Image segmentation algorithms based on information compression and graph structures

In this paper we propose a multistage computational procedure for segmentation of images that can also be used for partitioning of large process data sets. In the first step the original "raw" data set (e.g. the set of pixels from a given image) is compressed by use of the neural-gas unsupervised learning algorithm into compressed information model (CIM) that contains small predefined number of neurons. In the second step a graph structure is generated by using all the neurons as nodes of the graph and a number of consistent arcs. Two kinds of consistent arcs are defined here, namely crisp and fuzzy arcs that lead to the respective crisp and fuzzy graph structures. The crisp graphs use the Euclidean distance between the nodes as "arc lengths". The fuzzy graphs use weighted arcs with different "arc strengths", computed by using the weighs of the respective adjacent neurons. The third step identifies the number of the strongly connected elements (called also "connected areas") in the generated graph structure from the previous step. This is done by using the well known depth-first graph algorithm. Then each connected area corresponds to a respective segment of the given data or image. The proposed computational scheme is demonstrated and explained by several test examples of images with discussion about its practical application in different fields.

I. Introduction

Segmentation of images [1], [2], [3] is a computational procedure of splitting the original image into several distinct sub-images. Each segment (sub-image) is further on used for visualizing and highlighting of some details of a special interest from the original (whole) image that are important for the proper decision in the medical or fault diagnosis. The similar problem arises also in the partitioning of large operational data sets that have to be separated into certain portions of subsets. Each subset contains “similar” data that correspond to a certain type of operation of the system. This is considered as important pre-processing procedure that helps all further processing steps, such as data similarity analysis, classification, decision making and fault diagnosis of machines and other industrial systems.