Real-time Power Line Detection Network using Visible Light and Infrared Images
Hyeyeon Choi, Gyogwon Koo, Bum Jun Kim, Sang Woo Kim · 2019
In the electrical industry, power line detection is an important task for safety of autonomous drone, which assists inspection of the electrical facilities. In this paper, we proposed real-time power line detection network, which is based on deep learning. Building on existing fast segmentation convolutional neural network (Fast-SCNN), which showed competitive performance with fast processing time for image segmentation, we modified its classifier part by inserting additional convolution layers and transposed convolution layer to improve the performance. Furthermore, unlike previous studies that utilized one-kind of sensor image, stacked four-channel image of visual light (VL) and contrast enhanced infrared (IR) were introduced as an input of our modified network. Experimental results on our model showed that the mean intersection over union of multi sensor image (VL+IR) was 6.7% value higher than the Fast-SCNN with single sensor image (VL) and real-time processing speed of 24 frames per second on the drone captured outdoor dataset.