Statistical shape modelling: automatic shape model building.
Doctoral thesis, UCL (University College London).
Statistical Shape Models (SSM) have wide applications in image segmentation, surface registration and morphometry. This thesis deals with an important issue in SSM, which is establishing correspondence between a set of shape surfaces on either 2D or 3D. Current methods involve either manual annotation of the data (current ‘gold standard’); or establishing correspondences by using segmentation or registration algorithms; or using an information technique, Minimum Description Length (MDL), as an objective function that measures the utility of a model (the state-of-the-art). This thesis presents in principle another framework for establishing correspondences completely automatically by treating it as a learning process. Shannon theory is used extensively to develop an objective function, which measures the performance of a model along each eigenvector direction, and a proper weighting is automatically calculated for each energy component. Correspondence finding can then be treated as optimizing the objective function. An efficient optimization method is also incorporated by deriving the gradient of the cost function. Experimental results on various data are presented on both 2D and 3D. In the end, a quantitative evaluation between the proposed algorithm and MDL shows that the proposed model has better Generalization Ability, Specificity and similar Compactness. It also shows a good potential ability to solve the so-called “Pile Up” problem that exists in MDL. In terms of application, I used the proposed algorithm to help build a facial contour classifier. First, correspondence points across facial contours are found automatically and classifiers are trained by using the correspondence points found by the MDL, proposed method and direct human observer. These classification schemes are then used to perform gender prediction on facial contours. The final conclusion for the experiments is that MEM found correspondence points built classification scheme conveys a relatively more accurate gender prediction result. Although, we have explored the potential of our proposed method to some extent, this is not the end of the research for this topic. The future work is also clearly stated which includes more validations on various 3D datasets; discrimination analysis between normal and abnormal subjects could be the direct application for the proposed algorithm, extension to model-building using appearance information, etc.
|Title:||Statistical shape modelling: automatic shape model building|
|Open access status:||An open access version is available from UCL Discovery|
|UCL classification:||UCL > School of BEAMS > Faculty of Engineering Science > Medical Physics and Bioengineering|
Archive Staff Only