Related Experiment Videos
Contrast statistics for foveated visual systems: fixation selection by minimizing contrast entropy
Raghu Raj1, Wilson S Geisler, Robert A Frazor
1Department of Electrical and Computer Engineering and Center for Perceptual Systems, University of Texas at Austin, Texas 78712, USA.
Summary
The human visual system uses varying retinal resolution to process local contrast. An entropy minimization algorithm efficiently guides eye movements to maximize information gain about visual scenes.
Area of Science:
- Vision science
- Computational neuroscience
- Image processing
Background:
- The human visual system integrates a wide field of view with high-resolution foveal vision.
- Eye, head, and body movements are used to direct the fovea to salient visual areas.
- Efficient exploitation of varying retinal spatial resolution by central mechanisms is crucial.
Purpose of the Study:
- To understand the design requirements of central visual mechanisms.
- To analyze the impact of variable spatial resolution on local contrast in natural images.
- To explore how statistical properties of contrast might be utilized by perceptual systems.
Main Methods:
- Analysis of local contrast effects across retinal eccentricities in 300 calibrated natural images.
- Measurement of conditional probability distributions of local contrast at different blur levels.
- Derivation of an entropy minimization algorithm for selecting fixation points.
Main Results:
- Conditional probability distributions of "true" unblurred contrast are well-described by simple formulas.
- The entropy minimization algorithm optimally reduces contrast uncertainty.
- The algorithm effectively reduces mean squared error in image reconstruction.
Conclusions:
- Local contrast measurements alone can efficiently guide eye scan paths.
- Maximizing information gain about spatial structure is achievable through contrast-based fixation strategies.
- These findings offer insights into the neural mechanisms underlying visual attention and scene perception.