Enhanced image prior for unsupervised remoting sensing super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Super-resolution (SR) tasks traditionally rely on paired low- and high-resolution images, which are often difficult to acquire in real-world scenarios.
• Existing SR methods face limitations due to unknown or non-ideal image acquisition processes, hindering their practical application.

Purpose of the Study:

• To develop a novel unsupervised learning framework for super-resolution that does not require low-high resolution image pairs.
• To enhance the resolution of satellite imagery using a generative adversarial network (GAN) and an enhanced image prior.

Main Methods:

• Proposed the Enhanced Image Prior (EIP) framework, an unsupervised learning approach for satellite image super-resolution.
• Utilized a generative adversarial network (GAN) by feeding random noise maps for SR reconstruction.
• Converted a reference image into latent space to serve as an enhanced image prior, updating input noise with a recurrent strategy to transfer texture and structure.

Main Results:

• EIP demonstrated significant quantitative and qualitative improvements over state-of-the-art unsupervised SR methods on the Draper dataset.
• Experiments on SuperView-1 satellite images showed EIP's potential in enhancing remote sensing imagery resolution, outperforming supervised algorithms in certain aspects.

Conclusions:

• The proposed Enhanced Image Prior (EIP) framework effectively addresses the limitations of paired data in super-resolution tasks.
• EIP shows strong potential for practical applications in improving the resolution of satellite and remote sensing imagery.