Accurate Semantic Segmentation of Aerial Imagery Using Attention Res U-Net Architecture

Abstract

Semantic segmentation of photos holds importance in various fields such, as security, emergency response, agriculture, urban planning and surveillance. Researchers have shown increased interest in automating the analysis of photographs without intervention due to its critical role across multiple disciplines. A CNN variant called U-Net has been successful in segmenting photos beyond the domain. However U-Nets limited number of layers hampers its ability to extract information from photos accurately and can lead to flawed boundaries for objects with complex features. In this study we introduce an architecture called Attention Res U-Net that addresses challenges in semantic segmentation such as low accuracy, for high resolution images and incorrect object boundary recognition. Our approach focuses on refining object boundaries. This approach successfully balances the extraction of attributes with the preservation of space thereby significantly improving aerial image analysis.