Computing local edge probability in natural scenes from a population of oriented simple cells

Area of Science:

• Neuroscience
• Computational Vision
• Image Processing

Background:

• The extraction of object contours is a fundamental visual cortex computation, often attributed to V1 simple cells.
• The standard simple cell model (oriented linear filter + divisive normalization) performs poorly as a local edge detector on natural images.
• Fine edge discrimination likely relies on information encoded by local simple cell populations.

Purpose of the Study:

• To investigate the decoding problem of edge probability using local simple cell populations.
• To develop a more effective local edge detector for natural images.
• To gain insights into neural computations underlying edge perception.

Main Methods:

• Utilized Bayes's rule to compute edge probability from a surrounding filter population.
• Developed an efficient method for ground-truth edge labeling.
• Selected maximally informative and minimally correlated filters, focusing on orthogonal regions.
• Employed a customized parametric model for individual filter likelihood functions.

Main Results:

• Developed a population-based edge detector with zero parameters.
• The detector calculates edge probability based on summed surrounding filter influences.
• Demonstrated significantly sharper tuning and effective capture of fine-scale edge structure in natural scenes compared to linear filters.
• Predicted nonmonotonic interactions between neighboring visual cortex cells.

Conclusions:

• Population-based coding provides a more effective mechanism for edge detection than individual simple cell models.
• The developed model accurately captures fine-scale edge information in natural images.
• Predicted neural interactions suggest complex, stimulus-dependent communication between V1 cells.