Zdravko Bozakov, "Unsupervised Component Extraction for Design Optimization using Feature Analysis Methods", 2007.

Evolutionary algorithms offer an approach to alleviate some of the difficulties existing in traditional design optimization techniques. The choice of an appropriate design representation greatly influences the quality of achievable solutions. Representations which encode shapes using deformations applied to an initial design allow the number of optimization parameters to be decoupled from the complexity of the evaluated design. A novel technique called Free Style Deformation aims to further enhance this representation paradigm by encoding geometries using a combination of modules and deformations. The focus of this thesis is therefore to devise an unsupervised method for decomposing large collections of geometries into their salient parts. To achieve this goal techniques from the computer vision and object recognition domain are applied to three dimensional designs. Image decomposition algorithms, such as the Non-Negative Matrix Factorization, are evaluated. Limitations of the existing algorithms are identified and modification are developed which address the challenges specific to the 3D domain. Finally, a framework is proposed which is able to successfully extract features for use with free style deformation (FSD).