Research On Tiled Clothing Transformation Method Based On Generative Adversarial Networks

With the popularity and development of the Internet,e-commerce and online shopping have penetrated into personal life all the time.Among them,the purchase of the same product in the video or on the fashion website has gradually become a major craze in online shopping.How to accurately find the same item in the huge e-commerce database has become a major problem at present.The traditional method is that process the image containing the human body.First use image detection technology to cut out the clothing on the human body in the image,then put them into the database for clothing retrieval,and finally find a few clothes with the highest similarity for recommendation.Traditional method is caused by complex backgrounds,human skin,deformation of clothing,etc.,which will cause the results of clothing testing contain some noise.Such errors will accumulate in the clothing search.In this paper,the generative adversarial network is used to convert the human body images into a tiled clothing images.Only the white background and the tiled clothing are preserved in the tiled images,and the tiled clothes images are used for clothing retrieval.This method replaces the original clothing detection method and improves the retrieval accuracy of the clothes to some extent.At present,the generative adversarial networks has gained many important applications in the field of image transformation.Based on the existing generative adversarial network model and combining it with the clothing image field,this paper proposes three new generative adversarial network model,including:(1)adding a feedforward neural network for the classification of clothes.(referred to as the classifier),the loss generated by the classifier is back propagated to the generator,guiding the training and iteration of the generator;(2)In order to make the image transformation more user-friendly,from the buyer's point of view,the model adds a clothing category condition to the generator as a guide,simulating the buyer to provide category requirements.And using the triple loss function instead of the original loss function of the discriminator,so that the discriminator has its own discriminating ability and enhances the intensity of the discriminator loss;(3)Using the idea of cascading,the generator is divided into two phases.The image generated by the first phase being the input to the second phase generator.The second phase uses a two-layer generator and feature matching structure that produces more clothing detail.In order to reflect the superiority of the proposed model,a large amount of raw image data is collected by the crawler.Through a series of pre-processing operations,a one-to-one sample pair dataset consisting of a human body image and a tiled clothing image is constructed.Experiments show that the proposed method is superior to the existing generative adversarial network model in the quality of tiled clothing image.At the same time,using the generated images for clothing search evaluation,the results also show the usability and effectiveness of the method.