Contour Integration Models Predicting Human Behavior

Contour integration is believed to be a fundamental process in object recognition and image segmentation. However, its neuronal mechanisms are still not well understood. Psychophysical experiments showed that humans are remarkably efficient in integrating contours even if these are jittered or partially occluded. Therefore the brain requires a reliable algorithm for extracting contours from stimuli. Several recent publications demonstrated that the brain often uses optimal strategies to integrate sensory information. Hence in this thesis I want to tackle the question which contour integration model describes human contour integration best. Mathematically, contour ensembles can be characterized by a conditional link probability density between oriented edge elements, termed an association field. This association field can be used to generate contours or vice versa to extract a contour from a stimulus. While in most neuronal network models all inputs to a neuron are summed up, in such a probabilistically motivated neural network for contour integration the afferent input due to the visual stimuli and the lateral input from horizontal network interactions are multiplied. Long-range horizontal interactions in primary visual cortex link orientation columns with similar preferred orientations and are often assumed to be the neuronal substrate for the association field. Experimental findings in monkeys suggest isotropic long-range horizontal connections, spreading symmetrically into all directions from an orientation column. In contrast, probabilistic models require unidirectional lateral interactions, linking orientation columns in only one direction, in order to get optimal contour detection performance. Using stimuli generated from given association fields, our numerical simulations show that contour detection performance for both, probabilisticmultiplicative as well as additive models reaches human performance. Hence detection performance alone is insufficient to rule out either model class. However, psychophysical experiments with humans reveal that contour detection errors are not made randomly, but are highly correlated among different subjects. Thus a model describing contour integration in the brain should not only explain human contour detection performance, but should also reproduce these systematic errors made by humans. Comparison between misdetections of humans and mispredictions of the models on a trial-by-trial basis was used to evaluate different model dynamics and association fields. This suggests that unidirectional multiplicatively coupled horizontal interactions are required in order to explain human behavior. Furthermore, cortical magnification factors have to be taken into account and a fixed association field geometry for all stimuli is preferable instead of using for each contour the association field employed for the generation of this contour.