An Information Model of Image Segmentation Algorithm Based on Redundancy Minimization

This paper is devoted to the study of the information-theoretical approach to the problem of image segmentation quality. We consider a system that includes a segmentation algorithm with a parameter on which the number of segments depends, and a procedure for selecting parameter value that provides the segmentation quality measure with minimum. As a quality measure, we use an information redundancy index. To study the properties of the system, a new simplified mathematical model is proposed. It is shown that for the proposed model, the redundancy measure has a minimum. The validity of the model is confirmed by a computational experiment. An experiment conducted on images from the Berkeley Segmentation Dataset showed that a segmented image corresponding to a minimum of the redundancy measure gives the highest information similarity to the ground truth images available in the BSDS500 database.