Low-level features predict perceived similarity for naturalistic images.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Visual Perception

Background:

• Mechanisms of human visual scene organization are not fully understood.
• Perception of statistical regularities is studied using simple stimuli and low-level image features.
• Computational models exist for explaining perceptions based on low-level image features.

Purpose of the Study:

• Investigate how observers use low-level information in naturalistic scene regions for association.
• Determine the extent to which basic image features facilitate region association.
• Explore the role of luminance and structure similarity in scene perception.

Main Methods:

• Two experiments using a two-alternative forced choice method.
• Participants judged if image patches belonged to the same scene.
• Analysis included generalized linear multilevel models with low-level feature similarity predictors.

Main Results:

• Participants successfully associated image regions above chance using unaltered patches.
• Low-level feature similarity metrics (luminance, structure correlations) predicted responses.
• Performance decreased with thresholded or edge-only image regions, but models still correlated.

Conclusions:

• Image region associations can be explained by low-level feature correlations.
• Basic visual features are strongly linked to judgments about complex visual stimuli.
• Low-level image properties play a crucial role in coherent scene representation.