Study Of A Single Photo Of 3d Face Reconstruction

The model of 3D human face is widely applied in security certificate, film andcartoon, medicine science and other fields. In recent years, many results have beenachieved in 3D face construction with particular face information. However, gettingso much information is expensive in most cases. More importantly, it is difficult tocollect enough information. For example, this may occur in the application of videoretrieval for a particular person. So, this thesis discusses the method of 3D faceconstruction based on a single frontal face image.First, we establish the face database of the frontal and lateral face informationfrom a lot of collecting frontal and lateral face images. According to the key featuresof human faces described in anthropometry and anatomy, we define the points anddistances which need to be marked and measured. Using neural networks, we train thefrontal and lateral data derived from these images, and with the weights we canestimate the depth values of the feature points. The fitting and simulation results showthe feasibility of our method.Second, for BP algorithm convergent slowly, we give some discussions about themethods of searching the best shape factor, learning rate and the best combination ofthem to accelerate the learning of neural networks. Combining these methods, we getanother new accelerate algorithm with vector epsilon extrapolations. Numericalexperiments show that the new algorithm can save CPU time effectively and improvethe accuracy.Third, we define the feather points of the generic face model based on themeasure points defined in the face database, and use the interpolation based on radialbasis function to obtain special person's 3D face mesh. For the sparse mesh, weprovide a modification of Loop subdivision to hold the original shape of the model.The face model is more realistic after texture mapping.At last, the face database and face modeling system based on image are designedand implemented by Visual C++.NET and OpenGL.