A New Similarity Measure and Hierarchical Clustering Approach to Color Image Segmentation

Cluster analysis is an important task in data analysis and machine learning. Traditional clustering methods, such as partitioning and density-based approaches, have limitations in identifying natural clusters in datasets with elliptical and chained shapes. In this paper, we propose a novel hierarchical clustering algorithm for color image segmentation that addresses these limitations by quantifying the degree of overlap between clusters as a similarity measure for the merging process.

I. Introduction

Image segmentation consists of partitioning the image in homogeneous regions, where the pixels within these regions have coherent features such as gray level, color value, texture parameters, and depth to name a few. In some cases, image segmentation can have the purpose of extracting objects from the image based on the low-level physical properties that image pixels represent. A great deal of work has been done in image segmentation over the past decades and the proposed methods include region-based, contour-based, thresholding-based, watershed-based, and clustering-based methods.