Unsupervised Multi-domain Image-to-image Translation Method Based On Domain Label Self-learning

The rapid development of software and hardware and the rise of deep learning have promoted the rapid development of the field of computer vision.Image translation is an important research direction in this field.Its purpose is to learn a mapping between input images and output images,and it also has a milestone development.Through the method of image translation,the image can be changed in any style and content.There are large application scenarios in common super-resolution reconstruction,image colorization,image filling,and image translation,so it has become a hot research direction in the field of computer vision.However,the existing research is mostly in the direction of supervised,unsupervised and multi-modal single-domain models,and it is very difficult to extend to complex redundant image translation application scenarios.The most common is the exponential increase in the number of model parameters.At the same time,there will be the problem of unstable model training.There are various problems in the field of multi-domain translation.This paper proposes a new domain label self-learning unsupervised multi-domain image translation method.For multi-domain image translation,this paper designs an independent domain label for each domain,and through neural network and carefully designed domain representation vector loss,the domain label is learned automatically without additional information.In terms of model architecture,the method proposed in this paper has also been streamlined,using only a pair of generators,discriminators and N encoders,which is more concise and efficient than existing multi-domain algorithms.Through experimental comparison and analysis,the method proposed in this paper is generally superior to the latest multi-domain methods in terms of qualitative and quantitative indicators on multiple data sets,which can fully demonstrate the correctness of the proposed method.