Research On Multimodal Face Generation Technology Of Simulated Portrait

As a common criminal investigation method,simulated portrait technology plays a huge role in public security work,but the traditional manual painting simulation portrait technology is faced with the difficulties of high painting skills and difficult modification,so it is difficult to apply this technology on a large scale.Although the computer face combination simulation portrait technology through the facial features library can effectively improve the mapping efficiency,the lack of diversity and intuition of the combined simulation portraits restricts the further development of this technology.The powerful image processing capability of deep learning provides new ideas for simulating portraits.Therefore,based on the deep generative network,this paper adopts different face editing strategies for gender attributes representing discrete attributes,age attributes representing continuous attributes,and face contours,to convert simple hand-drawn face into intuitive,realistic multimodal face simulation portraits in a convenient,efficient,refined and comprehensive way.Specifically:First of all,aiming at gender attributes,this paper proposes a face gender attribute editing method based on feature encoder with the help of the principle of face editing using hidden variables.Experiments show that the method works well.Secondly,aiming at the age attribute,this paper optimizes the face age editing method based on the style transfer network with the help of the condition-guided face editing principle:on the one hand,the deep connection attention mechanism is used to design the face age estimation network,and the prediction The trained face age estimation network is used in the training of the age attribute editing network to ensure that the model achieves a good age editing effect;on the other hand,the loss function of the face editing model is optimized,and the cycle consistency loss function and reconstruction are introduced.Loss function to further ensure the consistency of face identity.Experiments have verified that the face age estimation network and the optimized face age editing model have achieved good results.Finally,based on the advanced "sketch-face" conversion platform,this paper builds a multi-modal face generation system for simulated portraits,and designs a method to edit gender attributes and age attributes in the two modules of the model’s latent variable manipulation area and style transfer network.which not only integrates face generation and face editing tasks,but also realizes face contour editing.Gender attribute slider and age attribute slider control the target of generating face attributes.