A Hybrid Approach for Image Segmentation in the IoT Era

Spectral clustering is a class of graph theoretic procedure, which is popular for finding natural groupings. Over the last decade, it has become a widely adopted tool – utilized in solving image segmentation problems, via normalized cut (NCut) methodology. Few challenges faced by image segmentation based on spectral clustering include its inability of processing large images due to high computational cost and memory requirements and its sensitivity to irrelevant and noisy data. This chapter presents an unsupervised image segmentation technique using spectral clustering, aimed at salient object detection, followed by extraction. The presented technique addresses all of the aforementioned challenges by means of a weighted binary tree-based fast spectral clustering (WBTFSC). The algorithm integrates dimensionality reduction with spectral clustering by introducing an effective preprocessor, comprising two fundamental steps of color quantization and unique pixels selection. The experiments, performed on color images using the proposed algorithm, show improved performance in extracting objects of interest with high accuracy. We also test the algorithm on several noisy images; the obtained results reveal better performance in comparison to few existing techniques.