Neural network model for generating subjective contours

The human visual system has the ability to recognize a virtual line even for a discrete object such as a dot pattern. Even if there is no difference in color or brightness, a clear subjective contour can still be perceived. This phenomenon can be formulated as an interpolation problem. This paper proposes a hierarchical visual information-processing model based on physiological data. The model then is applied to the problem of interpolation, e.g., the perception of a virtual line and a subjective contour. To form the subjective contour, the actual outline of the visual object, the points comprising the subjective contour and the points that move to form a subjective contour must be determined. This is achieved by parallel processing in the proposed model which is composed of complex cells and hyper-complex cells. The complex cells receive the outputs of two kinds of simple cells (S and L), and executes contour detection and interpolation of a dot pattern by global processing. The hypercomplex cell detects the end point and the point of large curvature through which the subjective contour is supposed to pass. All these processes are executed by parallel processing. When the results of processing are ready, the points comprising the subjective contour are determined by weighting. The model in this study is constructed to detect the subjective contour as well as the actual edge by assembling the information from these hierarchical processing.