Apparel Product Design Based On Generative Adversarial Networks

Generative Adversarial Nets(GAN)is a special deep learning model which performances through mutual game learning between the generator and the discriminator.In this thesis,GANs are employed to implement automatic apparel style fusion and transformation(take footwear as an example),in which two GAN model are designed.One is SFGAN(Style Fusion GAN)for apparel style fusion,another is SCGAN(Style Transformation of Clothing GAN)for apparel style transformation.The main work is as follows:1.Build the(footwear)apparel images dataset.The dataset contains five categories with a total of 11,466 images,including 3265 high heel images(thin: 1637;thick: 1628),865 slipper images,2971 sports shoe images,3633 casual shoe image(non-laced: 1971,laced:1662)and 732 leather shoe images(non-laced: 362;laced: 370).The resolutions of all the images in the dataset are set to256×256.2.By drawing on the idea of progressive growing network,a special SFGAN network is designed to achieve the task of apparel style fusion.The generator network of SFGAN has a depth of 34 layers,including 19 convolutional layers(1×1,3×3 convolution kernels),8 fully connected layers,and 7 upsampling layers;while the discriminator network of SFGAN has a depth of 22 layers,which contains 15 convolutional layers(1×1,3×3 and 5×5 convolution kernels),1 fully connected layer and 6 downsampling layers.The experimental results show that,(1)using the subjective identification of the human eyes,the SFGAN model generates an ideal effect accounting for 51.90%,which is 4.05% higher than the Style GAN model.(2)using FID(Fréchet Inception Distance),the SFGAN model reaches 40.33 points,which is 1.45 points lower than the StyleGAN model.3.By drawing on the cycle-consistent loss for reference,the SCGAN network is designed to achieve the task of apparel style transformation,in which the generator network has a depth of 18 layers,including 4 convolutional layers(3×3,7×7 convolution kernels),2deconvolution layers(3×3 convolution kernels)and 6 resnet?block(each of them includes two3×3 convolution kernels),and the discriminator network has 5 convolutional layers(5×5convolution kernels).An attention mechanism layer is added to the seventh and eighth residual layers of the SCGAN generator,and the traditional convolutional feature map is replaced with a feature map with attention to further improve the performance of the model.The experimental results show that,(1)using the subjective identification of the human eyes,the SCGAN model generates an ideal effect accounting for 56.00%,which is 7.50% higher than the CycleGAN model.(2)using FID(Fréchet Inception Distance),the SFGAN modelreaches 38.98 points,which is 2.96 points lower than the CycleGAN model.