Design And Implementation Of Image-based 3D Head Portrait Generation System

Since the invention of the camera in the fifteenth century,people have been eager to use their favorite photos as decorative or gift items.With the development of technology,the form of photos has also been changing: from the initial black and white photos,to color photos,to the popular electronic photo frames in the last decade or so,photos have become more and more expressive,but always stay on a 2D plane.With the development of computer vision technology,computer 3D modeling technology has gradually become a popular research topic in the field of computer vision and computer graphics with its more intuitive visual experience and richer expression ability,so”3D photo studio” has also become a more popular application area.Especially,as 3D printing technology becomes cheap and popular,it makes it possible to make 3D solid form photos.Among them,3D head reconstruction is of great research significance and practical value.However,it is more difficult to obtain 3D head portrait directly from 2D image reconstruction,mainly because the current head datasets are not highly accurate,and therefore cannot accurately represent the texture and geometric changes in the face region,which is an important criterion used to distinguish different 3D head portraits.And unlike the head dataset,many researchers have proposed many 3D face datasets with higher accuracy in recent years.Therefore,in this paper,we first reconstruct the corresponding 3D face model using the image-based face reconstruction algorithm,and then fit the 3D face model into the existing 3D head template to generate the final 3D head portrait.The work accomplished in this paper mainly contains the following aspects:(1)To address the problems that traditional multi-image 3D face reconstruction algorithms cannot fully utilize the information between different view images and the tedious optimization steps,this paper proposes a multi-image face reconstruction algorithm based on deep learning.The method uses a learnable fusion module so that the network can make full use of the information between different view images,and designs other loss functions such as perceptual loss and adversarial loss to help the network learn better and obtain a robust reconstruction network.Finally,the multi-image reconstruction algorithm designed in this paper improves the defects of the traditional multi-image face reconstruction algorithm in terms of reconstruction framework and optimization method,surpasses the existing multi-image face reconstruction algorithm based on deep learning,improves the reconstruction accuracy of 3D face model,and also provides a basis for the performance improvement of the subsequent single-image reconstruction algorithm.(2)In response to the problem that existing single-image reconstruction algorithms lack multi-view images to provide reliable geometric constraints and therefore lead to ambiguity in the reconstruction,this paper proposes a new training-testing framework for single-image reconstruction networks.First,the reconstruction network chosen in this paper remains a single-image reconstruction network.In the training phase,unlike the existing single-image reconstruction networks,this paper inputs pairs of images into the single-image reconstruction network,and through the multi-view geometric consistency and generative adversarial modules,the single-image reconstruction network can learn the complementary information between images of different views and eliminate the disparity problem caused by the reconstruction of images of different views.This multi-view image training method can effectively improve the learning effect of the single-image reconstruction network,so that in the testing stage,although the input is still a single image,a 3D face model with high accuracy can be reconstructed.The single-image reconstruction network training-testing framework proposed in this chapter can significantly improve the accuracy of the existing single-image reconstruction algorithm based on deep learning,so that its reconstruction accuracy approaches that of the multi-image reconstruction algorithm based on deep learning.(3)In this paper,we design and implement an image-based 3D head portrait generation system.The generation system designed in this paper contains multi-image face reconstruction,single-image face reconstruction and 3D head portrait generation modules.The system can accept user input face images and eventually allow users to obtain not only 3D face models but also corresponding 3D head portrait models,which can be applied to the field of 3D printing.The final 3D head portrait generation system designed in this paper makes the image-based face reconstruction algorithm proposed in this paper be applied in practice and expands the application prospects of the algorithm for face reconstruction.