Publication

Abstract : Accurate road extraction from satellite images is critical for applications such as urban planning, navigation systems, and disaster response. Recent advancements in computer vision has paved the way to achieving this task efficiently. However, maintaining both high accuracy and low computational complexity remains a challenge. To address this we propose a feature selective and lightweight UNet architecture (DepthLiteNet). It combines a reduced depth UNet structure with a novel Feature Selection Module (FSM) and aims to improve the precision of a standard UNet model along with reducing its complexity. The number of parameters is 35% less than standard UNet. Three different datasets — the Massachusetts Roads Dataset, the DeepGlobe Road Extraction Dataset, and the CHN6-CUG Roads Dataset — were used to validate DepthLiteNet. When compared to state of the art techniques, our results demonstrates notable and consistent increase in Intersection Over Union (IOU) by values 0.98%, 5.62%, and 2.7% for the Massachusetts Roads Dataset, the DeepGlobe Road Extraction Dataset, and the CHN6-CUG Roads Dataset respectively.