An Information Model for Digital Image Segmentation

This paper investigates an iterative information-theoretical method for segmentation of digital images. A system that includes a segmentation algorithm with a parameter that determines the number of image segments and a procedure for setting the value of this parameter that minimizes the information redundancy measure is considered. A new simplified mathematical model is proposed to analyze the properties of this system. It is shown that there exists a minimum of the redundancy measure for the proposed model. The adequacy of the model is confirmed experimentally. The computational experiment carried out on images from the Berkeley Segmentation Dataset (BSDS500) shows that a segmented image corresponding to the minimum redundancy measure has the highest informational similarity to ground truth segmentations available in BSDS500. We compared the image segmentation results provided by the EDISON system using the minimum information redundancy criterion and entropy criterion. The advantage of the minimum redundancy criterion is demonstrated.