A Comparative Study of Clustering-Based Image Segmentation Algorithms for Enhanced Visual Understanding

Abstract

Segmentation of visual information through breaking down an image into several distinct parts, known as image segmentation, is an important step in gaining a better understanding of a piece of visual information. This paper provides a detailed comparative analysis of different types of clustering-based image segmentation algorithms like K-Means, Mean Shift and Spectral Clustering, and evaluates their performance based on the BSDS500 dataset. Both qualitative and quantitative methods for analysis were utilised through the usage of region-based as well as boundary-based metrics. The results of this study indicated that Spectral Clustering was better than all others by virtue of Boundary Adherence. The other two algorithms, K-Means and Mean Shift, give different levels of performance on either end of that measure, but also indicate some trade-off decisions when considering computational efficiency with respect to the three compared methods of unsupervised segmentation.