End-to-End 3D Face Reconstruction Based On Multi-Objective Evolutionary Neural Network

Three-dimensional(3D)face applications have attracted a lot of attention in multimedia and image processing fields.However,the neural network-based 3D face reconstruction methods are still not mature enough,such as overfitting of the neural network itself,estimation of large pose faces,computational latency of the model and poor 3D face texture alignment performance.Therefore,in order to solve the above defects,three different neural network-based 3D face reconstruction methods are proposed in this paper,including a multi-objective evolutionary encoder-decoder model,a multi-objective evolutionary attention network model and a self-supervised network model based on knowledge distillation and multi-objective evolutionary.The main work of this paper is presented as follows.(1)To alleviate the impact of overfitting effect on 3D face reconstruction results,this paper designs a novel approach to construct 3D face,namely,multiobjective evolutionary 3D face reconstruction based on improved encoderdecoder network.The model introduces a regularization algorithm called feature map disturbance,which aims to enhance the generalization ability of the network.Based on this,we construct a multi-objective evolutionary 3D face reconstruction model with disturbance probability,size of disturbance blocks,distortion strength,probability step size and learning rate as decision variables and loss function and structural similarity(SSIM)as objective functions.In addition,four multiobjective evolutionary algorithms(NSGA-II,AGEII,NSLS and MOEA/D)are used in this paper to achieve the optimization of the proposed model.Extensive experimental results show that NSLS has better convergence and diversity of sets.Moreover,the present model achieves better performance compared to other stateof-the-art algorithms.Therefore,the proposed multi-objective evolutionary 3D face reconstruction model has outstanding 3D face reconstruction performance in terms of large poses and facial expressions.(2)Although the above method achieves good results in fitting some frontal faces,the reconstruction effect is poor in the case of faces with large poses.Therefore,based on this situation,a 3D face reconstruction method based on the attention network model is proposed in this paper.The model has the features of the above model,but also incorporates an additional attention network and an average face synthesis strategy.The use of the attention network enables the network to focus more on the main parts of the face during feature extraction,while the average face synthesis enables the network to make reasonable complements to the unseen parts of the face when returning to a large pose face.Finally,the proposed model is analyzed and compared with the above model to find higher reconstruction accuracy and robustness.(3)Considering that the above methods all use supervised networks for training and their requirements for dataset labeling are high.However,the diversity of faces in the current face dataset with 3D labels is still insufficient,and overly complex models increase the model latency,which makes the application of 3D face reconstruction a great challenge.Therefore,in this paper,we propose a self-supervised 3D face reconstruction method based on knowledge distillation and multi-objective evolutionary networks to further improve the network operation speed.The method compresses the network and improves the network speed by introducing a knowledge distillation framework and Adder Net technique.In addition,a new loss function is designed and a multi-objective evolutionary algorithm is used to achieve weight adaption.Teacher loss,student loss,and structural similarity(SSIM)are used as objective functions,which in turn leads to better reconstruction performance.Finally,by comparing with other excellent works,the experimental results verify that the proposed model in this paper has better performance and shorter running time.Moreover,the model has excellent3 D facial reconstruction performance under extreme lighting conditions.This also provides a new idea for future 3D face reconstruction methods.