Using a Stable Mass Spring Model as Trainable Dynamic Deformable Template

In this thesis we propose a Dynamic Deformable Template (DDT) using a Stable Mass Spring Model (SMSM). In our work, we combine the idea of deformable templates and physically based models in the proposed DDTs. A DDT is a template whose deformation is governed by its physical properties. We use two different strategies for generating a DDT. Handcrafted DDTs are those whose structural descriptions are specific to the shape of the object class and vary among the classes. Rectangular grid DDTs are structurally equal for all the shape classes and capable of sampling space around the object as well. Object specific shape information is introduced in to the templates as model forces. We also experimented with two different matching quality measures: global and local measures. The first one measures the quality of global fitting of the templates and the second one measures the quality of fitting of the templates at local level. The first measure can be applied to both types of templates. The second measure can only be applied to the rectangular grid DDTs as it requires comparison between corresponding parts of templates. We used handwritten digits from the MNIST database to investigate the capabilities of the proposed DDTs and the matching measures. We compared performances of trained and untrained DDTs. We investigated whether the deformation mechanism appropriately represents variation within a morphology class of a digit and whether training improves the performance of DDTs. The comparison between global and local measures for the template match is motivated by the question whether a useful measure for the template match exists. The findings of this thesis are that i) the idea of combining PBMs and deformable templates works successfully, ii) the general description of templates has more potential than the object specific description, iii) training does improve the performance iv) local measures are proved to be better template matching criteria than the global measure.