Interactive image segmentation of non-contiguous classes using particle competition and cooperation

Abstract

Semi-supervised learning methods employ both labeled and unlabeled data in their training process. Therefore, they are commonly applied to interactive image processing tasks, where a human specialist may label a few pixels from the image and the algorithm would automatically propagate them to the remaining pixels, classifying the entire image. The particle competition and cooperation model is a recently proposed graph-based model, which was developed to perform semi-supervised classification. It employs teams of particles walking in a undirected and unweighed graph in order to classify data items corresponding to graph nodes. Each team represents a class problem, they try to dominate the unlabeled nodes in their neighborhood, at the same time that they try to avoid invasion from other teams. In this paper, the particle competition and cooperation model is applied to the task of interactive image segmentation. Image pixels are converted to graph nodes. Nodes are connected if they represent pixels with visual similarities. Labeled pixels generate particles that propagate their labels to the unlabeled pixels. Computer simulations are performed on some real-world images to show the effectiveness of the proposed approach. Images are correctly segmented in regions of interest, including non-contiguous regions.