Research On 3D Face Reconstruction Algorithm Based On Deep Learning

With the rapid development of computer technology,3D face reconstruction technology has become a hot research topic in the field of computer vision.The three-dimensional face contains more identity feature information than the two-dimensional face,and can be applied to more abundant real life scenes,such as face recognition,film and television entertainment,medical beauty,etc.Due to the high cost of collecting 3D real face data,and the traditional 3D face reconstruction algorithm is inefficient and inaccurate,this paper proposes a 3D face reconstruction algorithm based on 2D face images,combining traditional 3D Morphable Model(3DMM)with deep learning algorithms,and uses deep neural networks to directly predict the 3DMM coefficients needed to reconstruct 3D human faces from 2D face images.The operation of the algorithm is simple and the cost is low,which has widely practical application prospects.Aiming at the problem that the traditional face alignment algorithm cannot detect the invisible face landmarks in the large pose face image,this paper proposes a joint 3D face alignment and 3D face shape reconstruction algorithm,that is,vertex sampling is performed on the reconstructed 3D face to obtain the corresponding face landmarks of the image.This paper improves several problems in the 3D Dense Face Alignment(3DDFA)algorithm:First,select DenseNet to replace the cascade regression network in the 3DDFA algorithm to improve the model's fitting ability.Second,a semi-supervised learning model is proposed,which uses a large number of 2D face images without 3D labels to solve the problem of insufficient training data.The alignment accuracy of this method on the public data sets AFLW and AFLW2000-3D has been improved by 1.5%and 2.9%,respectively.In addition,in order to solve the defect that the traditional texture mapping method cannot extract the texture of the invisible area of the face edge,this paper proposes a texture optimization algorithm to improve the texture at the edge of the 3D face.In order to further solve the problem of face texture reconstruction under the extreme situations such as large poses or occlusion in the input image,this paper proposes a weakly supervised learning algorithm that can reconstruct 3D face shape and texture information at the same time.In order to reconstruct face texture,this paper uses a differential renderer SoftRas to build a pixel-level weak supervision signal between the reconstructed 3D face and the 2D face image;in order to reduce the impact of occlusion on texture reconstruction,this paper uses a face segmentation algorithm to preprocess the face image to remove occlusion;in order to improve the reconstruction accuracy,a multi-level loss function is designed in this paper,without the need for three-dimensional label data,this algorithm can learn many kinds of weakly supervised information from 2D images directly.The qualitative and quantitative comparison experiments prove that the algorithm has achieved a certain improvement in reconstruction quality and accuracy,and can reconstruct a realistic 3D face.