Image Style Transfer Based On Dual Learning

Image style transfer is a challenging problem,especially for supervision methods.Because an image with a specific style often does not have the corresponding source image(such as a painting),so it is very difficult,laborious and time-consuming to collect such a large amount of data,and the price of manual labeling data is very high.Therefore,it is always a hard task how to achieve an unsupervised image stylization.Because the existing learning paradigm is difficult to use unlabeled data for learning,in order to solve this problem,relevant researchers have proposed a new method: dual learning.It differs from other learning paradigms where it utilizes the dual characteristics of AI tasks to learn,and obtains feedback information in the learning process,so these tasks can promote,learn and enhance each other.So far,dual learning has been proposed for three years.During this period,it has been widely used,such as text understanding,image expression and speech recognition.It can also be combined with auto-encoder,GANs and so on.This paper introduces dual learning to deal with the problem of unsupervised image stylization,and a kind of image style transfer methods based on dual learning is proposed.The main innovations are as follows.(1)An unsupervised method for global image style transfer is proposed.The core of this method is a dual learning framework,which consists of two tasks: the main task and the dual task.The main task completes the generation of stylized images and the dual task realizes the restoration from stylized image to content image.In order to calculate the losses in the two tasks,a deep convolution network VGG19 is used to obtain the content and style representation of images.The proposed dual learning algorithm is tested on some images collected from public sources,and it is compared with other methods visually.The experimental results show that the dual learning algorithm can get visually better stylized images.(2)A dual learning method is proposed to manipulate local facial attributes with deep convolution network VGG19.Different from the existing face generation methods,the proposed method realizes face generation according to a source face image and a given local face attribute value.The generated image can not only keep the identity of the source face image unchanged,but also have the given face attribute.The dual learning method can make the generated image's quality better.The proposed method is evaluated on an open face dataset and compared with other methods in terms of visual quality and specific index.The comparison results show that our method can get the images with good quality.