| In recent years,a large number of rapidly developing and practical technologies have emerged with the rapid development of artificial intelligence.Among them,data generation,as one of the key technologies in the development of artificial intelligence,has achieved extraordinary results in the application of images,text,speech and other fields.Research and development of generative models has become one of the important ways to achieve strong artificial intelligence.The data generation method applied to the field of image translation can not only give full play to the imagination of human artistic creation,but also greatly improve the quality of image translation and expand its application scope.This paper mainly studies the related issues of data generation models and applications in the field of image translation.First of all,the network structure of the data generation model in the current image translation task is complex,and the problem of poor style information transmission generally exists.Therefore,this paper proposes a multimodal unsupervised image translation network with a shared encoder(Sunit),which improves the style transfer ability by reusing the encoding ability of the discriminator into the style encoder,and removes the network redundancy,significantly reducing the amount of network parameters.In addition,this paper also designs a novel training strategy: the style encoder only uses the style reconstruction loss,and does not follow the generator for training,so that the training objective of the style encoder is clearer and the training is more effective.Second,in terms of artwork creation,existing generative networks are still blank in the field of multi-domain landscape style transfer.Therefore,this paper extends the research of image translation network on multi-domain and multi-modal landscape translation tasks,and proposes a two-stage landscape style transfer model(TMGAN).TMGAN imitates the creative process of human artists and decomposes the landscape style transfer process into:extracting content and injecting style,so as to better complete the mutual conversion of landscape paintings,oil paintings and realistic landscape photos.In addition,this paper produced a high-quality multi-domain landscape translation dataset,and innovatively designed subjective and objective evaluation experiments from two aspects of content and style similarity,which filled the gap in the evaluation index of style transfer ability.The Sunit network proposed in this paper is compact in structure,efficient in training,and improves the quality of multi-modal generated images,providing a direction for further optimization of multi-modal unsupervised generative network models.The multi-domain landscape style transfer model TMGAN proposed in this paper enriches the way of art creation,and innovatively proposes subjective and objective evaluation method,which can promote the further development of artificial intelligence technology in the field of art creation and art evaluation. |
