In the last 5 years generative algorithms for synthesising photorealistic images have become much more robust and harder to differentiate from real images, both by humans and computers. While much of the current research into the generation and subsequent detection of algorithm-based images has focused on face and object data, other important image domains have received less attention. One area that remains unexplored and open to exploitation is the generation of Earth Observation Data (EO), such as satellite and aerial imagery.
In partnership with the Defense, Science and Technology Laboratory (Dstl), this project aims to combine computer vision techniques with human factors research into expert knowledge to find novel methods for detecting AI generated aerial imagery. Many current detection methods rely on pretrained CNN based object recognition models but often falter in being able to correctly classify less common image types such as EO data due to the differences in features present in datasets of mainly objects and faces (e.g. EO data lacks central features such as facial features or single object). One novel approach to this problem is to utilize domain specific, human expert knowledge with our research finding that there is a significant correlation between self-identified expertise and performance in correctly classifying EO images as real or false.
The end goal of this research project is to identify and use expert knowledge to facilitate more accurate detection models and methods. The project also seeks to to further explore the differences and similiarities in computational and human detection in order to increase our understanding of how these methods work.
This author is supported by the Horizon Centre for Doctoral Training at the University of Nottingham (RCUK Grant No. EP/L015463/1) and DSTL.