Detection of Transmission Towers and Insulators in Remote Sensing Images with Deep Learning

Intelligent inspections of high-voltage electronic power grids with high-altitude aerial or satellite remote sensing images (RSIs) have attracted more and more attention. The detection of transmission tower and insulator in smart electronic power is of great importance. Traditional image recognition based methods have been proven to be difficult to complete this task effectively and efficiently, lots of deep learning based methods have been adopted due to their promising performance. However, collecting a large number of labeled aerial/satellite imaging data for deep learning requires a lot of manpower/financial costs and the deep models trained on small size of samples are often easy to over-fit. For this reason, it has practical significance to study the automatic detection of towers and insulators in the case of small samples. Aiming at the problem of object detection under small samples, a deep learning framework for simultaneous towers and insulators detection in RSIs based on Faster-RCNN and neural style image synthesis is proposed. Firstly, to alleviate the small sample size problem, a sample generation method based on neural style transfer and alpha channel image fusion techniques is proposed, which randomly combines the foreground towers and background images to expand the training data set. Secondly, upon the expanded training data, an object detection model for towers and insulators based on Faster-RCNN is further trained. Experiments show that the object detection model trained with the extended training data has better generalization performance and can better suppress false alarms.