Learning-based algorithm selection for image segmentation

Segmentation of nontrivial images is one of the most important tasks in image processing. It is easy for human being, but extremely difficult for computers. With the purpose of finding optimal segmentation algorithm for every image through learning from human experience, this paper investigates the manual segmentation process and thus presents a performance prediction based algorithm selection model to bridge the knowledge gap between images and segmentation algorithms. Derived from that model, a framework of learning-based algorithm selection system is proposed to automatically segment all images in a large database. A simulation system is designed to select the optimal segmentation algorithm from four candidates for synthetic images. The system is tested on 9000 images by comparing with the manual algorithm selection. The best algorithms are selected for 85% of the cases. If we also regard the second best algorithm as acceptable, more than 97% of images can be properly segmented. The satisfied result demonstrated that this study has provided a promising approach to achieve automated image segmentation.