Research On Aesthetic Image Synthesis Based On Generative Adversarial Network

As a carrier of a large amount of information,images play an irreplaceable role in various fields.With the improvement of living standards,people pay more attention to the aesthetic effect of images,and tend to choose images with aesthetic appeal.Image aesthetics can fit the human visual system and construct images of high aesthetic quality,which is a challenging subject in the field of image research.In this paper,we use the rules of adversarial learning to explore the generation of aesthetic images,and complete the related research on automatically generating images with aesthetic attributes and image aesthetic enhancement.The main work of this thesis is as follows:1.We propose a deep learning model which automatically generates aesthetic images.Most of the existing Generative Adversarial Networks are based on simple datasets training.The generated images are visually plausible in terms of semantics and aesthetics,which does not meet people's aesthetic needs.In order to improve the aesthetic effect of the generated image,the work is based on the advanced network DCGAN model,which introduces the aesthetic loss and the visual content loss,in which the aesthetic loss attempts to maximize the visual aesthetic of the image,while the visual content loss minimizes the similarity between the generated images and real images in terms of highlevel visual contents.In the experiment,the work was carried out on two public datasets,and qualitative and quantitative results demonstrated the effectiveness of the two loss.2.We improve a deep learning model that enhances the aesthetic quality of images.The popularity of smart tools like mobile has led to an explosive growth of images,but due to factors such as shooting methods and shooting tools,the images which users upload and shared are not aesthetically pleasing.In order to enhance the aesthetic effect of images,this work explores an end-to-end method that can automatically improve the aesthetic quality of the image.The inverse mapping network of the enhanced network reduces the limitation on the aligned photo pairs.What's more,it includes effective loss functions.Among them,color and texture loss is learned in an adversarial fashion,and the total variation loss enhances the smoothness of generated image.These loss functions improve the aesthetic quality of the image to some extent.Based on this,we improve the content loss.Extensive experiments on public datasets demonstrate the effectiveness of the method.