3D Face Modeling Based On Morphable Model

As the most important carrier for information exchange and emotion expressingin daily life, human face has been the focus of research and attention. Through face,we can not only obtain people's identity and race information, but also can get hiscurrent emotion state. With the development of computer vision science, the demandsfor human-computer interaction technology grow much higher. As the most importantpart of human-computer interaction technology, realistic3D face modeling certainlybecomes the focus of research. At present,3D face modeling technology has beenmade a considerable progress, and has been widely used in many areas, such ascomputer games, virtual interaction, medical technology, and public safe.Traditional3D face modeling methods have many deficiencies in modelingeffects and automatic modeling.3D face modeling method based on morphable modelis one of the best modeling at present. This method is constructed based on statisticallearning theory. Compared with other modeling methods, this method has betterperformance in modeling results and the automatic modeling.In this paper, theresearch works are carried out based on morphable model method and proposeappropriate solutions to corresponding flaws. The main works of this thesis areexamined in the following.1. Sample regularization based on combination model matchingRegularized3D face sample set is an important precondition for constructingmorphable modeling. At present, samples in different3D face database have bigdifference in topology, data structure, information coverage due to the differentbuilding purpose and sample deriving method. In order to implement linear operationon different samples, these initial samples should be regularized first to have themwith same topology, number of point and area, and can be expressed by a uniformvector. After in-depth study of face structure, a regularization method based oncombination model matching has been proposed in this paper. The combination modelis constructed based on regularized sample set which are derived by handwork.Samples which need to be regularized are matched with the combination model to gettheir regularization result. Since the combination model is a statistical model based onregularized samples, the regularized sample derived based on this model not onlymeets the geometric constraints, but also meets the reasonable constraints for face.2.3D face expansion based on Genetic Algorithm3D face samples are indispensable data resource for algorithm design, modeltraining and performance comparison about3D face research. Due to the restrictions of sample equipment and condition, samples in current3D face database are few andthe coverage range is relative small. To solve this problem, we propose a sampleexpansion method based on genetic algorithm. The basic idea is that the3D facesamples consist of limited fixed organs and a lot of new samples can be generated byregrouping different organs from different samples. Using this method can not onlygenerate a lot of3D face samples, but also increase the range covered by the sampleset.3.3D face modeling based on canonical correlation analysisThe assumption of the morphable model is that face space is a linear subspace.However, lots of studies show that human face is a nonlinear manifold embedded inhigh dimension space.3D face modeling method based on morphable model willcertainly ignore some details of face and affect the modeling results. In order tofurther improve the accuracy of modeling result, we proposed a nonlinear3D facemodeling method. The basic idea is to use the piecewise linear method to solve thecontradiction between the non-linear characteristics of face space and linearassumptions of the morphable model. This method is performed based on canonicalcorrelation analysis. The distance between2D face images and3D face samples arecalculated based on correlation distance and these correlative samples can be derivedin terms of this distance. When performing3D face modeling operation, a morphablemodel can be constructed based on these samples and the reconstructed sample can bederived by matching this model with input image. Since the model and input imageare highly correlated, a better reconstructed sample can be derived based on thismorphable model. Therefore,3D face modeling method based on canonicalcorrelation analysis can further improve modeling accuracy of morphable model.4. Modeling matching based on particle swarm optimizationThe process of3D face modeling based on morphable model is a matchingprocess. Due to the shape parameters, texture parameters, camera parameters andillumination parameters have been involved in matching process, the problem ofmorphable model matching is a large-scale and multi-parameters optimizationproblem. Particle Swarm Optimization is a swarm intelligence optimization algorithm,this algorithm is highly parallel and easy to implement. After in-depth study of thecharacteristics of modeling matching, a new multi-level modeling matching methodbased on particle swarm optimization to further improve the model matching speedand efficiency is proposed in this paper. Due to the particle swarm optimization is arandomly optimization algorithm with feedback, this method is robust to the matchingproblem of morphable model and sensitivity to the initial value. The proposed modelmatching algorithm used in this chapter can greatly improve the matching speed andmatching accuracy.