Double Discriminator Generative Adversarial Networks And Its Application In Detecting Nests Built In The Catenary

The catenary anomaly detection is of much importance in the electrification construction and operation management of high-speed railway.The detection method based on computer vision is a primary technical approach of catenary anomaly detection,which we have adopted a variety of data analysis methods to research.In the image-based detection of catenary anomalies,detection of bird nest anomalies is a typical situation.The bird nest built on the electrified railway catenary can easily cause serious faults such as contact net tripping and insulator breakdown.At present,the detection of bird nest mainly depends on manual inspection,which is costly and inefficient.Using video image processing to carry out automatic inspection is a very important technology development direction.However,the image data containing nests is only a small portion of total data,which makes nest detection a typical problem of imbalanced data classification.For using machine learning algorithm to solve imbalanced data classification,the learning ability of data features is of much importance.The Generative Adversarial Networks(GANs)can learn prosperous data features from unlabeled data,which has been widely confirmed and applied.In this paper,based on the application background of the catenary nest detection,we started the research work of the paper from three aspects,including image data preprocessing,object detection,and unbalanced data classification.There are a lot of images with fog,underexposure,or excessive issues in the catenary inspection video.Due to such issues,dark channel prior(DCP)are used to enhance the catenary inspection image,and the results are remarkable.In order to effectively extract the key section from the catenary inspection image,we use Faster R-CNN to detect as well as extract catenary pillar from the catenary inspection image.By using the transfer learning method,we have achieved desired detection accuracy based on a small number of labeled samples.Due to the limitation of GANs' structure and theory,it's not an ideal model for image classification.The GANs model is improved with respect to the image classification tasks,which is named the Double Discriminator Generative Adversarial Networks(DDGANs).With the DDGANs,the classification results of nest detection are satisfactory,and it is also an effective semi-supervised learning model.Experiments on the MNIST standard dataset show that the accuracy and convergence rate have been significantly improved compared with other models.