engleski

A neural model of the interaction between object recognition and figure-ground organization

Peterson and her colleagues provide intriguing empirical evidence that object recognition influences figure-ground organization. Familiarity effects contradict classical assumption that figure-ground is a prerequisite for object recognition. Vecera and O’ Reilly (1998) proposed a computational model based on interactive activation, which is able to assign figural status to familiar contour, but they failed to provide coherent account of the interaction between familiarity and bottom-up cues for figural assignment such as relative size, symmetry or occlusion (Peterson, 1999). Also, their model was based on a biophysically unrealistic computational mechanism (simulated annealing), which often produces spurious states that do not correspond with none of the two possible perceptual interpretations. We proposed a new neural network model in order to explain how bottom-up and top-down influences are combined into a unified perception of figure and background. The model is based on the recent findings about interaction between the ventral and the dorsal visual processing stream. The dorsal stream computes saliency based on boundary signals provided by the simple and the complex cortical cells. Output from the dorsal stream is projected to the surface network which serves as a blackboard on which the surface representation is formed. The surface network is a recurrent network which segregates different surfaces by assigning different firing rate to them. The figure is labeled by the maximal firing rate. Previous work showed that the proposed model correctly assigns figural status to the surface with a smaller size, a greater contrast, convexity, surroundedness, horizontal-vertical orientation and higher spatial frequency content. In the present work we showed that object recognition effect on figure-ground assignment might arise from increased saliency of familiar contour in the dorsal stream. Object recognition is implemented by spatial attentional gating and template matching of the part of the contour with memorized patterns. When there is a match between memorized template and bottom-up pattern, activity in the dorsal stream increased at the locations corresponding with the contour. On the other hand, when there is no match, activity in the dorsal stream reflects only bottom-up influences. We performed a set of computer simulations showing that model correctly assign figural status to familiar object when bottom-up cues are ambiguous. On the other hand, when there is a strong bottom-up cue in the input, familiarity competes with the bottom-up cue and final decision on figural status depends on the relative strengths of bottom-up and top-down signals. Furthermore, the model explains why object recognition is effective on luminance-defined stereo-disparity but not on random dot stereograms. Important advantage of the present model is that figural assignment is always consistent and there are no spurious states as in the model of Vecera and O’ Reilly (1998).

Figure-ground Organization; Object Recognition; Neural Network Model

nije evidentirano