides the style domains for the insulator according to color characteristics of background.Afterwards,it uses Cycle-GAN to generate images samples of the insulators.Finally,it establishes a classification network to verify effectiveness of the generated images for expansion,and explores influence of different image expansion scales on classification performances.It concludes that the generated insulator images can replace the real samples to some extent,and different expansion ratio of generated image has different effects on network performance.When the expansion sample accounts for 40%to 50%of the real sample,the classification network performance can be most effectively improved.