Research On Unsupervised Image-to-image Translation Methods For Natural Scenes

Image-to-image translation is a research hotspot in the field of computer vision,which can be used in art creation,video television production,multimedia applications and other fields.Image-to-image translation is a technique that changes the style of an input image while keeping its content unchanged.The traditional image-to-image translation algorithm has great limitations and is difficult to be extended to people’s daily life.With the rapid development of deep learning,the image-to-image translation algorithm based on deep learning has received extensive attention and has become a popular research direction.Among them,the image-to-image translation method based on the generative adversarial network has achieved a large number of successful cases in the field of image generation.Although the existing image-to-image translation algorithms have shown good style transfer performance,there are still problems such as insufficient detailed information of generated images,inaccurate styles and low diversity.Based on the above problems,this paper focuses on the study of natural scene image-to-image translation algorithm based on generative adversarial network.The main work is as follows:First of all,this paper proposes a natural scene image-to-image translation algorithm based on self-attention guided skip connections-NSIT-SASC,which is improved on the basis of the UNIT algorithm,combined with self-attention skip connections,adaptive instance layer normalization and orthogonal Jacobian regularization techniques.This paper introduces self-attention skip connections into the NSIT-SASC model to provide accurate and sufficient input image content information for image generation.At the same time,adaptive layer-instance normalization and orthogonal Jacobian regularization techniques are introduced to allow the model to adaptively select normalization techniques and improve its ability to disentangle style features,thereby improving the accuracy of image stylization.Experiments have proved that NSIT-SASC can achieve better results in the task of natural scene image style transfer.Second,this paper proposes an unsupervised multimodal natural scene image-to-image translation method-MDIT-SAOJ.In actual tasks,the image will show different styles due to the influence of time,weather,lighting and other factors.For the same input image,most image-to-image translation methods can only generate images of a single style.In response to this problem,this paper improves the DRIT++ algorithm and introduces orthogonal Jacobian regularization to improve the disentanglement ability of the model,so that the model can learn more accurate content and attribute features.At the same time,a self-attention module is introduced in the generator,which enables the model to model long-distance,multi-level dependencies across image domains.Experiments have proved that MDIT-SAOJ can achieve multi-modal natural scene image-to-image translation,and achieve better results than other multi-modal image-to-image translation algorithms.