Research On Craniofacial Reconstruction And Realistic Processing Based On Generative Adversarial Network

Craniofacial reconstruction is to reconstruct the face from the skull based on the craniofacial relationship to help researchers identify the research target.The existing computer-aided craniofacial reconstruction methods have insufficient details of the five sense organs and low realism,combined with artificial intelligence and image translation methods,to achieve a high-realistic Mongolian craniofacial reconstruction.Our method has promoted the application and promotion of craniofacial reconstruction in archaeology,medical science,criminal investigation,and other fields.The research in this paper mainly includes:(1)To solve the problem of 2D craniofacial image translation,the 3D craniofacial model was firstly preprocessed,and a data set containing 2210 images was constructed through orthographic projection and data augmentation technology to provide data for the experiment.(2)Aiming at the disadvantages of traditional methods with insufficient ability to learn deep information of craniofacial data and insufficient ability to express specific features of the data set,our research proposes a generative confrontation model for craniofacial reconstruction:CFR-GAN.In order to stabilize model training and expand the acceptance domain,the model is divided into two steps: coarse reconstruction and fine reconstruction.Coarse reconstruction reconstructs the overall structure content of the corresponding human face through the skull,and fine repairs and refines the facial features’ contour lines of the reconstructed face.The fine repair network is composed of an attention generator and a multi-level discriminator: the attention generator is a convolutional neural network with skip connections,which are used to generate facial images with the same identity attributes as the skull since the attention mechanism helps the generator capture the facial information of the feature map;the multi-level discriminators are composed of a face discriminator and three region-of-interest discriminators,which has ability to generate the authenticity of different parts of the face from different fine-grained definitions.The experimental results show that the CFR-GAN model synthesized image is high-fidelity and has a good craniofacial reconstruction effect.(3)To solve the problem of insufficient realism of the reconstructed face,this paper proposes a weakly supervised image translation method based on facial feature recognition: Deep Face2 Real.Generative Adversarial Networks are used as the basic mapping model,Gaussian noise is added to the generator to generate the details of the face,and the attribute discriminator is added to the discriminator to assist in the generation of important attribute features in the real face.In order to alleviate the problem of model collapse,the transfer learning method is used to transfer the facial feature distribution learned from the pre-trained source domain samples to the target model task to improve the generalization of the target model;a noise consistency loss is designed to guarantee the relative similarity of details between instances in the task.The experimental results show that the model overcomes the shortcomings of traditional methods of strong subjectivity and insufficient retention of important facial features,and can provide high-quality realism processing for reconstructed faces with different attributes.Based on the above algorithms,this paper constructs a complete system of craniofacial reconstruction and face photorealism,which has a complete set of reconstruction functions from craniofacial data preprocessing to craniofacial reconstruction,and then to face photorealism processing.