Evaluation of performance of VDSR super resolution on real and synthetic images

D. Vint, G. Di Caterina, J. J. Soraghan, R. A. Lamb, D. Humphreys

Research output: Contribution to conferencePaper

46 Downloads (Pure)

Abstract

This paper presents an evaluation of the suitability of the Very Deep Super Resolution (VDSR) architecture, to increase the spatial resolution of lower quality images. For this aim, two sets of tests are performed. The former being on real life images to determine the networks ability to improve low resolution images. The second test is performed on images of a resolution chart, and therefore synthetic. This is to analyse the frequency response of the network. For each test, three metrics are used to assess image quality. These are the Peak Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM) and Modulation Transfer Function (MTF). Experimental results show that the VDSR network is able to increase the quality of the images within the first test in all three metrics, therefore showing that the network is suitable for super resolution. The second test provides more information on the limitations of the network when given a high contrast image, and the resulting ringing effects it can create. Therefore results in PSNR/SSIM values are not improved over the low resolution images, however they have a higher MTF curve as well as more visually sharp images.

Original languageEnglish
Number of pages5
Publication statusPublished - 9 May 2019
EventSensor Signal Processing for Defence 2019 - Brighton, United Kingdom
Duration: 9 May 201910 May 2019

Conference

ConferenceSensor Signal Processing for Defence 2019
Abbreviated titleSSPD'19
CountryUnited Kingdom
CityBrighton
Period9/05/1910/05/19

Fingerprint

Optical transfer function
Image resolution
Image quality
Signal to noise ratio
Frequency response

Keywords

  • diffraction
  • deep learning
  • VDSR
  • SSIM
  • MTF

Cite this

Vint, D., Di Caterina, G., Soraghan, J. J., Lamb, R. A., & Humphreys, D. (2019). Evaluation of performance of VDSR super resolution on real and synthetic images. Paper presented at Sensor Signal Processing for Defence 2019, Brighton, United Kingdom.
Vint, D. ; Di Caterina, G. ; Soraghan, J. J. ; Lamb, R. A. ; Humphreys, D. / Evaluation of performance of VDSR super resolution on real and synthetic images. Paper presented at Sensor Signal Processing for Defence 2019, Brighton, United Kingdom.5 p.
@conference{4a8e46894a01432b8e119b3dae6d58b9,
title = "Evaluation of performance of VDSR super resolution on real and synthetic images",
abstract = "This paper presents an evaluation of the suitability of the Very Deep Super Resolution (VDSR) architecture, to increase the spatial resolution of lower quality images. For this aim, two sets of tests are performed. The former being on real life images to determine the networks ability to improve low resolution images. The second test is performed on images of a resolution chart, and therefore synthetic. This is to analyse the frequency response of the network. For each test, three metrics are used to assess image quality. These are the Peak Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM) and Modulation Transfer Function (MTF). Experimental results show that the VDSR network is able to increase the quality of the images within the first test in all three metrics, therefore showing that the network is suitable for super resolution. The second test provides more information on the limitations of the network when given a high contrast image, and the resulting ringing effects it can create. Therefore results in PSNR/SSIM values are not improved over the low resolution images, however they have a higher MTF curve as well as more visually sharp images.",
keywords = "diffraction, deep learning, VDSR, SSIM, MTF",
author = "D. Vint and {Di Caterina}, G. and Soraghan, {J. J.} and Lamb, {R. A.} and D. Humphreys",
year = "2019",
month = "5",
day = "9",
language = "English",
note = "Sensor Signal Processing for Defence 2019, SSPD'19 ; Conference date: 09-05-2019 Through 10-05-2019",

}

Vint, D, Di Caterina, G, Soraghan, JJ, Lamb, RA & Humphreys, D 2019, 'Evaluation of performance of VDSR super resolution on real and synthetic images', Paper presented at Sensor Signal Processing for Defence 2019, Brighton, United Kingdom, 9/05/19 - 10/05/19.

Evaluation of performance of VDSR super resolution on real and synthetic images. / Vint, D.; Di Caterina, G.; Soraghan, J. J.; Lamb, R. A.; Humphreys, D.

2019. Paper presented at Sensor Signal Processing for Defence 2019, Brighton, United Kingdom.

Research output: Contribution to conferencePaper

TY - CONF

T1 - Evaluation of performance of VDSR super resolution on real and synthetic images

AU - Vint, D.

AU - Di Caterina, G.

AU - Soraghan, J. J.

AU - Lamb, R. A.

AU - Humphreys, D.

PY - 2019/5/9

Y1 - 2019/5/9

N2 - This paper presents an evaluation of the suitability of the Very Deep Super Resolution (VDSR) architecture, to increase the spatial resolution of lower quality images. For this aim, two sets of tests are performed. The former being on real life images to determine the networks ability to improve low resolution images. The second test is performed on images of a resolution chart, and therefore synthetic. This is to analyse the frequency response of the network. For each test, three metrics are used to assess image quality. These are the Peak Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM) and Modulation Transfer Function (MTF). Experimental results show that the VDSR network is able to increase the quality of the images within the first test in all three metrics, therefore showing that the network is suitable for super resolution. The second test provides more information on the limitations of the network when given a high contrast image, and the resulting ringing effects it can create. Therefore results in PSNR/SSIM values are not improved over the low resolution images, however they have a higher MTF curve as well as more visually sharp images.

AB - This paper presents an evaluation of the suitability of the Very Deep Super Resolution (VDSR) architecture, to increase the spatial resolution of lower quality images. For this aim, two sets of tests are performed. The former being on real life images to determine the networks ability to improve low resolution images. The second test is performed on images of a resolution chart, and therefore synthetic. This is to analyse the frequency response of the network. For each test, three metrics are used to assess image quality. These are the Peak Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM) and Modulation Transfer Function (MTF). Experimental results show that the VDSR network is able to increase the quality of the images within the first test in all three metrics, therefore showing that the network is suitable for super resolution. The second test provides more information on the limitations of the network when given a high contrast image, and the resulting ringing effects it can create. Therefore results in PSNR/SSIM values are not improved over the low resolution images, however they have a higher MTF curve as well as more visually sharp images.

KW - diffraction

KW - deep learning

KW - VDSR

KW - SSIM

KW - MTF

UR - https://sspd.eng.ed.ac.uk/

M3 - Paper

ER -

Vint D, Di Caterina G, Soraghan JJ, Lamb RA, Humphreys D. Evaluation of performance of VDSR super resolution on real and synthetic images. 2019. Paper presented at Sensor Signal Processing for Defence 2019, Brighton, United Kingdom.