Gradual cluster elimination for color clustering in outdoor scenes

Abstract

Although color reduction is a common tool to simplify preliminary segmentation in color images, it does not show promising performance in case of outdoor images due to some complexities such as color variety, luminance effects, abundant texture details, and diversity of the objects in such images. In this paper, we propose a multi-stage color clustering procedure based on the well-known k-means algorithm, referred as GCE (Gradual Cluster Elimination). Here, a multi-resolution pyramid of the original image is exploited for deliberate expunging of texture details followed by a step-by-step clarification approach. Moreover, we gradually eliminate the apparent clusters while introducing new ones in each stage of the mentioned pyramid to consider all principal colors. The required similarity thresholds for color re-clustering are obtained automatically from multi-resolution images using their color distribution statistical characteristics. We have compared the performance of the GCE procedure and the standard k-means for color reduction on two outdoor datasets: University of Isfahan Data Set (UIDS), and Sowerby Image Dataset (SID) of British Aerospace. The experimental results have shown the advantages of the suggested procedure over traditional approaches, to name a few, improvement in the segmentation quality in terms of two well-known quantitative metrics: PRI and VoI, more accuracy and convergence speed, and simultaneously suppressing the over-segmentation and under-segmentation problems. The results of this research can be applied to many practical fields for segmentation in outdoor scenes, such as wearable computers, robotics, automatic vehicle control, and assisting the visually-impaired people. © 2013 ICIC International.