Towards Fully Automatic Optimal Shape Modeling
Shape models and the automatic building of such models have proven over the last decades to be powerful tools in image segmentation and analysis. This thesis makes contributions to this field. The segmentation algorithm typically uses an objective function summing up contributions from each sample point. In this thesis this is replaced by the approximation of a surface integral which improves the segmentation results. Before building a model the shapes in the training set have to be aligned. This is normally done using Procrustes analysis. In the thesis an alignment method based on Minimum Decsription Length (MDL) is examined and the gradient of MDL is derived and used in the optimization. When trying to build optimal models by optimizing MDL there is a tendency for the parameterizations to put most of their weight on small parts of the shapes by doing a mutual reparameterization. In this thesis this problem is solved by replacing the standard scalar product with a formula that is invariant to mutual reparameterizations. This is shown to result in better models. To evaluate the quality of shape models, the standard measures have been generality, specificity and compactness. In this thesis, these measures are shown to have severe weaknesses. An alternative measure called Ground Truth Correspondence Measure is presented. This measure is shown to perform better. Typically, shape modeling assumes that the training set consists of images where the shape has been segmented as a curve or a surface as a preprocessing step. In this thesis a method is introduced that does not need preprocessed manually segmented data, automatically handles outliers/background and missing data, and still produces strong models. The algorithm makes all the decisions about what to include in the model and what to consider as background and about what points in the different images are to be considered to be corresponding. This results in patch-based shape and appearance models generated fully automatically.
Source Type:Doctoral Dissertation
Keywords:MATHEMATICS; Benchmarking; Parameterization Invariance; Interpretation; Segmentation; Alignment; Shape Modeling; MDL
Date of Publication:01/01/2008