Expression Robust3D Face Reconstruction And Recognition

3D face reconstruction and recognition is one of the hottest research topic in the field ofcomputer graphics and pattern recognition.3D face reconstruction technology can restore thephotographs of human faces, which can be applied to games and other fields, while face recognitionbased on3D reconstruction can completely get rid of the bottleneck of2D face recognition, which hasa widespread prospect. This thesis tries to study3D face reconstruction technique based on a singleimage, and recognition method based on reconstruction. The main work is as follows:Firstly, this thesis introduced a illumination robust face detection algorithm, and normalized the3D face database. We proposed a Triangulated Binary Pattern(TBP) as a face descriptor, whileintroduced a windows scanning strategy for detection in the picture. In pre-processing step, we dealwith3D face database by pose correction and position data normalization and projected all the3Dsamples into2D plane to get training samples and testing samples.Secondly, this thesis introduced a coarse-to-fine3D face sparse reconstruction method. Wetreated3D reconstruction as a coarse-to-fine process, reconstructed the temple model based on keypoints first and modified the temple model based on other points to get the final result. The methodcan improve the reconstruction precision and reality to a certain extent.Thirdly, this thesis introduced a sparse reconstruction method based on Compressed Sensing.Compressed Sensing theory is firstly used to estimate the similarity between testing and prototypeface samples, and a modified morphable model is then built on the selected prototype samples withlarger similarity. Secondly, the model is deformed based on facial salient points. Finally, combiningthe shape recovered by the modified model and the shape obtained by RBFs interpolation. Theproposed method can effectively improve the reconstruction speed and accuracy.Fourthly, this thesis introduced a reconstruction method based on3D Morphable Model andEdge Map. Edge Map is defined by points which represent main region on face, and the optimaldeforming coefficients is selected by matching between local edge line formed by the projection ofedge map and input feature points. This method can effectively reconstruct faces under any pose withhigh speed. Based on this method, a3D face reconstruction system was built using VC++.Lastly, this thesis also gave a preliminary discussion of reconstruction based face recognition,and studied the combination of2D face recognition and reconstruction based face recognition.Experimental results show the effectiveness of these recognition methods.