Robust segmentation and object classification in natural and medical images.

Ph.D. Thesis Lin Yang


Abstract

Image segmentation and object classification are two fundamental tasks in computer vision. In this thesis, a novel segmentation algorithm based on deformable model and robust estimation is introduced to produce reliable segmentation results. The algorithm is extended to handle touching object and partially occluded image segmentation. A multiple class segmentation algorithm is described to achieve multi-class "object cut". The accurate results are achieved using the appearance and bag of keypoints models integrated over mean-shift patches. An affine invariant descriptor is proposed to model the spatial configuration of the keypoints. Besides working with 2D image segmentation problem, a robust, fast and accurate segmentation algorithm is illustrated for processing 4D volumetric data. One-step forward prediction is applied to generate the motion prior based on motion modes learning. Two collaborative trackers are introduced to achieve both temporal consistency and failure recovery. Multi-class classification algorithms using a gentle boosting is used to classify three types of breast cancer. The algorithm is Grid-enabled and launched on the IBM World Community Grid. We will introduce a fast and robust image registration algorithm for both 2D and 3D images. The algorithm starts from an automatic detection of the landmarks followed by a coarse to fine estimation of the nonlinear mapping. The parallelization of the algorithm on the IBM Cell Broadband Engine (IBM Cell/B.E.) will also be explained in details.

The thesis contains 106 pages.


Return to Theses