What is super-resolution reconstruction?
The aim of the super-resolution reconstruction (SRR) techniques is to improve the quality and increase the resolution of images (upscale images) while restoring as many details as possible from the source image. Additional goals include: preserving sharp edges, limiting the number of unwanted image distortions after the transformations, producing a visually appealing image.
The SRR developed by KP Labs is based on deep neural networks and advanced statistical methods offering top reconstruction performance. What’s more, application of evolutionary algorithms allowed us to improve these methods further.
After
Before
SRR applications
Satellite imaging
Imaging from nano-satellite constellations or other low to medium resolution imagery
Machine vision systems
Production monitoring and inspection and control systems
Pattern recognition
Bar codes (1D) or QR codes (2D) and OCR applications
SRR comparison
The SRR methods can use a single, low-resolution image (single-image SRR) or a number of images showing the same scene or object (multiple-image SRR).
| Single-image SRR | Multiple-image SRR | |
|---|---|---|
| Focus on | Improvement of image appearance | Fusion of images and extraction of new information |
| Number of input images | One | Even only a few up to tens and more |
| Overall quality improvement | Little | Significant |
Do you want to find out more?
SRR technique was the result of 3 projects
Satellite image spatial resolution enhancement (SISPARE)
The objective of the SISPARE project was to implement the algorithms for super-resolution reconstruction and to validate them for satellite images.
Super-resolution reconstruction of satellite images using deep convolutional neural networks (SUPERDEEP)
The objective of the project was to explore the capabilities of deep neural networks for super-resolution reconstruction of satellite images.
Deep learning-based multiple-image super-resolution for Sentinel-2 data (DEEPSENT)
The objective of this project is to enhance the capacities of super-resolution reconstruction applied to multispectral Sentinel-2 images, especially if multiple images of the same region, captured at a different time, are available.