Regularization of Building Boundaries in Satellite Images using Adversarial and Regularized Losses

TL;DR

This paper introduces a neural network-based method for refining building boundaries in satellite images, producing regularized and visually appealing footprints while maintaining accuracy comparable to Mask R-CNN.

Contribution

The proposed approach combines adversarial and regularized losses to improve boundary regularity in satellite image segmentation, outperforming traditional models in boundary quality.

Findings

Achieves boundary regularization comparable to Mask R-CNN in accuracy

Produces more visually regular and pleasing building footprints

Maintains high segmentation performance with improved boundary quality

Abstract

In this paper we present a method for building boundary refinement and regularization in satellite images using a fully convolutional neural network trained with a combination of adversarial and regularized losses. Compared to a pure Mask R-CNN model, the overall algorithm can achieve equivalent performance in terms of accuracy and completeness. However, unlike Mask R-CNN that produces irregular footprints, our framework generates regularized and visually pleasing building boundaries which are beneficial in many applications.