Research On Unsupervised Image-to-Image Translation With Generative Adversarial Networks

Image-to-Image Translation is the task of translating one possible representation of a scene into another,which is one of the research hotspots in the field of computer image processing and its mainly application is in transportation,medical care and daily life etc.This thesis studies and improves the unsupervised image translation method based on Generative Adversarial Networks to improve the content consistency and image domain consistency of generated images and input images,and enhance the clarity and authenticity of translated images.The specific contents are as follows:(1)Study unsupervised image translation in single image domain.In order to solve the problem of unsupervised image translation in single image domain based on Generative Adversarial Networks,which is caused by unstable training process and large change of irrelevant domain,this thesis based on dual learning proposes an unsupervised image-to-image translation method with self-attention and relativistic discriminator adversarial networks.In the generator of the model,self-attention mechanism is designed to build long-short-range dependency for image generation tasks.Skip-connection between low and high convolution layers help reduce the loss of feature information in irrelevant image domain.In the discriminator,spectral normalization is used to prevent the gradient disappearing caused by the mutation of the discrimination ability to enhance training stability.In the loss function,the selfreconstruction consistency is added on the basis of loop reconstruction to focus on target image domain change.The relativistic adversarial loss is designed to guide the zero-sum game between generator and discriminator.Experiments verify that this method improves the quality of the translated image.(2)Study unsupervised image translation in multiple image domain.Aiming at the problem that the discrete form labels in the multi-image domain unsupervised image translation based on Generative Adversarial Networks result in the single diversity of image domain expression and the poor stability of the discriminator,an unsupervised image translation method based on fusion category label and dual-scale discrimination is proposed.The domain controller is added to the model generator to generate the target image domain style information,and the adaptive instance normalization is combined with the original image feature information to guide the image translation of the multiple image domain,and the translation image style diversity is increased;The classification structure is added,the image domain category is identified,and the two discriminators are used to verify the authenticity and category of the original image and the scaled image respectively to enhance the stability and improve the quality of the translated image.In the loss function,the classification loss is used to establish the relationship between the input label and the output image.The hinge relativistic adversarial loss is designed to guide generator and discriminator to complete the multiimage domain unsupervised image translation.Experiments show that this method enhances the diversity of image domain expression and translation effects.