September 2021
Volume 21, Issue 9
Open Access
Vision Sciences Society Annual Meeting Abstract  |   September 2021
Visual Information Fidelity with better Vision Models and better Mutual Information Estimates
Author Affiliations & Notes
  • Jesus Malo
    Universitat de Valencia
  • BENYAMIN KHERAVDAR
    Khajeh Nasir Toosi University of Technology
  • QIANG LI
    Universitat de Valencia
  • Footnotes
    Acknowledgements  MICINN grant DPI2017-89896, and GVA grant Grisolia/2019/035
Journal of Vision September 2021, Vol.21, 2351. doi:https://doi.org/10.1167/jov.21.9.2351
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Jesus Malo, BENYAMIN KHERAVDAR, QIANG LI; Visual Information Fidelity with better Vision Models and better Mutual Information Estimates. Journal of Vision 2021;21(9):2351. https://doi.org/10.1167/jov.21.9.2351.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

*INTRODUCTION*: Visual Information Fidelity [1] is a subjective image distortion measure based on: (i) a noisy observer model, and (ii) a method to quantify mutual information. Specifically, given two images A (original) and B (distorted), and their responses through the model R(A) and R(B), the visual fidelity is defined as the ratio between the amount of information about A that can be extracted from R(B) and the corresponding amount that can be extracted from R(A). The information about X that is available from Y is quantified via the mutual information, I(X,Y). Therefore: VIF(A,B) = I(A,R(B)) / I(A,R(A)). The original VIF used a too crude vision model (linear wavelet pyramid + Gaussian noise) and a too crude mutual information estimate (based on a simplified image statistics model). In this work we explore how better vision models and better information measures may lead to a better explanation of image quality psychophysics. *METHODS*: we improve the linear vision model in original VIF by using divisive normalization stages for brightness and contrast masking [2], we use a more accurate psychophysical estimate of neural noise [3], and an estimate of mutual information that does not rely on parametric assumptions [4]. *RESULTS AND CONCLUSION*: the correlation with human opinion with images from the TID-2008 database improves from 0.78 (using the original formulation of VIF) up to 0.90 using some of the proposed improvements. This suggests that an appropriate consideration of the neural noise and a non-parametrical measure of mutual information are critical in assessing subjective image quality. *REFERENCES*: [1] H. Sheikh, A. Bovik. IEEE Trans. Im. Proc., 15(2):430–444, 2006 [2] M. Martinez, P. Cyriac, T. Batard, M. Bertalmío & J. Malo. PLOS ONE, 13(10):1–49, 10 2018. [3] J.Esteve, G.Aguilar, M.Maertens, FA.Wichmann & J.Malo. https://arxiv.org/abs/2012.06608, 2020 [4] J. Malo. J. Math. Neurosci., 10, 18 (2020). https://doi.org/10.1186/s13408-020-00095-8

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×