EIP

EIP performs unsupervised super-resolution of remote sensing imagery by using an enhanced image prior and a Generative Adversarial Network (GAN) to reconstruct high-resolution images without paired low/high-resolution datasets.

Key Features:

• Unsupervised Learning Framework: Operates without paired low/high-resolution image datasets, enabling super-resolution when paired data are unavailable.
• Generative Adversarial Network (GAN): Employs a GAN that takes random noise maps as input and uses adversarial training to reconstruct satellite image super-resolution.
• Enhanced Image Prior: Converts a reference image into a latent-space representation that guides texture and structural information in the generated super-resolved images.
• Recurrent Updating Strategy: Updates input noise in the latent space using a recurrent strategy so generated high-resolution images progressively incorporate more detailed features from the reference image.

Scientific Applications:

• Performance Improvement: Demonstrated significant quantitative and qualitative improvements over existing unsupervised SR methods.
• Draper Dataset Experiments: Extensive experiments on the Draper dataset validated effectiveness in enhancing image resolution.
• SuperView-1 Satellite Imagery: Application to SuperView-1 satellite imagery highlights potential to improve remote sensing imagery resolution beyond supervised algorithms.

Methodology:

Unsupervised training without paired low/high-resolution images; GAN architecture with random noise map inputs and adversarial training; reference image conversion into a latent-space enhanced image prior; recurrent updating of input noise in the latent space.