Power Line Segmentation in Aerial Images Using Convolutional Neural Networks

Visual inspection of transmission and distribution networks is often carried out by various electricity companies on a regular basis to maintain the reliability, availability, and sustainability of electricity supply. Till date the widely used technique for carrying out an inspection is done manually either using foot patrol and/or helicopter operated manually. However, recently due to the widespread use of quadcopters/UAVs powered by deep learning algorithms, there have been requirements to automate the visual inspection of the power lines. With this objective in mind, this paper presents an approach towards automatic autonomous vision-based power line segmentation in optical images captured by Unmanned Aerial Vehicle (UAV) using deep learning backbone for data analysis. Power line segmentation is often considered as a first step required for power line inspection. Different state-of-the-art semantic segmentation techniques available in the literature have been used and a comparative analysis has been done in terms of the Jaccard index on two different power line databases. This paper also presents a new power line database captured using UAV along with the baseline results. Experimental results show that out of the four deep learning-based segmentation architectures used in our experiments the Nested U-Net architecture out-performed others in terms of line segmentation accuracy in various background scenarios.