Multi-domain Image-to-image Translation With Generative Adversarial Networks

Image translation refers to mapping the input image from the source domain to the target domain so that it has the specific attributes of the target domain.The image attributes can be defined according to the basic characteristics of the image or the specific application scene and target object in the image.For example,the basic attributes such as the resolution and color of the image can be defined,and the season,weather attributes of the scene image,the expression,age attributes of the face image can also be defined.Image translation has a wide range of applications in many image processing tasks,including image style transfer,image super-resolution,face attribute editing,and so on.The multi-domain image-to-image translation problem to be solved in this paper is to realize the translation of an image from the source domain to any other domain for a given multiple image domains.The generated image must not only obtain the attributes of the target domain,but also retain the original image's attribute-independentdetails.It is mainly used in the translation of scene images in multiple weather and season domains to achieve tasks including image de-noise,image batch editing,and virtual scene simulation.Generative Adversarial Networks(GAN)is based on the idea of zero-sum game.Through the process of mutual confrontation between generator and discriminator,it assists the generator to learn the distribution of real samples and has a prominent performance in image generation.Therefore,most methods of image-to-image translation are currently implemented based on GANs.This paper proposes a multi-domain image-to-image translation model based on generative adversarial networks,and realize supervised and unsupervised method with different dataset.The main innovations are as follows:(1)After preprocessing to obtain a batch of weather dataset in pairs,a multi-domain image-to-image translation method based on paired dataset is proposed.The model is based on the Att GAN network architecture.The squeeze-and-excitation block is added to the generator to assign weights to the feature maps to make full use of the effective intermediate feature maps.The pre-trained VGG19 is used to extract the high-dimensional spatial features of the true and false samples,and the perceptual loss is calculated and combined with the L1 loss to further constrain the generated space.Alarge number of comparative experiments have proved that these two improvements can effectively improve the quality of the generated images,and achieve the best results in the application of bad weather to good weather translation.(2)A multi-domain image-to-image translation method based on unpaired dataset is proposed.The model is based on generative adversarial networks,combined with the domain coding mapping and the multiply discriminators methods in Single GAN.The main innovation is to add a spatial feature transform layer to the generator,and use the segmentation semantic information of the image to guide the image generation process.By training the model on unpaired weather dataset related to traffic scenes and comparing with other similar models,it proves that our method can generate high-quality and diverse results.At the same time,the model is applied to the season and transient-attributes dataset,and get the conclusion that the model is extensible.