Research And Implementation Of Image Translation Based On Adversarial Learning

Through a lot of research and experiments,it is found that there are two problems in image translation: one is the inability to generate multi-modal and multi-class data,and the other is the inability to translate images based on context perception.Firstly,a semi-supervised multi-modal multi-class image translation is proposed to balance the relationship between sample labels and multi-class translation.In this work,a cross-domain auto-encoder is proposed to learn to decouple the content encoding that is invariant in the potential domain and the style encoding of the specific domain.Style code matches the prior distribution so that we can randomly sample from the prior distribution space and generate a series of meaningful samples.Through adversarial learning between domain classifier and category classifier,content code can be embedded in multi-class joint data distribution space,so that multi-class data can be generated simultaneously.Therefore,multi-modal and multi-class cross-domain images are generated by joint decoding of latent content code and sampling style code.Finally,the network is designed and tested.Semi-supervised experiments show that the proposed framework has the ability to generate high-quality and diverse images with fewer label samples.Further experiments under unsupervised conditions show that the network has advantages in learning decoupled representation and domain adaptation.Next,this paper proposes a context-aware image translation task to solve the problem that the current image translation work can not complete the translation of a specific region while maintaining the same region unrelated to the target of translation.Specifically,a new attention module is proposed to capture the relationship between various features in the context,so that only specific scene objects can be automatically concerned on unsupervised image translation.This module can be integrated into different image translation networks to improve their context-aware translation ability.In addition,there is a linear relationship between the computational cost of the module and the size of the image.Experiments on the dataset from day2 night also show that the module is insensitive to the increase of image resolution.