A deep convolutional neural network trained for lightness constancy is susceptible to lightness illusions

Area of Science:

• Computer Vision
• Computational Neuroscience
• Psychophysics

Background:

• Human visual perception accurately estimates surface reflectance under varying illumination.
• Developing image-computable models for human lightness perception has been challenging.
• Deep convolutional neural networks (CNNs) show promise in estimating surface reflectance.

Purpose of the Study:

• Evaluate a CNN for modeling human lightness perception.
• Test CNN susceptibility to classic lightness illusions.
• Compare CNN performance against human observers and traditional models.

Main Methods:

• Implemented and trained a CNN using supervised learning for surface reflectance estimation.
• Tested the CNN on stimuli designed to elicit lightness illusions (e.g., argyle, Koffka, simultaneous contrast).
• Conducted human psychophysical experiments and compared CNN outputs with classic lightness models.

Main Results:

• The CNN effectively removed lighting effects like shadows and shading.
• The CNN qualitatively predicted human perception of most tested lightness illusions.
• The CNN outperformed classic models in reflectance estimation and matching human lightness perception.

Conclusions:

• CNNs offer a powerful framework for modeling human lightness perception.
• The findings support a normative approach to understanding visual processing.
• Deep learning models show significant potential for advancing research in visual perception.