BatIK-DG: Improved deblurgan for batik crack pattern generation

Abstract

As a traditional handcraft in China, batik is valuable for both decoration and fashion design due to its unique crack pattern. However, previous researches on batik crack pattern generation remain space in promoting verisimilitude. In recent years, Generative Adversarial Networks (GAN) show great ability in realistic image generation, and DeblurGAN performs well in image deblurring task, showing good potentiality in image generation. In this paper, we improve DeblurGAN for batik crack pattern generation by several methods. First, transposed convolution is replaced with resize-convolution. Second, Hybrid Dilated Convolution (HDC) is added to the network. Third, RaLSGAN is utilized instead of WGAN-GP, and several additional loss functions such as L1, L2 and so forth are tested respectively. Our network is trained on BIFT-Batik dataset and performs well to some degrees both in visual verisimilitude and IQA indexes including PSNR and SSIM. We name our network as Batik-DG (Batik-DeblurGAN).