Generative Adversarial Network And Its Application In Neural Style Transfer

Generative Adversarial Networks(GAN)have become a hot research topic in Deep Learning.GAN are composed of a generator denoted by G and a discriminator denoted by D.The training of both D and G uses unsupervised methods,and the performance will improve over time through adversarial training.Eventually,it reaches the point where D fail to distinguish the real data and the data generated by G.Therefore,GANs has been successfully applied in many fields including Image Processing,Computer Vision,Natural Language Processing,and speech recognition for its strong ability of data generation.This article focuses on the application of GANs in cross-domain Neural Style Transfer.The main contents are as follows:(1)We propose a Generative Adversarial Networks model(SSIM-GANs)which integrates the Structural Similarity Index Measurement(SSIM)and the Least Squares loss function.Since the generator learns the idea of Deep Residual Network(ResNet),SSIM-GANs is an end-to-end Res-StyleNet designed for cross-domain Neural Style Transfer.In addition,the Instance Normalization and Reflection Padding techniques are employed in the generator to eliminate artifacts for improving image quality.In order to solve the problems of instability in network training and the occurrence of Mode Collapse,this paper adopts the method of rebuilding the input data of the generator and utilizing the SSIM to measure the difference before and after reconstruction.(2)We propose the SSIM-GANs model for the task of Neural Style Transfer.Two groups of experiments are carried out with CUHK sketch-photo dataset and a Chinese ink-painted style dataset(named beihong-photo which is constructed by our lab),Experimental results show our method achieves a state-of-art performance,compared with the most popular algorithms including DualGAN,CycleGAN,Pix2 Pix and the classic GAN.To speed up the convergence,this paper applied the improved Adam optimization algorithm.