The Research And Implement Of Feature Transplant On Mesh Model Of Human Face

This paper is a research on the problem of feature transplant, existing transplanting algorithms, their defects and improvements.Face mesh models are usually used in 3D computer animation. Because of the complexity of geometry features on human face, it is usually used the feature transplanting algorithm to construct fine face models. Feature transplanting can reconstruct a face model with good topology from the geometry information of scanned faces, which will greatly promote the automation of compute modeling and animation producing.The input of feature transplant algorithms includes two mesh models, the fine model that structured by modelers is named template model and the scanned model with bad connectivity is named target model. Existing feature transplant methods are mainly following two threads which separately named Laplacian feature transplant and Iterative Closest Point (ICP) algorithm.Laplacian feature transplant is highly depended on mesh topology:template mesh and target mesh must be fine enough and can not contain data hole, or the feature transplant will fail. The cause of this problem is that Laplacian feature transplant algorithms need a parameterization step, but the reconstructed models always have too many data holes to be parameterized. This disadvantage badly reduced the practicability of Laplacian feature transplant method.Though feature transplanting using ICP algorithm is surface preserving and automatic hole filling, the algorithm is trapped into local extremum and causes large error at the area of feature points. This is because ICP algorithm is extremely sensitive to the correctness of nearest point search. If the search returns a bad position, ICP can not give out a good deformation. Combining the advantages of the above two methods and fixing their disadvantages, we present the Twin-step ICP algorithm based on Laplacian smooth. This algorithm is intense feature preserved and automatic hole filling.The main contributions of this paper are:1)Point out the problem of ICP algorithm at intense feature points and optimize the initial condition of it by extracting Laplacian feature texture. Our algorithm improves the result of ICP at intense feature points.2)Implement feature texture mapping by ICP algorithm to avoid parameterization step in Laplacian feature transplant. Reduce the dependence of mesh topology so that Laplacian feature transplant algorithm can run on the target mesh that unable to be parameterized.3)Enhance the adaptive capacity of mesh geometry and topology that our algorithm maintains maximum features when it is transplanted from a thick target model to a coarse template model. This improve makes our algorithm better to deal with practical problems...