Proc. Asilomar Conference on Signals, Systems and Computers, Nov. 2-5, 2014, pp. 948-954, Pacific Grove, California USA.

Spatial Domain Synthetic 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 - Poster Presentation - ESPL Synthetic Image Database

Abstract

Natural Scene Statistics (NSS) has been applied to natural images obtained through optical cameras for automated visual quality assessment. Since NSS does not need a reference image for comparison, NSS has been used to assess user quality-of-experience, such as for streaming wireless image and video content acquired by cameras. In this paper, we take an important first step in using NSS to automate visual quality assessment of synthetic images found in video games and animated movies. In particular, we analyze NSS for synthetic images in the spatial domain using mean-subtracted-contrast-normalized (MSCN) pixels and their gradients. The primary contributions of this paper are

creation of a publicly available ESPL Synthetic Image database, containing 221 color images, mostly in high definition resolution of 1920 x 1080, and
analysis of the statistical distributions of the MSCN coefficients (and their gradients) for synthetic images, obtained from the image intensities.

We find that similar to the case for natural images, the distributions of the MSCN pixels for synthetic images can be modeled closely by Generalized Gaussian and Symmetric Alphha Stable distributions, with slightly different shape and scale parameters.

Question and Answer Session

The following is a reconstruction by the first author of the questions and answers during the poster presentation:

Question #1: How do you define artifact in a synthetic image? The fact that it is an artificial image might itself be an artifact according to some viewers.
Answer #1: We are concerned more about the visual quality acceptability. Unrealistic content does not necessarily mean a visually displeasing image. We are trying to judge image quality without taking into account the content.
Question #2: Isn't it natural that MSCN coefficient distribution of photorealistic graphics images will be same as natural images?
Answer #2: Yes, it is natural. But till date there has been no analysis on a large database to prove the amount of model match. Also, we can have two images having identical MSCN coefficient distribution, but one of them is a natural image and the other one is a highly non-realistic graphics image. We are trying to show that the distributions are valid irrespective of the content, be it natural or artificial, as long as the images are devoid of distortions.
Question #3: Is your database representative enough?
Answer #3: We have tried to maintain a good variation in content by looking at images from a large number of animation movies and video games. As in the case of any supervised learning approach, it is impractical to take into account every graphics image ever rendered, but we have tried to maintain a variety in illumination conditions, natural and man made objects, single (or multiple) salient objects, clear (or not) foreground-background separation, focus at different depths etc.
Question #4: How do the statistics vary with respect to change of size of the local window?
Answer #4: We tried 3x3, 5x5, 7x7, and 9x9 local windows. Statistics were pretty much the same; i.e., they led to nearly identical distribution of the MSCN coefficients. We reported the results with 7x7 windows in the paper.

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 01/31/16.