Research And Implementation Of 3D Face Reconstruction Algorithm Based On Tensor Decomposition

With the rise of artificial intelligence technology,computer vision and other multimedia application technologies have also received more and more attention from industry and scientific research institutions,and related intelligent products have gradually entered people’s daily lives.3D face reconstruction and face analysis are important areas of computer vision.The most important data in the calculation and analysis process is the facial features and contour information of the face.Accurately and efficiently let the neural network learn the apparent and structural semantic information of the face is an important prerequisite for face key point detection and 3D face reconstruction.3D face reconstruction based on a single image is a long-standing challenging problem in the field of computer vision.At present,most face key point detection and 3D face reconstruction methods only focus on the apparent information of the face and lack the learning of the detailed information of the face structure,making it difficult for the neural network to infer in the case of large poses and occlusions.The key points of the occluded part are therefore difficult to reconstruct a fine 3D face model.In recent years,the end-to-end 3D face reconstruction method has been favored by many scientific researchers due to its concise reconstruction characteristics,but at the same time,its drawbacks have gradually emerged.Existing end-to-end methods need to consume a lot of computing resources and it is difficult to reconstruct rich 3D face details.This paper aims to propose a new depth tensor learning method to solve the fineness problem of redundant calculation and 3D face reconstruction,especially the restoration of facial details in an unconstrained natural environment.Specifically,we propose a dual-stream convolutional neural network combined with face super-resolution method,which can effectively restore the 3D position information of the image.The method also combines an attention fusion mechanism,which can learn the individual attention mapping of each feature subspace,learn multi-scale and multi-frequency features while effectively learning crosschannel information,and obtain the most discrimination in different local areas Features enhance the consistency and correlation between attention areas.At the same time,we added an improved CP decomposition method of higher-order tensors to compress CNN.This method can effectively eliminate redundant information and better maintain the original characteristics of the data.This reduces the computational cost and greatly improves the processing speed of a single image.We use the t-SVD method to limit the rank of the modified CP decomposition to effectively approximate the rank of the tensor,and use weighted iteration to compensate for the decomposition loss.The main work of the thesis is as follows:1.In order to solve the problems of the original PRNet method in detail reconstruction,that is,the surface of the reconstructed 3D face model is rough,and the details such as expressions are not clear.enough.This paper proposes a multi-scale detail enhancement method to obtain the local rich details of the image.Combining the attentional subspace mapping information for feature fusion,using more complementary advantages of different levels of features,effectively enhancing the detailed information of the image data.2.This paper introduces the TDNet network as the basic network.TDNet is both a two-stream neural network and an efficient end-to-end 3D face alignment framework.Through deep separable convolution.tightly connected convolution and lightweight channel attention mechanism,a stable network model is established.We use channel and spatial attention mechanisms to extract important facial and structural semantic information.The tensor decomposition method is used to compress the convolutional layer of the convolutional neural network to reduce the computational overhead of redundant information.Improve the accuracy and training speed of network regression.3.We designed a face key point detection and face reconstruction platform,and integrated the algorithm of this paper.It is designed to allow ordinary users to log in to this platform to perform face key point detection and face reconstruction,so that the algorithm proposed in this paper has been practically applied.