Segmentation of natural scenes: Clustering in colour space vs. spectral estimation and clustering of spectral data

Abstract

In this paper, two approaches are implemented and compared in order to determine which one offers the better segmentation quality for natural scenes: one is finding the best colour space for colour-based segmentation, and the other is segmenting by using the spectral data obtained by estimation from sensor responses. Eight colour spaces including perceptual and non-perceptual spaces are evaluated, and pseudo-inverse spectral estimation method is used for obtaining the estimated spectral reflectances from three simulated sensor responses. K-means and spectral clustering using Nyström approximation are used for segmenting an image after an adaptive quantisation step by mean-shift method is applied on the image. The segmentation results are evaluated by measuring the degree of matching with the segmentation benchmark using a similarity metric based on Jaccard index, and the segmentation benchmark is created by manual labelling. Results show that using estimated spectral data for colour image segmentation of natural scenes can achieve better or equally good results as the best colour space among the tested eight colour spaces.