A fast and robust ensemble evolving pixel cloud-based image segmentation approach

The existing cluster-based image segmentation algorithms have the burden of iterative computation caused by the change of cluster centers and are sensitive to noise. In this paper, we present a fast and robust ensemble evolving pixel cloud-based image segmentation approach. The concept of pixel clouds by clustering pixels of the same pattern around their focal pixels is proposed. The following attributes distinguish the proposed algorithm: (1) The pixel clouds are evolvable according to the global densities of the incoming pixels and the number of pixel clouds is automatically determined. (2) The focal pixels of pixel clouds are dynamically updated with the highest local densities by using the recursive density estimation, which avoids redundant distance calculations when a new pixel arrives. (3) A multiscale morphological gradient reconstruction operation is employed to merge or filter meaningless pixel clouds, especially in noisy images, which helps to adaptively polish neighboring pixel clouds and compact the pixel clouds. (4) An ensemble structure is introduced to fasten the image segmentation speed by splitting the whole image into multiple independent sub-images, in which the pixel clouds are independently formed and evolved. Comprehensive experiments on natural images, remote sensing images and medical images reveal that the proposed approach surpasses the state-of-the-art algorithms in both segmentation accuracy and computational efficiency. Even for the noisy images, the proposed approach demonstrates more robust performance.