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

The purpose of image-to-image translation is to find the mapping between the input domain and output domain.According to the types of datasets,image-to-image translation based on generative adversarial network(GAN)can be divided into supervised learning and unsupervised learning.The unpaired dataset used in unsupervised learning for image-toimage translation is relatively easy to obtain in real-world scenarios,so it has great application prospects.GAN can achieve image translation tasks in the absence of paired images.However,GAN still has unsatisfactory generation results in the training process,and the traditional GAN may lead to no clear connection between input and output images.Therefore,this thesis proposes two GAN models with one-sided and two-sided structures.The idea of attention mechanism and contrastive learning is integrated into the model,and the image translation performance of the model is further improved in the Horse→Zebra,Cat→Dog datasets.The main work is as follows:(1)An unsupervised image-to-image translation model based on the attention mechanism is proposed.The attention mechanism is used to extract the effective information of the channel and spatial dimension,so that the model can focus on the part of the image that needs attention.According to the idea of contrastive learning in CUT,the Patch NCE loss is used in GAN to establish the connection between input and output images,and further improve the image translation ability of the model.(2)A robust two-sided GAN algorithm is proposed,which introduces the idea of contrastive learning into the two-sided GAN for efficient mapping between unpaired image pairs.The model uses a more high-powered attention mechanism to extract the effective information in the horizontal and vertical directions of the input features,which can obtain more effective spatial information in the task of image-to-image translation.Therefore the model can be more robust in different image translation tasks.