3D Face Modeling From A Single Photo Based On Neural Network

Technology of 3D human face modeling based on graphic photo has been a hot spot, and had many applications on the fields of film and television animation, game design, facial expression recognition, etc. Currently, there are many methods about modeling of 3D face, but the improvement is needed on the effect and efficiency, etc.This paper presents a new method of reconstruction of 3D human face from a single photo. The core algorithm is that neural network is adopted to be an approximation of mapping function, which input is the coordinates of facial feature points plane, and the output is the depth coordinates It makes the depth of information as close as possible real effect.First, BP neural network is adopted as a mapping model. The connection weights are trained where the genetic algorithm aided, while learning methods of the NN is optimized. The tests showed BP had better effectively.Second, RBF neural network is adopted as a mapping model. Regularization network is used to improve the over-fitting problem, and better results are achieved. Comparison of performance between two neural networks by experiments, RBF network training speed is faster, better robustness, but the speed of modeling of genetic neural network is faster.Third, the sample database of face photo is established. This research defines standard model feature points and shape cell, proposes a transformation algorithm based on local shape of specific face model. For two kinds of neural networks, deep fitting algorithm is proposed respectively. For the bilinear surface interpolation problem, NURBS interpolation of surface reconstruction is proposed so as to complete surface reconstruction with less computation. Finally, based on NURBS interpolation properties, cylindrical texture mapping model is applied to complete texture mappingFinally, Experiment is performed to prove the validation of algorithm, and effect of realistic 3-D face is realized.