Contour Integration Models Predicting Human Behavior

Abstract

Contour integration is believed to be a fundamental process inobject recognition and image segmentation. However, its neuronalmechanisms are still not well understood. Psychophysical experimentsshowed that humans are remarkably efficient in integrating contourseven if these are jittered or partially occluded. Therefore thebrain requires a reliable algorithm for extracting contours fromstimuli. Several recent publications demonstrated that the brainoften uses optimal strategies to integrate sensory information.Hence in this thesis I want to tackle the question which contourintegration model describes human contour integration best.Mathematically, contour ensembles can be characterized by aconditional link probability density between oriented edge elements,termed an association field. This association field can be used togenerate contours or vice versa to extract a contour from astimulus. While in most neuronal network models all inputs to aneuron are summed up, in such a probabilistically motivated neuralnetwork for contour integration the afferent input due to the visualstimuli and the lateral input from horizontal network interactionsare multiplied.Long-range horizontal interactions in primary visual cortex linkorientation columns with similar preferred orientations and areoftenassumed to be the neuronal substrate for the association field. Experimental findings in monkeys suggest isotropic long-rangehorizontal connections, spreading symmetrically into all directionsfrom an orientation column. In contrast, probabilistic modelsrequire unidirectional lateral interactions, linking orientationcolumns in only one direction, in order to get optimal contourdetection performance.Using stimuli generated from given association fields, our numericalsimulations show that contour detection performance for both,probabilistic-multiplicative as well as additive models reacheshuman performance. Hence detection performance alone is insufficientto rule out either model class. However, psychophysical experimentswith humans reveal that contour detection errors are not maderandomly, but are highly correlated among different subjects. Thus amodel describing contour integration in the brain should not onlyexplain human contour detection performance, but should alsoreproduce these systematic errors made by humans. Comparison betweenmisdetections of humans and mispredictions of the models on atrial-by-trial basis was used to evaluate different model dynamicsand association fields. This suggests that unidirectionalmultiplicatively coupled horizontal interactions are required inorder to explain human behavior. Furthermore, cortical magnificationfactors have to be taken into account and a fixed association fieldgeometry for all stimuli is preferable instead of using for eachcontour the association field employed for the generation of thiscontour.