Deblurring of Optically Aberrated Satellite Imagery With Deep Learning (UNET)
Abstract
Satellite imaging performance can degrade due to optical aberrations. To maximize a satellites imaging output over its useful lifespan, deep learning presents a cost-effective alternative to traditional adaptive optics for deblurring satellite images. This is because deep learning is essentially a post-processing technique that relies on algorithms and a large dataset. This research focuses on applying deep learning algorithms based on the UNET Convolutional Neural Network, which is widely used in the bio-medical imaging field, to deblur optically aberrated satellite imagery. The XVIEW dataset, which is composed of images taken by the Worldview-3 satellite, is used. The XVIEW images are then simulated with optical aberrations (defocus and spherical) using Zernike polynomials. The blurred images are subsequently deblurred with UNET and UNET variants (UNET++and UNET3+) before final performance evaluation with various image quality metrics. The results showed that (1) UNET algorithms can effectively deblur optically aberrated satellite images, and (2) UNET3+ modified with additional convolutional layers (deep-UNET3+) provided the best deblurring performance. Based on the positive results, this thesis recommends that the UNET algorithm be applied on actual field cases of optically aberrated satellite imagery and be further developed to perform better even on super-resolution applications.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 2022
- Accession Number
- AD1201754
Entities
People
- Jun J. Siew
Organizations
- Naval Postgraduate School