Research Of Virtual Wearing Technology Based On Generative Adversarial Networks

With the maturity of the e-commerce industry,shopping on the Internet has become a new consumption habit.However,the selection only through the product pictures and parameters displayed cannot provide a good shopping experience for consumers,which is particularly obvious in the sales of apparel products.The virtual wearing system can make up for the lack of wear when shopping online,and consumers can see the effect of the wearing through the network.The generative adversarial networks is a network model designed from the perspective of game theory,the process of learning from each other through the generator model and the discriminator model gives it the ability to simulate the real data.Therefore,this paper uses the generative adversarial networks to conduct virtual wearing technology research,and trains the network model to generate virtual wearing effect.The main work of this thesis includes the following aspects:1.Research and implement a virtual wearing algorithm based on VGG-GAN.Combined with the characteristics of the VGGNet model,the designed generative model has a depth of 12 layers,contains 10 convolutional layers and 2 fully connected layers,the convolutional layers all uses 3×3 small convolution kernels.The discriminative model is an 8 layers full convolution structure,using ELU nonlinear activation function.The wearing effect generated by this model was evaluated,among which the relatively ideal wearing effect accounted for 55%of the total number generated,and the test duration of a single image was 1.85s.2.Research and implement a virtual wearing algorithm based on ResNet-GAN.Combined with the characteristics of the ResNet model,the designed generative model has a depth of 15 layers,contains 6 residual block,a total of 13 convolutional layers and 2 fully connected layers,the first convolutional layer uses a 7×7 convolution kernel,and the remaining convolutional layer uses a 3×3 small convolution kernel.The discriminative model is an 10 layers full convolution structure,using ELU nonlinear activation function.The wearing effect generated by this model was evaluated,among which the relatively ideal wearing effect accounted for 74.5%of the total number generated,and the test duration of a single image was 1.76s.3.Design and implement a virtual wearing algorithm based on the Django framework.This thesis designs the virtual wearing webpage by building the Django framework and configuring the database,and connects the ResNet-GAN model with the Django project.The average time spent running the webpage to display a wearing effect is 2.6s.