Visualization of power lines recognized in aerial images using deep learning


26th IEEE Signal Processing and Communications Applications Conference (SIU), İzmir, Turkey, 2 - 05 May 2018 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/siu.2018.8404196
  • City: İzmir
  • Country: Turkey
  • Keywords: deep learning, convolutional neural networks, saliency map, power line detection
  • Anadolu University Affiliated: Yes


Aerial power lines used to conduct electrical energy pose a significant risk of accidents for aerial vehicles. In this study, we propose a method to highlight the power lines in an aerial image. The visual feedback provided by this method is expected to aid the pilot in avoiding power lines. To implement the method, a convolutional neural network is trained to classify images into ones that contain power lines and not. In the case that this network infers that power lines are present in an image, the partial derivative of the classification loss function with respect to the image is used to produce a type of saliency map. By superposing this saliency map on the original image, the power lines are highlighted. Various backpropagation methods were tested to produce the saliency map, and it has been observed that guided backpropagation has performed best in this task. Since the network used in the proposed method is only trained for classification, the locations of the power lines in the dataset do not need to be annotated.