The Study And Application On Clothing Generation Method Based On Conditional Generative Adversarial Networks

Nowadays,the world has entered the information era,through which e-commerce has achieved great development,and it has increasingly become an indispensable part of peoples’ and especially young people’ daily lives.Therefore,purchasing clothing online is gradually being accepted by each individual.However,e-commerce platforms include numerous clothing merchants and various types of clothing,which make it difficult for consumers to find suitable clothing and greatly affects the shopping experience of consumers.In addition,it also has some drawbacks for buying clothing online though it brings us convenience.A case in point is that it is not capable of trying on them,which is quite unlike in an offline store that can make consumer to decide whether to buy or not based on the effect of the clothes on.On the other hand,various e-commerce platforms are competing for the lead as the amount of users is an important indicator to measure the popularity of various e-commerce platforms.The greater the number of users,the greater the flow,and the greater the potential value.As a result,how to use the massive amounts of data readily available on the Internet to improve consumers’ buying experience and attract new customers is of paramount importance.In the past few years,generative adversarial networks in deep learning have achieved great development,and it becomes more and more closely connected with the clothing industry.The application of generative adversarial networks to generate clothing to solve the problems of the clothing industry has also become the mainstream.Consequently,it is a wise choice for e-commerce platforms to improve their performance by taking advantage of generative adversarial networks and the massive amount s of data available on the Internet to solve the dilemma of consumers’ poor purchase experience and to enhance consumers’ desire to buy.Therefore,research on clothing generation technology is of great significance for both e-commerce platforms and consumers.Motivated by the application background of virtual dressing and clothing retrieval,we in this paper,combine conditional generative adversarial networks and attention mechanism in deep learning to generate high-quality tiled clothing images according to the clothing worn on the human body.The main contributions of this paper are as follows:1.We design a high-resolution clothing generation method based on conditional generative adversarial networks.This method is mainly based on the pix2 pix HD model.In detail,this high-resolution image generation model is applied to the clothing dataset for the first time and aiming at the problems of cumbersome model training process,long training time and loss of texture in the generated results when the model generates high-resolution clothing images,an improved version of pix2 pix HD model is proposed.The improvement is mainly reflected in three aspects:(1)Improve the loss function.Added perceptual loss to improve the perceptual similarity between the generated result and the ground-truth.An improved version of feature matching loss is proposed and different weights are assigned to feature maps of different layers within the discriminator when calculating feature matching loss.(2)Improve the training strategy.The end-to-end training method is adopted,without step-by-step training,which simplifies the training process.(3)Improve the network model.We abandon the multi-scale discriminator with a huge amount of calculation,and adopt the Markovian discriminator to reduce the amount of parameters and shorten the training time.The improved model not only avoids the cumbersome and time-consuming training of high-resolution clothing generation models,but also solves the problem of texture loss in the generated results,greatly improving the performance of high-resolution clothing image generation.In both qualitative and quantitative performance evaluation,the improved model we proposed is superior than the original model,and the model training time is shortened by nearly one day.2.Based on the problems encountered in first research,conditional generative adversarial networks are combined with attention mechanism and a multi-stage attention-based category-supervised high-quality tiled clothing generation method is proposed.This method is inspired by the multi-stage image generation method,which deals with different problems at different stages.In the first stage,a coarse picture of the tiled garment is generated.(1)Introduce the spatial transformer module to overcome the problem of deformation of the generated result.When the shape of the input human body picture deviates greatly from the target shape,shape transformation is performed on both the input and the output to reduce the deformation of the generated clothing picture.(2)Add category supervision information to increase the controllability of the categories of generating clothing.When there are multiple types of clothing for the input human body image,the wrong type of clothing may be generated.Therefore,the generated result is transmitted to the classifier,calculating the classification loss to constrain generation direction,and handle the ambiguity of network.In the second stage,high-quality tiled clothing is generated from the coarse results of the first stage.(3)A dual-path generator based on attention mechanism is proposed.The shape information and category information contained in the coarse clothing generated in the first stage are merged with the detailed information contained in the input human body picture.(4)The channel attention module is introduced to replace the skip connections,and the self-attention module simulates the global dependency between pixels.The channel attention module selectively pays more attention to the channels where more valuable information is located,and filters redundant information such as limbs and faces.The self-attention module overcomes the weakness of the small convolution kernel which only has the local receptive field and obtains the correlation of long-distance region.The clothing generation experiments conducted on the clothing dataset we established demonstrate that our proposed method is superior to other similar methods in both quantitative and qualitative results.The ablation experiment and other additional experiments further prove the novelty of our method.3.Apply the high-quality tiled clothing our method generated to practical applications.Using state-of-the-art virtual try-on method CP-VTON,the results also prove that our network has a high practical value.4.A supervised image-to-image translation dataset is established.The dataset contains more than 34762 pairs of pictures and 10 types of clothing.Each pair of pictures contains corresponding type.The number of pictures of each category is controlled at about 2000 to avoid uneven samples.90% of the image pairs are employed as the training set,and the rest are used as the test set.