Research And Application Of Face Manipulation System Based On Deep Learning

Face manipulation aims to manipulate single or multiple attributes of facial images to generate new facial images with the required attributes,while retaining other original details,and has a wide range of application needs in many fields such as intelligent monitoring,entertainment and social networking.Traditional face manipulation needs to be done with professional image editing software,but because ordinary users often lack professional image editing skills,this approach does not meet the needs of most ordinary users.At present,many companies have launched targeted face manipulation applications,such as the Meitu Xiuxiu produced by Meitu,the sprouting and light-faced cameras launched by Byte Dance.These products meet the needs of ordinary users to a certain extent,but there are still some defects,such as the loss of original details after editing and adjusting the image,the operation is not simple enough,and the corresponding functions need to be further optimized.In order to solve the above problems and obtain a more realistic and natural face image editing effect,this paper designed a face image editing system with the help of deep learning network.The system is composed of three parts: face detection network and face key point detection network,face segmentation network and face manipulation network.For face detection and face key point detection,in order to achieve higher accuracy,face detection and face key point detection generally rely on independent detection networks.However,the network designed in this way often has more parameters,which makes the network inference more computational and slower.Considering the strong correlation between the two tasks of face detection and face key point detection,this chapter has designed a lightweight face and face key point joint detection network-Flash Net,using multi-task learning to The two tasks are modeled under the unified neural network framework,and the structured knowledge distillation technology is used to further improve the detection accuracy of Flash Net.For face segmentation,this paper proposes a lightweight face segmentation network-Flash Seg Net.The network uses a fast downsampling method to reduce the amount of network operations,and obtains large context information through a global pooling operation,and uses the attention mechanism and feature fusion to enhance the ability to express features.In addition,this paper uses a structured knowledge distillation method for semantic segmentation to improve network accuracy.Aiming at the problem of the imbalance of the number of sample categories,this paper designed a weight adjustment coefficient to control the gradient return intensity during the training process to improve the segmentation accuracy of the category containing a small number of samples.For the face attribute editing part,this paper uses a conditional generation adversarial network to generate face images.In order to obtain high-definition face images,a coarse-to-fine generator structure and a multi-scale discriminator structure are used.In order to decouple the face attributes so that each part of the attributes will not affect each other when editing,this paper uses the semantic segmentation mask as the generation condition and proposes a new semantic region adaptive normalization for the face attribute editing task Module to improve the quality of the generated face.The core of this article is to use deep learning tools to achieve faster and more natural face manipulation.A large number of experimental results show that the accuracy of the algorithms proposed in this paper can meet the requirements of the face manipulation system,and users can achieve a higher degree of freedom in face manipulation.