Predicting Eye Fixations Under Distortion Using Bayesian Observers

TL;DR

This paper explores how image distortions like compression artifacts influence human visual attention by applying Bayesian models to predict eye fixation patterns, revealing that such artifacts can distract viewers and affect perception.

Contribution

It introduces a Bayesian framework to predict eye fixations under image distortions, specifically applying MAP and ELM models to JPEG artifacts, which is a novel approach in this context.

Findings

Compression artifacts influence visual attention patterns.

Bayesian models can predict eye fixations on distorted images.

Distortions like blocking and ringing affect perceptual focus.

Abstract

Visual attention is very an essential factor that affects how human perceives visual signals. This report investigates how distortions in an image could distract human's visual attention using Bayesian visual search models, specifically, Maximum-a-posteriori (MAP) \cite{findlay1982global}\cite{eckstein2001quantifying} and Entropy Limit Minimization (ELM) \cite{najemnik2009simple}, which predict eye fixation movements based on a Bayesian probabilistic framework. Experiments on modified MAP and ELM models on JPEG-compressed images containing blocking or ringing artifacts were conducted and we observed that compression artifacts can affect visual attention. We hope this work sheds light on the interactions between visual attention and perceptual quality.