Development of a clustering algorithm for universal color image segmentation

Abstract

Image segmentation is an important stage of many real-world image applications in the domain of computer vision as a core method for understanding and analyzing digital images. It is aimed to segregate the most salient objects in an image by clustering homogenous regions based on the characteristics of image pixels. Segmentation of salient objects is a complex process because of the existence of numerous inherent characteristics of images that can impede the performance of the process. Due to these diverse image characteristics, a model that is suitable for one category of images is essentially inappropriate for other image categories, which makes image segmentation an open problem. Myriads of classical segmentation algorithms have been developed over the years, yet generalization and universal optimum performance are far from ideal levels. Clustering algorithms have been developed in recent times for the effective segmentation of images. However, the performance of the majority of the existing clustering-based segmentation algorithms substantially relies on the selection of an optimal number of initial clusters. Incorrect cluster count selection may result in uneven highlighting of the target object and be susceptible to under- or oversegmentation of images. This opens an avenue to fully discover a universal clustering algorithm for image segmentation that would be appropriate for manifold classes of images. In this study, the color histogram clustering algorithm has been proposed to automatically determine a suitable number of clusters that indicates homogenous regions in an image. The aim was to segment the most salient object from its surrounding regions using color histogram clustering that characterizes homogenous regions based on primitive features. The segmentation algorithm starts with histogram clustering-based on the quantized RGB color image to automatically identify the clusters that correspond to the homogenous regions in the image. The perceptual homogeneity of the input RGB color image is achieved by the transformation to L*a*b* color model based on four primitive features. The primitive features are color contrast, contrast ratio, spatial feature, and center prior that are extracted to compute the descriptor of each cluster. The cluster level saliency score is then computed as a function of the four primitive features extracted from the color image. The cluster level saliency is used to compute the final saliency score of each pixel to highlight the target object. The color histogram clustering method of this study combines the Otsu thresholding algorithm with the saliency map to represent the segmented image in a silhouette format. Morphological operations are finally performed to remove the undesired artifacts that may be present at the segmentation stage. Hence, this present study has introduced a novel, simple, robust, and computationally efficient color histogram clustering algorithm that agglutinates color contrast, contrast ratio, spatial feature, and central prior for efficiently segmenting the target objects in diverse image categories. The performance of the proposed algorithm was evaluated using the widely used metrics of precision, recall, F-measure, mean absolute error, and overlap ratio on six different categories of images selected from five benchmarked corpora of MSRA10K, ASD, SED2, ImgSal, and DUT OMRON. Moreover, 1000 images from ECSSD, 4447 images from HKU-IS, and 1500 images from COCO datasets were selected to validate the performance of the algorithm on more complex natural datasets. Experimental results have indicated that the proposed algorithm outperformed 30 bottom-up non-deep learning and seven top-down deep learning salient object detection algorithms. The performance of the proposed algorithm was further evaluated on four medical image datasets and the effects of image preprocessing were comprehensively investigated. The performance of the proposed image segmentation algorithm was analyzed in terms of accuracy, sensitivity, specificity, and dice similarity on 10015 images from HAM10000, 2594 images from ISIC2018 dataset, and 200 images from the PH2 dataset against six supervised and six unsupervised benchmark segmentation algorithms. The performance of the proposed algorithm was further validated on the segmentation of 1145 leukocyte nuclei images from the Raabin-WBC dataset in terms of accuracy, sensitivity, specificity, Dice similarity, and Jaccard index. In total, 22307 images with a variety of properties were used to test the performance of the proposed algorithm. In addition, the effects of image preprocessing on the performance of the proposed algorithm were further investigated in this study. The statistical results obtained have shown that the proposed algorithm is free from image preprocessing, and demonstrated its application on a wide class of images without any bounding to the heterogeneous characteristics of the input images. The novelty of the work reported in this thesis has demonstrated that the proposed algorithm is superior to the investigated supervised deep learning and prominent unsupervised segmentation algorithms in terms of quantitative results and visual effects.