Proc. IEEE Int. Conference on Image Processing, Sep. 25-28, 2016, Phoenix, Arizona USA.

Visual Attention Guided Quality Assessment of Tone-Mapped Images using Scene Statistics

Debarati Kundu and Brian L. Evans

Embedded Signal Processing Laboratory, Wireless Networking and Communications Group, The University of Texas at Austin, Austin, TX 78712 USA
debarati@utexas.edu - bevans@ece.utexas.edu

Paper Draft - PowerPoint Slides - PDF Slides - Software Release

Abstract

Measuring visual quality, as perceived by human observers, is becoming increasingly important in the many applications in which humans are the ultimate consumers of visual information. This paper assesses the visual quality in mapping of high dynamic range (HDR) images to standard dynamic range (SDR) images with 8 bits/color/pixel. In previous work, the Tone-Mapped image Quality Index (TMQI) compares the original HDR image with the rendered SDR image. TMQI quantifies distortions locally and pools them by uniform averaging, in addition to measuring naturalness of the SDR image. For SDR images, perceptual pooling strategies have improved correlation of image quality assessment (IQA) algorithms with subjective scores. The primary contributions of this paper are:

integrating local information-based pooling strategies in the TMQI IQA algorithm,
measuring image naturalness by using mean-subtracted contrast-normalized pixels, and
testing the proposed methods on JPEG compressed tone-mapped images and tone-mapped images for SDR displays using subjective scores.

Questions and Answers

Question #1. Have you considered non-correlation-based IQA measures, e.g. RMSE and T-Test?

Answer: Although we only used correlation-based IQA measures in this paper, we compute root mean squared error (RMSE) RMSE values and reduced chi-squared test of subjective ratings vs. predicted ratings in follow-up papers on HDR image quality assessment:

D. Kundu, D. Ghadiyaram, A. C. Bovik and B. L. Evans, "No-Reference Quality Assessment of High Dynamic Range Pictures", IEEE Transactions on Image Processing, submitted July 3, 2016.
D. Kundu, D. Ghadiyaram, A. C. Bovik and B. L. Evans, "Large-scale Crowdsourced Study for High Dynamic Range Pictures", IEEE Transactions on Image Processing, submitted May 25, 2016.

Question #2: How could you better handle noise in measurements and uncertainty in correlation coefficients?

Answer: This paper uses two previous subjective evaluation studies of 20 participants for tonempapping operators [Yeganeh & Wang, 2013] and 27 participants for compression [Narwaria, Da Silva, Le Callet & Pepion, 2013]. With a smaller number of opinion scores, there can be a wider variability in the ability of an automated algorithm to predict subjective results. After we had submitted our ICIP 2016 paper in January 2016, we collaborated with Alan Bovik and Deepti Ghadiyaram here at UT Austin to compile the ESPL-LIVE HDR database with over 1800 HDR images and obtain over 300,000 opinion scores from more than 5000 participants in a crowdsourced study using Amazon Mechanical Turk. Each HDR image has at least 40 evaluations.

Question #3: How did you choose the parameters in the proposed naturalness measure?

Answer: We used the parameters from [Yeganeh & Wang, 2013] as is. We could have followed the approach in the [Yeganeh & Wang paper and trained the parameters on a large set of natural images to determine the parameters in the naturalness measure.

Question #4: What is the range of values for the proposed quality measure?

Answer: The minimum value is 0. The maximum value is slightly greater than 1 due to the beta^delta1 term where delta1 is 0.7088. The parameter beta is the exponent of the fit to a generalized Gaussian density of the mean subtracted contrast normalized pixels of the tonemapped image. For natural images, beta can be greater than 1; in fact, beta commonly takes values between 0.8 and 1.4.

COPYRIGHT NOTICE: All the documents on this server have been submitted by their authors to scholarly journals or conferences as indicated, for the purpose of non-commercial dissemination of scientific work. The manuscripts are put on-line to facilitate this purpose. These manuscripts are copyrighted by the authors or the journals in which they were published. You may copy a manuscript for scholarly, non-commercial purposes, such as research or instruction, provided that you agree to respect these copyrights.

Last Updated 10/01/16.