Image Segmentation/Registration: a Variational Framework for 2-D and 3-D Applications

Abstract

Segmentation lands itself in the middle level of a computer vision system that it extracts boundary features in images. An accurate extraction of features will lead to the success of later higher level processes such as classification and recognition. Registration, on the other hand, is intimately intertwined with segmentation. An accurate allocation of edges, which are used as feature points, may increase the performance of registration. Whenever a single modality is not sufficient for segmentation and the resort to multi-spectral images is needed, perfect alignment of these multi-spectral images will also ease the segmentation task.

In this thesis we propose segmentation and registration methods corresponding to different real applications. In the first biomedical application we propose a constrained Mumford-Shah type energy functional incorporated with an information-theoretic view and tuning weights. This model characterizes higher-order statistical properties of data and give a probabilistic flavor to our segmentation. It successfully segmented T1-Maps and T1-weighted images in both 2-D and 3-D. Validation by experts manual segmentations also shows our method outperform most other techniques. Moreover we propose a joint radiofrequency (RF) -inhomogeneity calibration method to correct the non-uniformity of RF filed for accurate T1-Map generation.

We propose a multi-phase joint segmentation and registration technique (MPJSR) for mid-range layered imageries in the second application. Our method in particular may bring the objects of interest in a pair of layered images into perfect alignment and delineate the boundaries simultaneously. Furthermore, based on our technique, we tackle the tracking problem for layered videos. By calculating a constrained optical flow between consecutive frames, a prediction for the contour location may be made in the next frame to expedite and increase the segmentation performance.

In the third application we took a step further after segmentation and registration. The study of canonical views of multiple range images launches from the assumption that the segmentation and registration between multiple range images are done. We propose using two methods: minimum description length (MDL) and compressive sensing approaches, to determine the canonical views for a 3-D object.