Research On Image Style Transfer Method Based On Generative Adversarial Networks

The distinction between image art style and target-specific attributes is a visual task that can be performed through abstract knowledge,and through this knowledge can complete the process of more complex image re-creation.Image style transfer is a challenging topic in computer vision,and there are many difficulties in problem definition and implementation,like how to define and represent the style in a mathematical approach,how to modify the style of the image.In recent years,the development of deep learning methods has promoted various breakthroughs in the field of image style transfer,with the deep learning method guiding the model to discriminate and modify the image style.Many of existing work is based on the loss function of image style or content representation,in contrast,this paper proposed an improved general image generator for image style transfer tasks based on these research.The main work of the thesis includes the following:1.Taking the abstract representation of the deep learning feature as the start-ing point,the related characteristics of image features and semantic con-tent in convolutional neural networks are studied.And how to modify the image content by image feature interpolation;2.Based on the connection between image deep feature and semantic con-tent,a deep feature transfer method is proposed,which can be used as a general image generator structure for image style transfer tasks.The experimental results shows that the CycleGAN improved by our method performs better in fidelity and diversity,as well as the acceptance in hu-man study.3.Realized a deep feature transfer model with faster convergence.We show the generalization ability of our method by the results on other kind of style transfer datasets.