An intuitive way of color image segmentation is through clustering in which each pixel in an image is treated as a data point in the feature space. A feature space is effective if it can provide high distinguishability among objects in images. Typically, in the preprocessing phase, various modalities or feature spaces are considered, such as color, texture, intensity, and spatial information. Feature selection or reduction can also be understood as transforming the original feature space into a more distinguishable space (or subspaces) for distinguishing different content in an image. Most clustering-based image segmentation algorithms work in the full feature space while considering the tradeoff between efficiency and effectiveness. The authors’ observation indicates that often time objects in images can be simply detected by applying clustering algorithms in subspaces. In this paper, they propose an image segmentation framework, named Hill-Climbing based Projective Clustering (HCPC), which utilizes EPCH (an efficient projective clustering technique by histogram construction) as the core framework and Hill-Climbing K-means (HC) for dense region detection, and thereby being able to distinguish image contents within subspaces of a given feature space. Moreover, a new feature space, named HSVrVgVb, is also explored which is derived from Hue, Saturation, and Value (HSV) color space. The scalability of the proposed algorithm is linear to the dimensionality of the feature space, and our segmentation results outperform that of HC and other projective clustering-based algorithms.
Color Image Segmentation: Kernel Do the Feature Space