Research On Key Technology Of Unsupervised Single-view Face 3D Reconstruction

Single-view 3D face reconstruction is a research hotspot in the field of computer vision,aiming at inferring the 3D geometry of the face from a single face image and realizing the stereoscopic modeling of the face.This technology has a wide range of application prospects in computer graphics,virtual reality,face recognition,human-computer interaction and other fields.The key lies in how to obtain true and accurate 3D face information through a single image.Although current methods using a priori model reconstruction perform well in singleview 3D face reconstruction tasks,in unsupervised non-a priori methods,the predicted textures suffer from appearance blurring and the returned facial geometry details are incorrect and noisy due to their lack of a priori model constraints.Methods to solve these problems usually rely on the appearance conditions of the input images or on multi-view datasets of the required input objects.Therefore,this paper conducts an experimental study on the problems in unsupervised non-a priori methods,both in terms of input images and models,as follows:(1)To address the problem of requiring multi-view datasets of input objects in non-priori single-view image reconstruction tasks.In this paper,we design and use U-Net discriminator to generate new view image of input image.The input image and the rendered image regressed by the renderer with the new view are fed into the U-Net discriminator.Combining the input image,the encoder will classify the pixels into real and false regions at the pixel level for the rendered image,and the decoder will learn the global and local differences between the real and false images based on the input signal,and output an image that is consistent with the resolution of the rendered image and real at the pixel level.This image can be used as a conditional constraint on the input image as a way to build multi-view consistency.(2)In order to solve the problem that the geometric details of the face are incorrect in regression,this paper designs the Rotational Cycle Consistency.The input image is discriminated by U-Net to generate a new view real image,and the new view is rendered back to the original view,thus constituting a mutual translation between two domains performed in cyclic consistency.The resulting result should be as similar as possible to the input image,and both are combined with the rendered image of the input image regression,using 2D feature loss for continuous perceptual refinement.In this way,the multi-view consistency loss is implicitly constructed,thus improving the model’s ability to resolve the geometric information of the input image.(3)To solve the problem that the reconstructed model depends on the appearance condition of the input image and the prediction texture is fuzzy,in order to obtain more real and accurate texture information of the occlusive parts of the face,this paper designed the Texture Align Cycle Consistency.For the input image and the new view image of the input image,the albedo and illumination information between the two is exchanged and then rendered into a new image,which is then predicted and exchanged again and compared with the original image for 2D features.During the exchange process,the details of the albedo representation are encouraged to be stored in the same place to make the shared information between images recover the true texture from the invisible position in the original view for the view generalization effect.In this paper,experiments are conducted on the CeleA and 3DFAW dataset,and the experimental results show that the improvements proposed in this paper on the input image and model design can predict more realistic information from a single view,reconstruct a face model with better details,and effectively solve the self-obscuring problem of the input face image.