Learning Graph Neural Network For Face Reconstruction

Face reconstruction is widely applied in face recognition,facial expression analysis,facial animation and other fields.The existing face analysis algorithms use regular grid to extract face local features,which not only extracts useful information,but also increases the influence of invalid information.It will weaken the description ability of local features.At the same time,face images often have a relatively fixed structure.For different 2D face images,the texture of the same position is similar.Considering these two reasons,we propose a new graph neural network for face reconstruction.This method transforms the input 2D face image or the 3D face model into their graph representation,so that,for all images and models,each node in the same position of the graph corresponds to the same information.In order to reduce the difficulty of solving face reconstruction problem,our graph neural network algorithm trains a shallow neural network for each node in the graph.Thus the original global face reconstruction problem is decomposed into a number of simpler local reconstruction problems.In order to extract local information of each node in the graph structure efficiency,we consider both the texture similarity geometric similarity in our graph neural network.Compared with the methods which only consider the geometric similarity,our method has greatly improved the local feature description ability.Finally,by using the depth first search algorithm,the algorithm realizes the parameter transmission among all sub networks,which speeds up the convergence speed of the network,and improves the accuracy of regression.This paper mainly applies the graph neural network to two kinds of face reconstruction problems:(1)single image based facial image super-resolution;(2)single image based 3D face reconstruction.In the experiment of face super-resolution,the proposed method outperforms all other algorithms in comparison with the latest face super-resolution methods,and achieves the best result.Meanwhile,in comparison with the natural image super-resolution algorithm,which based on deep network,the graph neural network achieves the same result as the deep network,and performs excellently on the high drop sampled images,which surpasses all other algorithms.For 3D face reconstruction algorithm,we firstly presents a new self-reinforced method for 2D facial landmark detection algorithm,which utilizes a characteristic number based global geometry discrepancy and texture feature based local appearance discrepancy.It can expand the number of training samples via a semi-supervised strategy and improve the accuracy of cascade regression based and deep leaning based facial landmark detection algorithm by 10%.We also proposed a conformal mapping based 3D facial landmark detection algorithm,which simplifies the 3D problem to the 2D problem.The algorithm obtain minimum mean error in all comparison methods.For the 3D face reconstruction experiments,graph neural network is tested under different illumination,pose and facial expression.The result shows our algorithm can well recover global and local geometric structure from only a single facial image.In the quantitative experiments,our method overpasses the latest Lm2 Vertex and 50-layers residual network and obtains the highest the reconstruction accuracy.