Research On Deep Learning Algorithm For 3D Face Reconstruction Based On Single Imag

In recent years,with the development of 3D digital industry,3D face reconstruction has become one of the research mainstreams in the field of computer vision,which attracts the attention of more and more researchers.The single-image based 3D face reconstruction has become the mainstream of research because of its simple data acquisition and easy modeling.The traditional single-image based 3D face reconstruction model has problems of low reconstruction accuracy and high modeling cost.However,with the development of deep learning,the single-image based 3D face reconstruction method has greatly improved the accuracy of the reconstructed 3D face and has been widely used in the 3D face reconstruction task.In this paper,we conduct a study on some problems in the existing single-image 3D face reconstruction models based on deep learning,and propose two different 3D face reconstruction methods,which effectively improve the performance of the models.A 3D face reconstruction method based on feature and shape context information is proposed.To address the existing problem of insufficient fusion of contextual information in3 D face reconstruction based on UV position map,this paper adds constraints between facial feature contexts and between shape contexts in the reconstruction of 3D faces.Specifically,the method uses the UV position map to characterize the 3D face shape,and extracts features and returns the UV position map to the input face image by the encoder-decoder network respectively.On the one hand,in order to enrich the face information in the encoder features,a feature context correlation modulator is designed in the feature encoding process to perform feature resolution and correlation on the global features.Hence,the feature maps output from the encoder module contain face context information and the network adaptively focuses on the features in the face region,thus improving the accuracy of the global shape of the reconstructed 3D face.On the other hand,in order to obtain more accurate 3D face shape,the shape context vector is constructed using the constraints of the face shape,and the reconstructed 3D face is constrained using the shape context vector,thus improving the 3D face reconstruction accuracy.The experimental results on AFLW2000-3D,AFLW-LPFA datasets prove the effectiveness and superiority of the proposed method.A self-supervised fine-grained 3D face reconstruction method based on the face detailed mask is proposed.To address the problem of inaccurate local details of existing 3D face reconstruction methods,the detailed mask is used to improve the accuracy of local face details.Specifically,a detailed face mask generator is used to generate the face detailed mask.Aside from the basic loss functions such as Face shape constraint,image reconstructing consistency loss,image identity perception loss,etc.,the face detailed mask is used to give refinement constraints on the face region and self-supervised constraints on the face detailed mask,so as to improve the local accuracy of the reconstructed 3D faces.Qualitative and quantitative experimental results on the AFLW2000-3D and MICC Florence datasets demonstrate the effectiveness and superiority of the proposed method.