Research On Clothing Image Synthesis Based On Generative Adversarial Network

With the rapid development of artificial intelligence in the clothing industry,technologies such as virtual fitting,fashion display and virtual reality have been widely used.Online clothing purchases have gradually become a trend.When shopping online,online users generally want to be able to view multiple photos of fashion models in different poses.In order to meet the needs of users,clothing image synthesis technology can be used to enhance the user’s shopping experience,while also reducing business costs.Restricted,it is difficult to handle complex distributed image synthesis,and for the rapid development of deep learning,generative confrontation networks are widely used in the field of image synthesis,but traditional generative confrontation networks have defects such as blurring of generated images and distortion of training.For the above problems,this thesis updates the mechanism and builds a clothing image synthesis system in conjunction with a generative confrontation network.The main tasks are as follows:(1)In order to solve the problem of loss of detailed information caused by human body pose estimation based on traditional convolutional neural networks,the attention hourglass network for human body pose estimation is proposed here.The traditional method uses coordinate regression to extract the closed nodes of the two-dimensional human body.However,as the network deepens,the information of the human body joint points will be slowly lost.In order to overcome the problem of the loss of information leading to the reduction of the accuracy of the predicted joint points,here in the traditional convolution preserved a neural network based on the mechanism of the feature extraction module.The module is composed of a deep separable convolution unit and a channel attention mechanism unit.It is replaced with the residual module in the traditional hourglass network,and the deep separable convolution is used instead Conventional convolution greatly reduces model training parameters and computational complexity.When training the target hourglass network,a feature matching loss function is added to solve the problem of gradient disappearance.(2)In order to solve the problem of the loss of human body and clothing edge information in clothing images,this paper proposes a mechanism-based semantic generation network.Although the traditional Bicycle GAN network performs feature fusion in a multi-scale manner,the edge information cannot be effectively extracted in order to better capture the hierarchical semantic information.Here,the attention mechanism is added to the last layer of the semantic generation network,and the last layer of the network Add Softmax regression to the channel of a layer of feature maps,put more attention on the edges of the human body and clothing,and fully extract semantic information.(3)A new type of clothing image synthesis framework is proposed.The size generator trains the network through the multi-scale generator and discriminator to obtain the field of view of the texture,which can make the texture generator generate images with consistent classification,and the multi-scale discriminator can generate finer texture details by the texture generator and high-quality robe generation,Large-scale realistic clothing images.In this paper,comparative experiments and result analysis are carried out on the MPII dataset,Deep Fashion dataset and Market-1501 dataset.Compared with other mainstream methods,the method proposed here has higher accuracy in predicting joint points,and is more accurate in image synthesis quality and Significant improvement in quantitative evaluation indicators.