摘要
针对不可抗力因素造成无人机航拍绝缘子图片模糊、绝缘子目标检测率较低的问题,提出了一种基于Wasserstein距离优化的生成式对抗网络(WGAN)图片去模糊的绝缘子目标检测方法。首先在WGAN训练过程中引入残差网络,使得生成的绝缘子图片更加清晰;其次在损失函数中引入Wasserstein距离以保证训练过程的稳定性;最后通过优化模型的训练过程,使得生成的绝缘子图片细节还原度更高。绝缘子图片去模糊化实验结果表明,所提方法在结构相似性与峰值信噪比等评价指标上均高于基于卷积神经网络与深度多尺度卷积神经网络等图像去模糊算法。另外,将利用所提方法生成的绝缘子图片与模糊绝缘子图片划分为3组,采用改进的基于区域建议的卷积神经网络目标检测算法分别进行目标检测实验,精确度均值分别提高了5.77%、6.73%与5.98%,有效提高了绝缘子的目标检测率。
Due to the force majeure factor,the aerial images of insulator took by drone is fuzzy and the object detection rate of insulator is low. Aiming at this problem,an insulator object detection method based on image deblurring by WGAN(Wasserstein-Generative Adversarial Network) is proposed. The residual network is introduced in the WGAN training process,which makes the generated insulator picture clearer. Then the Wasserstein distance is introduced into the loss function to ensure the stability of the training process.Finally,by optimizing the training process of the model,the reduction degree of generated insulator pictures’ details is improved. The results of defuzzification experiments show that the proposed method is superior to traditional deblurring algorithms such as CNN(Convolutional Neural Network) and DM-CNN(Deep Multiscale Convolutional Neural Network) in terms of structural similarity and peak signal-to-noise ratio. In addition,the insulator image and the fuzzy insulator image generated by the proposed method are divided into three groups,the target detection experiment is carried out by Faster-RCNN(Faster Region-Convolutional Neural Network) target detection algorithm,with the average accuracy increased by 5.77 %,6.73 % and5.98% respectively,which improves the insulator target detection rate of the insulator effectively.
作者
王德文
李业东
WANG Dewen;LI Yedong(School of Control and Computer Engineering,North China Electric Power University,Baoding 071003,China;Engineering Research Center of Intelligent Computing for Complex Energy Systems,Ministry of Education,Baoding 071003,China)
出处
《电力自动化设备》
EI
CSCD
北大核心
2020年第5期188-194,共7页
Electric Power Automation Equipment
基金
国家自然科学基金资助项目(51677072)。
