Research On Face Attribute Generation And Style Transfer Based On Deep Learning

Face attribute generation and style transfer,as the research focus of computer vision,have been widely used in many fields such as face data expansion,human-computer interaction,virtual robots,and face restoration.In 2014,Goodfellow first proposed to generate a confrontation networkGAN.Its model task is to realize the fake picture,which makes the face attribute generation and style transfer have been developed rapidly.More and more people are researching the network,and many variant networks have been obtained through the improvement ofGAN.Improvements in different aspects have greatly improved the effect of the network.However,there are still some problems in the process of generating face attributes.For example,the occlusion of the face in the picture,the angle of the face and the color difference of the background will affect the authenticity and visual effect of the generation,and cannot make the generated picture truly fake.purpose.This article makes the following improvements to address the above problems:(1)Through the improvement of the normalization layer module in the network layer,the network pays more attention to the change of the attributes of the image,eliminating the interference of attributes and other factors that cause the distortion of image generation,and effectively solving the poor generation effect caused by the face angle problem in the image.At the same time,different loss functions are combined to train samples with smaller differences,in order to make the attributes of the same type closer,which can effectively distinguish the entanglement between the background and the attributes.(2)By introducing the idea of interpolation in the attribute tag to adjust the change range of the attribute,so as to control some of the interference caused by the attribute and the attribute after the attribute is changed,which causes the distortion and blurring of the image after it is generated.(3)Introduce the idea of style transfer to change and merge the face attributes of several images,which can effectively make the attribute generation more realistic.Through the ingenious design of the loss function in the style transfer,the original image and the style map are merged and generated.If the difference is small,the quality of the generated image will be high.(4)The experimental results show that the effect of the 13 attributes generation trained in this article has been improved,and the overall effect score is better than that of mainstream generation networks such as StarGAN and AttGAN.