Year: 2024
Type: conference-paper
Abstract: This paper investigates the application of generative models in simulating radar images, highlighting their capacity in generating realistic images from text description. To overcome the usual limitation of image size constraint, we propose a model operating through a series of blocks to progressively enhance the image’s resolution and details by leveraging a pre-trained latent diffusion model. This cascading architecture, structured into three levels, is augmented by integrating a latent scaling technique, enabling a gradual improvement in the scale of the image, thus allowing the generation of larger images. Our method also stands out for incorporating the ControlNet model, maintaining content consistency during the multi-resolution process. After training on aerial SAR images from the ONERA SETHI sensor, we compare our architecture with alternative methods for large-scale image generation using larger pre-trained models and outpainting methods. The results demonstrate a clear enhancement in the scale of simulated radar images while maintaining alignment with the contexts described by the text descriptions.
Authors: Solène Debuysère, Nicolas Trouvé, Nathan Letheule, Elise Colin, Olivier Lévêque
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4404531660