Research And System Development Of Facial Makeup Transfer Technology Based On Deep Learning

Makeup transfer is an important problem in the field of computer vision,which involves transferring a person’s makeup from one image to another,and has high research and practical application value.Although the existing research has achieved outstanding results in the makeup transfer task,there are still various problems.Existing methods either adaptively normalize deep content features according to style to match their global statistics,or fuse deep style features into deep content features without considering feature distribution.On this basis,this paper combines the advantages of AdaIN and attention mechanism,and proposes a makeup transfer model based on AdaAttN.The main work content of this paper is as follows:1.Use AdaAttN as the feature conversion module in the makeup transfer task.Specifically,our model adaptively shifts the feature distribution at each point by using an attention mechanism while considering both low-level and high-level features.It is divided into three steps in total,one is to calculate the attention map generated using content features and style features from the shallow layer to the deep layer;the second is to calculate the weighted average and standard variance map of the style features;Adapt normalization to achieve feature distribution alignment.This not only improves the quality of makeup transfer,but also ensures the controllability and flexibility of the model.2.Optimize the feature extraction module of perceptual loss.This paper pre-trained a ResNet-50 on a large face recognition dataset VggFace2 to extract high-level features of images.Compared with general models such as Vgg19 or InceptionNetV3 trained with ImageNet,models trained with face recognition datasets can better reflect the differences between different face identities.3.Use data augmentation strategies.The input image and its corresponding segmentation mask are rotated by 30 degrees,60 degrees and 90 degrees respectively.This method can not only increase the amount of training data,but also allow the model to learn faces from different angles,thereby improving the generalization ability of the model.4.Model deployment.Using TensorRT for optimization and acceleration,and packaging it into an SDK and integrating it into a WeChat applet not only significantly improves the reasoning speed of the model,but also reduces the threshold for using the model.5.System development.The model in this paper is integrated into the WeChat applet,and the requirements analysis,general design,detailed design,system implementation and testing of the applet are introduced in detail.