Expression Invariant3D Face Reconstruction

3D face reconstruction has drawn wide attention in the Computer Graphics and PatternRecognition fields since the last three decades. With the extensive application of computer technologyand pattern recognition,3D face reconstruction has continuously been applied to solve practicalproblems.3D Morphable Model (3DMM) was the mainstream in3D face reconstruction these dayswhich had a series advantages of high reconstruction accuracy and strong realistic. Besides, changesin the expression will also have influence on the face recognition. This thesis tries to study thecombination of the3D face reconstruction and expression factors to overcome such influence, and hasdone some work in expression invariant3D face reconstruction. The main work is as follows:First of all, we have proposed3D face correspondence based on improved Active Shape Model(ASM) algorithm. To create a standardized3D face database was the fundamental work for3DMMmethod. Via the idea of feature location method of classic ASM, we strengthen the optimizationfunction. An improved ASM-based3D faces correspondence algorithm was used to achieve thepixel-wise correspondence between3D faces, which improve the accuracy and efficiency of thestandardization process.Secondly, considering the poor robustness of the2D face recognition on the factors of expression,we have proposed Local Edge Distance alignment model to accomplish expression invariant3D facereconstruction from expression variant2D face images, which overcome the factors of expression in adegree. During the expression invariant3D face reconstruction,3D Sparse Morphable Model wasadopted to improve the efficiency of the3D reconstruction process.Thirdly, in order to overcome the3DMM constrained by the statistical model. Shadinginformation of2D image is fully used as prior knowledge to constrain each surface’s normal locatedin expression invariant region, which can effectively avoid the effect of the statistical model.Finally, experiment were conducted on standard2D face database, real-life image and networkimage to test our proposed algorithm, we also built platform based on VC++. Experimental resultscertify the effectiveness and superiority of our proposed methods.