Application Of Generative Adversarial Network In Fashion Style Transfer

In recent years,with the rapid development of computer hardwares and technologies,the proportion of digital images for information exchange in people's daily work and life is becoming higher and higher.Some research work related to image analysis has also been paid more and more attention by more and more researchers and has a very rapid development at the same time.The style transfer of clothing is to modify the clothing in people's photos with the required style.In the case of inputting the basic clothing image and the target style image,the main goal is that the synthesized clothing image should be able to fuse the style of the given reference style image while maintaining the original structure and shape of the clothing.It has important research significance and practical application value to carry out the research.With the wide application of neural networks in image processing,computer vision,and other related fields,especially the proposal and rapid development of generative adversarial networks(GANs),a series of image translation work related to this has attracted the attention of many researchers.It is an interesting and meaningful task to automatically transfer fashion style of fashion images.In this paper,when the model receives an input image,combined with the corresponding instance information(such as object segmentation mask),we use a Cycle GAN based deep neural network structure to achieve the style transfer of clothing segments with the background preserved,which can ensure the style transfer under the premise of clothing shape unchanged.Through the input fashion image and the required clothing style,the model generates a person image with the desired clothing style by an end-to-end method.The main work and research contents of this paper are as follows:1.Firstly,the paper introduces the research background and significance of style transfer of specific clothing segments while the background remains unchanged;then,it analyzes and summarizes the research status of the generative adversarial network itself and its application in the field of image translation;last,the generative adversarial networks and image translation principle related to the content of this paper are explained,and the related models and algorithms of neural style transfer and clothing style transfer are analyzed.2.Based on the Cycle GAN model,this paper proposes a clothing style transfer model.Most of the existing style transfer models can only transfer the style of a single object in a relatively simple background,and can not complete the related work well in the face of complex background and variable instance shape.The research similar to our usually need to cut the corresponding region,transfer the style first,and then paste it back to the corresponding position of the original image.The overall process is very cumbersome and the application is limited.In this paper,when the generator and discriminator receive the input the image,the segmentation mask corresponding to the instance information in the image is added to achieve the style transfer of the clothing in the specific area,and at the same time,the whole background outside the scope of the instance remains completely unchanged before and after the image translation.The experimental results show that the proposed model can change the corresponding clothing style on the premise of ensuring the style and shape of the original clothing while keeping the backgrounds completely unchanged except for the specific clothing area.At the same time,the comparative experiments also show that the method can effectively complete the style transfer task of the corresponding clothing segment,and the overall performance of the generated image is good.