Research On Optimization Method Of 3D Synthetic Model Based On Deep Network Model

With the rapid development of the display technologies,3D shape data is becoming another important and popular media type,And it will be an important tool to change people's production and life.However,most of the existing 3D shape acquisition methods are either expensive or expertise-dependent.These problems have caused great difficulties to the popularization of 3D model data.The method of 3D model synthesis based on deep neural network model makes it possible to obtain a large number of 3D model data.But the data synthesized by deep neural network need to solve two difficulties: 1)the 3D model has geometric characteristics,the model matrix has multilayer structure,and it is difficult to fit the deep neural network model;2)the results of deep neural network model synthesis can predict that there must be a lot of noise,so a good surface optimization scheme is needed.Therefore,In order to make the threedimensional model synthesized by the deep network model have a good effect,it is urgent to solve the two major difficulties of the 3D model synthesis method,which seriously affect the effect of the 3D model based on the deep network model,and are also the focus of the development of the 3D model synthesis method.In order to address these issues,two links are designed in this paper.the first one is the automatic synthesis of 3D point cloud data,which is mainly obtained by the deep neural network model;the second one is the optimization of the synthesis results.In this paper,we have done the following work:In this paper,we present an image representation for the 3D point cloud model to bridge the gap between the feature-lacking 3D shapes and the powerful deep neural network learning tools.To achieve this,with the training set,we first extract the radial curves for each 3D point cloud surface,and reform the curves into an RGB image matrix,which enable to apply the classical Generative Adversarial Network model for the image synthesis.Then,we propose a refining process to transform the GAN output image into a natural 3D synthetic model.Due to the low-resolution of the coarse 3D point cloud models by using the synthetic deep neural networks,we propose a gradient based optimization approach for the coarse models.We further propose a gradient based optimization method to interpolate the 3D point cloud models for high quality 3D surface models.Our experimental results demonstrate the capability of our method which can generate the 3D faces with different expressions.And Our experimental results also show the effectiveness of our smoothing method.