Multi-Attribute 3D Face Reconstruction Algorithm And Parallelization Research Based On Unconstrained Conditions

At present,3D face reconstruction has been applied in many scenes.Such as the generation of real-life avatars of game characters,simulation of face shaping,and virtual image generation.Traditional face reconstruction methods mainly express 3D face attributes through optimization algorithms.These methods only focus on the overall representation of the face but lack specific attention to each attribute of the face.This will lead to the lack of geometric details in the reconstructed 3D face when facing unconstrained conditions.This paper presents a method to improve the result of 3D face reconstruction under unconstrained conditions.At present,3D face alignment and face reconstruction in large poses,extreme expressions,occlusions,etc.Those are still challenging problems.This paper proposes a new deep learning model for these problems to solve the 3D face reconstruction and face alignment under unconstrained conditions.The main work of this paper is as follows:1.When a single face image is under unrestricted conditions.Due to the lack of full capture of the attributes such as the pose,expression,and identity of the face in the traditional network,the reconstructed face lacks accurate geometric information.In order to solve the above limitations,this paper proposes a multi-attribute regression network(MARN)for face reconstruction.Among them,the method of multi-attribute regression can fully capture the characteristic information of pose,expression and identity attributes improve the learning ability of the network for each attribute.At the same time,three loss functions are introduced to constrain the attributes of identity,expression,and pose respectively to improve the accuracy of 3D face reconstruction and face alignment.2.The traditional single-branch regression network lacks the ability to extract more discriminative features from the input image,especially in the case of large poses,extreme expressions,and partial occlusion.In the network learning of the face's identity,expression and pose attribute parameters,it is difficult to capture the low-level local details of the face and the high-level semantic information,so that the reconstructed face lacks obvious geometric details.In order to solve the above limitations,this paper proposes an attribute fusion branch regression network(AFBRN)for face reconstruction.This method can allow the network to fully capture the discriminative facial features of the face,and improve the network's ability to learn low-level detailed features and high-level semantic information.At the same time,in the merged branch networks,the various attributes of the face share information with each other in the adjacent branch networks to achieve feature interaction to enhance the network's learning of local geometric details and improve the reconstruction of 3D face geometric details.3.In the process of face reconstruction,the 3D points constitute the face represented by the grid.In the traditional 3D face reconstruction algorithm,due to the large amount of 3D point data in the face reconstruction process,the matrix calculation is complicated,and the amount of network parameter calculation increases,thereby reducing the running speed and reconstruction efficiency of the algorithm.In order to solve this problem and effectively improve the performance of the algorithm,through the overall analysis of the algorithm structure,the time-consuming part of the network training is parallelized.Secondly,this paper also uses the principal component analysis method to reduce the dimensions of the multi-dimensional variables of the training data.Under the premise of not causinL,waste of resources.the training of the convolutional neural network is carried out in a data parallel manner and the network tasks are divided in a targeted manner to improve the training speed of the network.