Segmenting Roads from Aerial Images: A Deep Learning Approach Using Multi-Scale Analysis

Abstract

Road map generation requires frequent map updates due to the irregular infrastructural changes. Updating a manual road map is a lengthy process, whereas using aerial or remote sensing (RS) requires less time for the update. However, road extraction becomes more complex due to the similar texture appearance of building top roofs, shadows, and occlusion due to trees. The occluded roads appear as discontinuous road patch in segmented image of updated maps. In this paper, we propose a deep learning method that uses multi-scale analysis for road feature extraction. The dilated inception module (DI) in the up and down sampling paths of network extracts the local and global texture patterns of the road. Furthermore, we also utilize the pyramid pooling module (PP) which has average and max pooling to study the global contextual information under the shadow regions. In the proposed architecture, first, the road in the aerial images is segmented along with the tiny non-road segments. Next, the post processing, which exploits the geometrical shape features, is utilized for filtering the tiny non-road noises. The performance of proposed network is validated on using the publicly available Massachusetts road data by comparing with the other models available in literature.