Research On Image-to-image Translation Based On Generative Adversarial Networks

Image-to-image translation has been widely used in real life and has been a research hotspot in the computer vision.It can be understood as changing particular aspects of the image meanwhile maintaining the basic structure and features of the original image.Traditional image translation algorithms often require a large amount of paired data for model training.In addition,many existing methods lack scalability and robustness when dealing with more than two domains transformation tasks.Therefore,after analyzing the existing image translation algorithms,this paper proposes a new model for cross-domain image-to-image translation.Recently,Generative Adversarial Networks(GANs)have aroused widespread attention in the academic community,and GANs have become one of the most mainstream technical means of image-to-image translation tasks.Inspired by the idea that deep neural networks learn the hierarchical feature representation,this paper proposes a GAN-based unsupervised transformation network(UTN-GAN)for unsupervised cross-domain image-to-image translation tasks.In UTN-GAN,we designed an auto-encoder reconstruction network to extract the hierarchical representation of the source domain image by minimizing the reconstruction loss.At the same time,in the two sets of GAN in UTN-GAN,the weights of the layers encoding and decoding high-level semantic information are also shared to ensure that the output image can maintain the basic structure and features of the input image.We conducted experiments on multiple image translation tasks and performed qualitative and quantitative experimental analysis with several advanced methods.The experimental results and discussions indicate that the proposed UTN-GAN algorithm is effective and competitive.Considering that many existing methods lack scalability and robustness when dealing with more than two domain transformation tasks.Therefore,based on UTN-GAN,this paper proposes a framework named UMT-GAN that can realize unsupervised multi-domain image translation.By combining modules with specific function,including encoder,reconstruction network,translation network and discrimination network,UMT-GAN can increase or decrease the corresponding translation network and discrimination network according to the target requirements.So that image-to-image translation in multiple domains can be performed simultaneously.And the UMT-GAN has good scalability.Through multiple sets of experiments,we verified the effect of UMT-GAN and also compared it with some advanced algorithms,indicating that UMT-GAN can perform stable training and can obtain competitive image translation results.In the style transfer,the two key points are how to effectively learn the style information of the image and how to effectively merge the style information into the semantic information.This paper proposes a multi-domain image style transfer network to transform images into a variety of different artistic styles.We designed an expert-style network to extract the style codes of the input images in different target domains,containing the unique information of the respective domains through a set of bidirectional reconstruction losses.At the same time,a transfer network is designed,which recombined the extracted style codes and the semantic information of the source domain images to generate a new image.With the adaptive instance normalization,the transfer network realizes the style transfer from the source domain to the target domain.In the test phase,the source domain images and the target domain images are directly input into the content encoder and the style encoder,and the target results can be obtained through the transfer network.Experiments show that the model can effectively combine the content of any photo with the style of many artworks to produce new images.