Surface Harmonic Mapping Based On Deep Learning

High-quality mapping between surfaces plays an important role in engineering and medical fields,and harmonic mapping can handle arbitrary topological surfaces.Harmonic maps have differential homeomorphism and are close to conformal maps.Harmonic mapping is not only simple and intuitive,but also supported by perfect theory.It has great universality and applicability.Mapping a three-dimensional face surface to a unit disk is a harmonic mapping based on a topological disk.Although there are mature algorithms to achieve harmony mapping,under the situation that in-depth learning has become a hot research topic,whether the neural network in-depth learning can learn harmony mapping has not been paid attention to until now.Understanding harmony mapping from the perspective of in-depth learning is the main purpose of this paper.This paper mainly uses the deep convolution neural network model to learn the harmonic mapping.The data set is the three-dimensional face grid data and the two-dimensional face disk data generated by the traditional harmonic mapping algorithm.The information of points is extracted from the three-dimensional face mesh and made into point cloud.The point cloud is used as the input data of the neural network model.The input point cloud is aligned with a feature.Then the feature of the three-dimensional face is extracted by the operation of multi-layer convolution neural network.The global feature and the local feature extracted are integrated as the feature of the three-dimensional face extraction.Then we use a 1*1 convolution core to perform multi-level convolution operation,and finally map the extracted face features onto a two-dimensional face disk.For the trained model,a new face on the unit disk will be generated by testing the new three-dimensional face,which shows that the harmonic mapping can be learned by using the neural network.At the same time,compared with the face disc generated by the traditional algorithm,the effect of reconciliation mapping using the traditional algorithm is very close.