Image Translation Based Generative Adversarial Networks

Image translation is translating image in one domain to another domain.The image translation algorithm based on supervised learning is widely used in the fields such as super-resolution enhancement,image complementation and style transformation.Most of these algorithms use a deep convolutional network and achieved good results,but requires a large number of pairs of training data,which greatly limits the scope of its application.In practical image translation tasks,it is very difficult and costly to obtain the large number of pairs of images to train a model.Therefore,how to reduce the requirement of paired images and how to use non-paired images to enhance the performance of the model is a significant issue.In this paper,we propose an image translation framework based on semi-supervised learning.By referring to the recent advances in machine translation and image generation,two semi-supervised image translation algorithms are proposed.First of all,we propose a new generative adversarial network: multi-scale discrimination generative adversarial networks(MSD-GAN)and apply it to semi-supervised image translation.The algorithm use non-paired data to train the model by multiscale discriminator and reduce the need for paired data.In addition,we apply the idea of dual learning to semi-supervised image translation.We propose a new image translation algorithm that combines the generative adversarial networks with dual learning to exploit non-paired data to train the model,which greatly expands the algorithm application scenarios.Finally,we apply the algorithm in this paper to the face makeup-removal task.The algorithms was evaluated on multiple data sets and the experimental results show that our algorithm shows excellent performance which is comparable to that of supervised learning algorithms.