基于DAGAN的电气设备小样本红外图像生成技术与应用

Infrared Image Generation Technology and Application of Small Sample of Electrical Equipment Based on DAGAN

面向电力设备的红外图像由于其高度敏感性和机密性,缺乏公开数据集,严重制约了面向电力红外图像的深度学习目标检测网络的发展。针对该问题,使用改进后的数据增强生成对抗网络(DAGAN)对数据进行扩充。首先改变其生成器结构,提出一种全新的生成器结构MUL-U-ResNet3+,该结构对网络结构和残差模块都进行了优化,U-Net结构优化为MUL-U-Net3+结构,并将其残差块改变为SandGlass模块,另外其损失函数的梯度惩罚改变为随机梯度惩罚,从而降低网络训练的难度。使用某220 kV和某500 kV变电站的数据对该算法进行测试,试验结果表明数据扩充后的样本在目标检测中有了更好的效果。

Due to its high sensitivity and confidentiality,the infrared image for power equipment lacks public data sets,which seriously restricts the development of deep learning target detection network for power infrared image.To solve this problem,The data was enriched using an improved Data Augmentation Generative Adversarial Network(DAGAN).Firstly,the generator structure was changed,and a new generator structure MUL-U-ResNet3+was proposed.This structure optimized the network structure and residual module.The U-Net structure was optimized to MUL-U-ResNet3+structure,and the residual block was changed to SandGlass module.In addition,the gradient penalty of the loss function is changed to random gradient penalty,thus reducing the difficulty of network training.In this paper,the algorithm was tested with the data of a 220 kV and a 500 kV substation.The experimental results showed that the expanded samples have better effect in target detection.