基于改进SinGAN的电力线巡检异物数据增强技术

Data enhancement technology of power line inspection foreign object based on improved SinGAN

针对电力线异物识别模型能使用的数据集较少,并且传统单幅自然图像的生成式模型(SinGAN)模型生成数据与异物识别模型匹配度不高、质量不佳、耗时过久的问题,提出了改进SinGAN模型。在改进SinGAN模型基础上加入仿射变换单元、大小变换单元进一步增强数据集,加入图像滤波单元提高电力线异物识别模型所需数据质量。并通过改进SinGAN反向传播训练过程和SinGAN的单精度生成器结构提升模型生成质量,减少所用时长。实验结果表明,经50次实验后,改进SinGAN的平均弗雷谢特起始距离(Fréchet inception distance, FID)为91.375,平均训练时长1.21 h。分别比传统SinGAN降低了27.247%和87.31%。改进SinGAN与其他主流生成式对抗网络相比有更好的异物数据生成能力,可以增强电力线异物识别模型所需数据,具有优越性。

Aiming at the problem that the power line foreign object recognition model can use fewer data sets, and the traditional SinGAN model generated data does not match the foreign object recognition model, the quality is poor, and it takes too long, the improved SinGAN model is proposed. Based on the improved SinGAN model, an affine transformation unit and a size transformation unit are added to further enhance the data set, and an image filtering unit is added to improve the data quality required by the power line foreign object recognition model. By improving the SinGAN back propagation training process and SinGAN’s single-precision generator structure, the quality of model generation is improved and the time spent is reduced. Experimental results show that after 50 experiments, the average Frechet starting distance score(FID) of the improved SinGAN is 91.375, and the average training time is 1.21 h. Compared with traditional SinGAN, it is reduced by 27.247% and 87.31% respectively. Compared with other mainstream generative adversarial networks, improved SinGAN has better foreign object data generation capability, which can enhance the data required by the power line foreign object recognition model, and has superiority.