A model of the *perception of partially occluded and fragmented figures
A computer model was built to implement and test a new theory of the perception of partially occluded figures. A key component of this contour-based theory is that an occluded figure may be perceived if its bounding contours can be separated from those bounding the occluding figure(s). Further, any contour property can potentially be used for such classification. This implies that the human visual system uses global processing in the perception of partially occluded figures. The model uses an exponential pyramid architecture to achieve such global processing-based classification. First, the pyramid analyzes all image contours with respect to an arbitrary number of contour properties. Here, the properties of contour orientation and contour size were implemented and tested. For a given property, the pyramid then tries to compute a cutoff value for dividing image contours into those which belong to the occluded figure and those which do not. Next, the pyramid uses this cutoff value to classify image contours and remove those not belonging to the occluded figure. Thus, the output of the pyramid is that subset of contours judged to be “intrinsic” to the occluded figure. If the model is psychologically plausible, then the amount of intrinsic contour recovered by the model from an image containing occlusion should correlate with human ability to correctly perceive the partially occluded figure in the image. Two sets of simulations and psychophysics, one testing the model with the property of contour orientation and the other with contour size, showed a strong relationship between model and human performance. This led to two new psychophysical experiments, which further tested aspects of the model. First, since the model uses global processing, subjects were tested to determine how much of the image is actually analyzed. In other words, how large is ‘global’? Results showed that the answer to this question varies between subjects. Second, the model determines the cutoff value for classification from the image itself. Can human observers do the same or do they use other cues such as expectation? Results showed that humans can indeed determine a cutoff from the image. ^
Major Professor: Zygmunt Pizlo, Purdue University.