GAN-Based Object Removal in High-Resolution Satellite Images

TL;DR

This paper introduces a GAN-based method for object removal in high-resolution satellite images, addressing realism issues and data scarcity, and demonstrates the difficulty of detecting such forgeries.

Contribution

It presents a novel approach using conditional GANs trained on Canny feature images for realistic object removal and creates a new forged satellite image dataset.

Findings

Forged images are highly challenging to distinguish from real images.

The proposed method improves realism in satellite image manipulation.

Fake image detectors struggle to identify the forged satellite images.

Abstract

Satellite images often contain a significant level of sensitive data compared to ground-view images. That is why satellite images are more likely to be intentionally manipulated to hide specific objects and structures. GAN-based approaches have been employed to create forged images with two major problems: (i) adding a new object to the scene to hide a specific object or region may create unrealistic merging with surrounding areas; and (ii) using masks on color feature images has proven to be unsuccessful in GAN-based object removal. In this paper, we tackle the problem of object removal in high-resolution satellite images given a limited number of training data. Furthermore, we take advantage of conditional GANs (CGANs) to collect perhaps the first GAN-based forged satellite image data set. All forged instances were manipulated via CGANs trained by Canny Feature Images for object removal. As part of our experiments, we demonstrate that distinguishing the collected forged images from authentic (original) images is highly challenging for fake image detector models.