Image-to-image Translation With Generative Adversarial Networks

As an information carrier,image is an indispensable way to express emotions and communicate in modern life.Similar to machine translation in natural language processing,image-to-image translation not only need to generate dual images in the target domain according to the images from source domain,but also need to constrain the generated images remaining similar to images from source domain during the translation.Whether the supervised translation algorithms with aligned training pairs or unsupervised translation algorithms with unpaired images has implemented image-to-image translation between two domains.However,in multi-domain image translation tasks,existing models need to be trained separately between two domains of all each time.This training method is quite cumbersome and time-consuming.This thesis proposes two solutions for image-to-image translations for multiple domains to reduce the numbers of training models and improve the image qualities of translation results.Firstly,for the tasks on facial attribute transfer,an unpaired image-to-image translation based on conditional projection in this thesis.This method divides the dataset into different domains according to the different attribute values of the face,and marks them with labels.Then,it trains a discriminator to discern the differences between the feature learned from the generated model and the target feature,and calculates the similarity between the feature information learned by generator and conditional information.It alleviates the problems of overfitting in the generator and facilitates the generator to generate more realistic and natural images.It also improves the accuracy of translations.Secondly,referring to the conditional generation model,this thesis proposes a cross-domain image-to-image translation based on disentangled representation.This algorithm uses a domain label to represent domain imformation,and combines domain labels which is a conditional information and disentangled representation.Then it can control the domain label and translate the image into any desired target domain.The use of domain labels can reduce the number of training models.This algorithm can perform image-to-image translations for multiple domains using only a single model,and reduce the training time.In addition,the input data contains images with different labels,which substantially expands the amount of training data,so that the generator can learn more features.The model also introduces random variables during the training process,which make the generated results diverse.The effectiveness of the proposed methods is verified by comparison with existing translation and experiments on multiple datasets.