基于条件生成式对抗网络和AlexNet-BiLSTM模型的变电设备缺陷检测

Detection of Substation Equipment Defect Based on Conditional Generative Type Adversarial Networks and AlexNet-BiLSTM Model

针对巡检机器人拍摄变电设备图像含噪严重、图像模糊和分辨率低影响设备缺陷检测的问题,提出一种基于条件生成式对抗网络和AlexNet-BiLSTM的变电设备缺陷检测方法,实现变电设备缺陷定位与辨识。首先,通过条件生成式对抗网络将模糊图像转换成清晰图像;其次,为了避免大量超参数的设置,提高网络的训练速度,引入迁移学习思想,采用变电设备图像训练预训练的AlexNet网络,通过AlexNet网络提取图像的高维特征向量,利用双向长短时记忆网络(bi-directional long short-term memory, BiLSTM)对提取的特征向量进行分类;最后,在R-CNN框架下完成变电设备缺陷的标注和辨识。试验结果表明,所提方法复原的图像主观视觉效果良好,客观评价指标高,提高了变电设备缺陷检测准确率。

According to inspection robots shooting substation equipment with serious noise,blurred images,and low-resolution affecting device defect detection problems,a substation equipment defect detection method is proposed based on conditional generative adversarial network and AlexNet-BiLSTM to realize substation equipment defect localization and identification.Firstly,the blurred images are converted to clear images by the conditional generative type adversarial networks.Secondly,in order to avoid a large number of hyperparameter settings and improve the training speed of the network,the migration learning idea is introduced.The images of substation equipment are used to train pre-trained AlexNet networks.The AlexNet network extracts the high-dimensional feature vectors of the image,and the extracted feature vectors are classified by using bi-directional long short-term memory(BiLSTM)network.Finally,the labeling and identification of substation equipment defects are completed in the R-CNN framework.The test results show that the proposed method has a good subjective visual effect of image recovery,a high objective evaluation index,and enhances substation equipment defect detection precision rate.