| 引用本文: | 芦肇基,沈艳霞,谭永强.基于多结构融合WGAN的模糊绝缘子图像复原方法研究[J].电力系统保护与控制,2024,52(22):166-175.[点击复制] |
| LU Zhaoji,SHEN Yanxia,TAN Yongqiang.A WGAN blur insulator image restoration method based on multi-structure fusion[J].Power System Protection and Control,2024,52(22):166-175[点击复制] |
|
| 摘要: |
| 为解决因不可抗力因素导致无人机航拍绝缘子图像发生运动模糊的问题，提出一种基于多结构融合Wasserstein生成对抗网络(Wasserstein generate adversarial networks, WGAN)的模糊绝缘子图像复原方法。针对模糊图像复原问题，对基于Wasserstein距离的生成对抗网络加以改进，在损失函数中引入梯度惩罚项优化训练过程，保证模型训练的稳定性并提高图像复原质量。在生成网络中融入空洞卷积残差网络和卷积注意力机制，加强神经网络对图像有效特征的学习。实验结果表明，通过与其他算法比较，所提方法在峰值信噪比和结构相似度两种指标上均高于其他算法。对不同算法生成的图像进行比较，证明了该方法能有效提取图像细节特征，提高模糊绝缘子图像的复原质量。采用YOLOv5s目标检测算法进行实验，证明了所提方法对目标检测的准确率有所提升。 |
| 关键词: 绝缘子图像复原 生成对抗网络 残差网络 卷积注意力机制 深度学习 |
| DOI：10.19783/j.cnki.pspc.240115 |
| 投稿时间：2024-01-25修订日期：2024-06-14 |
| 基金项目:国家自然科学基金项目资助(61573167，61572237) |
|
| A WGAN blur insulator image restoration method based on multi-structure fusion |
| LU Zhaoji1,SHEN Yanxia1,TAN Yongqiang2 |
| (1. School of Internet of Things Engineering, Jiangnan University, Wuxi 214122, China; 2. Nanjing Power Supply
Company, State Grid Jiangsu Electric Power Co., Ltd., Nanjing 210000, China) |
| Abstract: |
| There is a problem of motion blur caused by force majeure factors in the unmanned aerial vehicle aerial photography of insulator images. Thus a blur insulator image restoration method based on improved Wasserstein generative adversarial networks (WGAN) with multi structure fusion is proposed. An improved generative adversarial network based on Wasserstein distance is proposed to solve the problem of blur repair, and a gradient penalty is introduced into the loss function to optimize the training process. This ensures the stability of model training and improves the quality of image restoration. A dilated convolution residual network and convolutional attention mechanism are integrated into the generating network to strengthen the learning of effective features of images by the neural network. The results of experiment show that both the peak signal-to-noise ratio and the structural similarity index measure of the proposed method are higher than with other algorithms. The comparison of images generated by different algorithms proves that this method can effectively extract the detailed features and improve the quality of deblur image restoration. YOLOv5s is used for detection experiments. These demonstrate that the method presented enhances the accuracy of object detection. |
| Key words: insulator image restoration generative adversarial networks residual networks convolution attention mechanism deep learning |
