POWER LINES DETECTION AND SEGMENTATION IN MULTI-SPECTRAL UAV IMAGES USING CONVOLUTIONAL NEURAL NETWORK

In this paper, detection, and segmentation of power line in Unmanned Aerial Vehicles (UAV) multi-spectral images using convolutional neural network is proposed. Initially, the multi-spectral images captured from UAV were calibrated and pre-processed, following which they were fed into deep CNN for semantic segmentation to perform a binary classification; each pixel was assigned either of the two classes - "power line" or " no power line". Semantic segmentation was performed with different networks such as U-Net, SegNet and PSPNet. Qualitative (visual inspection) and quantitative analysis of the results showed that U-Net outperformed other networks with an overall accuracy of around 99% with a competitive execution latency, making it useful for real time analysis of power lines from UAV data.