Research On Simulation Of The Batik Crack Pattern Based On GAN

Batik is a traditional Chinese folk art.As a national intangible cultural heritage,it owns a high artistic value and cultural connotation.Using computer technology to simulate the unique crack pattern of batik is both meaningful for designers to quickly obtain the batik pattern and helpful for the spread and development of batik culture.The Generative Adversarial Network(GAN)is a neural network structure that has developed rapidly in recent years.It has superior performance over ordinary neural networks and has been widely used in image generation and other fields.So,applying GAN on the simulation of the batik crack pattern is feasible to some degrees.This paper trained and tested several representative networks,including SRCNNEx,FSRCNN,U-Net,pix2 pix,Cycle GAN and Deblur GAN on the self-built batik dataset.And further improvements were carried out in Deblur GAN including replacing transposed convolution with resize-convolution,adding three layers of atrous convolution,using Ra LSGAN instead of WGAN-GP as adversarial loss and complementing the loss function with L1 Loss.The improved network performs quite well in simulating the batik crack pattern and is named as Batik-DG.This paper then implemented the improved residual modules in Inception-Res Netv2 and ESRGAN on Batik-DG and discussed their effects on improving the quality of batik crack pattern.The networks are respectively named as Batik-IR and Batik-ES.This paper finally researched on the miniaturization of the network.Based on the technologies such as Deep Separable Convolution and Linear Bottleneck/Inverted Residual Block from Moblie Net-v1 and Mobile Net-v2,the attempt to miniaturize Batik-DG,Batik-IR and Batik-ES is carried out and resulted in both less network parameters and faster inference speed while maintaining the performance in simulation.