|Shape model deformable models structural models biometry content based image retrieval sketches|
This thesis proposes a new shape model called the Active Shape Structural Model (ASSM). The ASSM combines both statistical and structural a-priori knowledge about shape variation. The statistical a-priori knowledge models co-variations between two or more parts of the shape structure (e.g. co-deformation, joint articulation). The structural a-priori knowledge specifies which structural parts can be statistically related. The a-priori knowledge enables the ASSM to model a larger class of problems than structural or statistical models alone. Pure statistical models would have to use a complex distribution function to model shapes consisting of articulated parts like the human body. Pure structural models can decompose complex shapes into parts but cannot validate this decomposition against the allowed co-variations between those parts. Combining both structural and statistical a-priori knowledge results in interesting properties of ASSM such as multi-resolution of part variation depending on its context, completing missing structures and resolving conflicting interpretations using the shape's largest context. These properties of ASSM are demonstrated on two applications: Sketch recognition and ant recognition. Sketches demonstrate ASSM well because they have clearly defined structures that exhibit statistical variation for a single user, multiple users and depending on the co-variation with other parts in the sketch. The structural co-variation between multiple users was used in a new application called biometric recognition algorithm. In this case the structural relationship between drawing primitives are used as the secret information between the user and the biometric system. Experiments show that the ASSM can utilize well it's prior knowldge in recognizing, and correcting sketches as well as achieving good discrimination between users in biometric sketches. The ASSM was compared to a pure statistical representation and shown to be capable of effeciently reprsenting valid states of training data. After demonstrating the ASSM framework within the domain of online sketches, it was next used for ant segmentation. This is because ants both have articulated parts and different structural templates are needed to represent different ant types. Experiments show that co-variation between parts can be succesfully used for both template selection and finding effeciently the articulated parts. All these applications show that utilizating prior knowledge in the form of co- variation between shapes templates can lead to a better repersntation, reconstrcution, recognition, and correction of shapes.