Unsupervised 3D Face Reconstruction Method Based On Feature Fusio

At present,with the rise of 3D animation,3D games,intelligent medicine and other industries,3D face reconstruction technology has become one of the hot research directions in computer vision and computer graphics.A large number of researchers began to pay attention to the development of 3D face reconstruction.Because of the high cost of using precision instruments for 3D face reconstruction,3D face reconstruction based on single face image has become the mainstream of research in this field.The traditional 3D face reconstruction methods based on single face image have some problems of low accuracy and low efficiency.With the development of deep learning methods,the 3D reconstruction methods based on deep learning for single face images not only greatly improves the accuracy of the reconstructed 3D face model,but also greatly improves the efficiency of reconstruction Therefore,deep learning method has been widely used in 3D face reconstruction.However,due to the use of a simple codec network structure for model training,the high level of the network lacks the underlying structural information,resulting in inaccurate shape and lack of details of the reconstructed face.This paper makes a related research on the above problems,and proposes two different feature fusion unsupervised 3D face reconstruction methods,which effectively improve the performance of face model reconstruction.In order to solve the problem of inaccurate shape of the reconstructed face model,a 3D face reconstruction method based on channel multi-scale feature fusion is proposed.This method can fully integrate the feature information of different channels of the feature map and pay attention to the face regions of different scales,so that the network can learn more global feature information,and then improve the accuracy of face shape reconstruction.Specifically,the feature maps of different channels are grouped and fully integrated so that the network can learn more feature information.In addition,the large kernel attention mechanism is introduced into the encoder of each decoupling factor,which can fully extract effective facial feature information on a single channel and suppress the background noise of the image.Qualitative and quantitative experiments on BFM,3DFAW and Photoface face datasets demonstrate the effectiveness of the proposed method.In order to solve the problem of lack of details in the reconstructed face model,a 3D face reconstruction method based on context detail feature fusion is proposed.This method can combine the high-level semantic information of the network and the underlying structural information to supplement the missing face detail information,so as to reconstruct the face model with rich details.Specifically,the upper layer feature information is fused into the lower layer network from top to bottom in the coding network,and the coding information of the corresponding layer is sent to the decoder,so that the decoded depth map has rich detailed information.In addition,in the coding of each decoupling factor,the channel and space mixed attention mechanism is introduced to suppress the image background noise,so as to fully extract the effective facial feature information;in the network training phase,the loss of facial key points is also introduced to constrain the generated face shape.Qualitative and quantitative experiments on BFM,3DFAW and Photoface face datasets demonstrate the effectiveness of the proposed method.