Humans make efficient use of natural image statistics when performing spatial interpolation.

Area of Science:

• Vision Science
• Computational Neuroscience
• Image Processing

Background:

• Visual systems adapt to natural image statistics for effective task performance.
• Understanding how the human nervous system utilizes these statistics is crucial for vision science.

Purpose of the Study:

• To investigate which statistical properties of natural images are learned and exploited by the human visual system.
• To compare human performance in estimating missing image pixels against computational models.

Main Methods:

• Human participants estimated intensities of missing pixels in natural images.
• Performance was compared to simple heuristics and optimal observers with varying statistical knowledge.

Main Results:

• Human estimation accuracy surpassed simple heuristics.
• Performance matched optimal observers using local contrast statistics, predicting error patterns.
• Humans did not match optimal observers using absolute intensity statistics.

Conclusions:

• The human visual system efficiently exploits local contrast statistics in natural images.
• Neural mechanisms likely focus on relative intensity rather than absolute intensity for pixel estimation.
• Context beyond local image patches offers minimal improvement for human estimation accuracy.