De-noising and Road Extraction from High-resolution Images using Fully Convolutional Network

Abstract

Automatic road extraction from high-resolution images has gained research interest due to a wide range of applications like autonomous driving, map updates, and traffic management. Roads have different shapes and patterns in different geographic regions. So it becomes challenging to classify the road and non-road (background) pixels in remote sensing images. These issues have been addressed by proposing a fully convolutional network by utilizing an encoder-decoder network. The work was carried out in two phases: pre-processing and road detection from high-resolution images. Firstly, it uses an adaptive median filter to remove the noise components present in the image. Later these images are applied to FCN for detection of roads. The FCN consists of convolutional layers along with batch normalization, pooling, and activation layers. The architecture generates the feature map and classifies the predicated value either as roads or non-roads. It also makes use of skip connections between the encoder and decoder to avoid the loss of information. By using high-resolution images the network extracts the roads with an accuracy of 88.5%.