Road detection from high resolution images using fully convolutional network

Abstract

Automatic road extraction has become an important topic due to its wide range of applications like map updating, traffic management, and automatic driving. Roads do not follow the same shape and pattern in all regions. Classification of road and non-road (background) pixel in high resolution (HR) image gives false segmentation many times. These two problems have been addressed using a fully convolutional network using encoder and decoder architecture. It consists of convolutional layers with batch normalisation (BN), activation, and pooling layer. Encoder and decoder help to generate feature maps and predict individual pixels to their class. This network also uses skip connections between encoder and decoder which prevent loss of image information. The network extracts road segments through HR images with high correctness value.