3D Face Reconstruction Based On NRSFM Algorithm

3D face modeling is a hot research topic in the field of computer vision. With the development of science technology, the 3D face modeling is more diverse. The face information acquisition of most 3D face reconstruction algorithm methods need shooting faces in some fixed postures, hence it limits their application range. Structure from motion algorithm can recover 3D face information from face video random shooting without 3D face priori information. In recent years, this algorithm has been improved and can calculate 3D face accurately. On this basis, further research of this algorithm has been studied for structure from motion algorithm and overcomes the shortcomings in 3D face modeling. Specific improved parts are shown as follows:(1)This thesis designs a new sieve method to solve the problem that structure from motion algorithm is sensitive to feature points. Using direction and scale selectivity of Gabor filters extracts texture features of image blocks where feature points is located, to evaluate the reliability of feature points and remove the wrong matching points.(2)Occlusion, the wrong matching points, and other factors will cause missing values on 2D feature point sequences to reduce the accuracy of 3D information. To solve this problem, this thesis designs an adaptive column space fitting algorithm to estimate the missing values. Especially, it is better at estimating these missing values in small sample image sequence.(3)Non-rigid structure from motion with rotation invariant kernels can recover accurately 3D structure, but parameter variation will affect performance of the algorithm. This thesis designs an improved rotation invariant kernel function to measure the similarity among 2D feature point sequences to reduce the influence of parameter variation.All proposed algorithms in this thesis are tested on datasets to validate the effectiveness of the aforementioned methods. Finally, these methods are applied in 3D face modeling system to verify validity and feasibility of the system.