High-Resolution Aerial Image Restoration with Latent Diffusion Models

This paper aims to improve the performance of diffusion models in high-resolution unmanned aerial vehicle (UAV) aerial image restoration tasks. We propose an efficient image restoration algorithm based on diffusion models, named Latent-IRSDE, which consists of two parts: pretrained encoder-decoder networks and diffusion-based image restoration networks. Diffusion-based image restoration networks have the remarkable ability to generate high-quality aerial images with low-quality visions. Pretrained encoder-decoder networks transform high-resolution aerial images into small latent images. We leverage the frozen internal representations of pretrained encoder-decoder networks to perform an efficient high-resolution aerial images restoration. We select some data from an open-source dataset to create our IR-UAV2UAV dataset for evaluating our method. The experiments show that our proposed method can restore high-resolution aerial images and achieve fastest inference time in quantitative comparisons of image deblurring and deraining tasks on IR-UAV2UAV dataset with images sized at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1920 \times 1080\ \text{pixels}$</tex>.