SS4: Statistical Modeling and Analysis
of Subjective and Objective Image Quality Assessments


Statistical modeling in image quality assessment is the use of mathematical models and statistical assumptions to represent a faithful abstraction of data from responses in a subjective quality assessment experiment. Such a statistical model is a collection of probability distributions on the set of all possible outcomes of the experiment. It is useful to generate sample data, make predictions, and can be applied for learning of objective quality metrics and judging their performance.

The method of pairwise comparison has been used for image quality assessment together with a statistical model, introduced by the psychometrician L. L. Thurstone in 1927 and called the law of comparative judgment, to order and position images on an interval scale of perceived quality. However, most popular are direct single stimulus quality ratings on a visual analog scale or with the 5-level absolute category scale, yielding mean-opinion scores (MOS). Recently, also for these direct rating approaches, statistical models were introduced to take into account subject biases and rating inconsistencies, and already adopted in the ITU standards P.913 and BT.500.

Statistical models for image quality assessment may provide several useful features:

  • Computing a model: To compute model parameters, Bayesian and maximum likelihood estimation can be used.
  • Applying a model: Under a model, each response in a quality assessment experiment for a direct rating, a pair or triplet comparison has a defined likelihood or probability.
  • Checking a model: Given a model, derived from a dataset, one can analyze the null-hypothesis that the data is a sample of the model, including derivation of a p-value.
  • Selecting a model: Given a design space of different model-types, established statistical methods can be used to select the most appropriate one for a given dataset.

This special session aims at discussing the further development and application of statistical models for subjective and objective image quality assessment.


We invite researchers and professionals to submit long and short papers covering different topics related to statistical models in QoE, including:

  • comparison of distribution types for perceptual quality
  • modeling biases beyond individual subjects
  • statistical models for diverse kinds of quality rating protocols (like quality ruler, Samviq, …)
  • statistical models for assessment of JND
  • characterization of ACR-type distributions by parametrized models (like the GSD)
  • merging of databases generated by different quality assessment types
  • performance and quality assessment of IQA datasets
  • adaptive sampling techniques for large-scale QoE subjective tests
  • deep learning of quality metrics from datasets with statistical models for quality
  • comparison of quality metrics based on benchmarks with statistical models
  • efficient methods for computation of large-scale models
  • software frameworks for simulation studies
  • multidimensional image quality models
  • statistical approaches for QoE datasets denoising