Abstract
Contour integration is believed to be an important step in human visual processing and object recognition, and has been shown to be performed very efficiently by the visual system.
We modeled contour integration using a Bayesian neural network. In this model the connection structure is given by an association field, a probability density describing the link probability between different edge elements. To evaluate the model we compared its predictions to human contour detection performance. Stimuli consisted of curvilinearly aligned contour elements which were drawn from an association field and a background of randomly oriented Gabor patches. Hence one would expect the model to perform most accurately when using the same connection structure for detecting the contour as was used to create the contour.
However, this would mean that the connection structure employed by the human brain had to change with the statistical properties of the contour while one might expect that the brain uses only one connection topology which is learned from visual experience and hence might be adapted to the statistical structure of natural images.
We analyzed several possible connection structures including the ideal association field for each single stimulus, connection topologies resembling the contextual interactions found in electrophysiology, and an association field extracted from natural images. In addition we looked at different association field symmetries and ranges.
As human contour detection performance was reached with several association fields, we compared correlations between human contour detection errors and errors by the model on a trial to trial basis. While the association field used for contour generation leads to highest performance, this correlation analysis suggests that the brain uses a single unidirectional association field linking edge elements in only one direction, rather than stimulus adapting association fields or bidirectional association fields which allow contours to abruptly change directions.
This work is supported by the German Academic Exchange Service (DAAD).