Design And Implementation Of Arbitrary Style Transfer Methods

Arbitrary style transfer aims to render the content image according to reference style image to make the content image own the style features such as color,stroke,and texture while preserving its original semantic structure.Thanks to the self-learning and flexibility of neural networks for extracting image features,neural style transfer methods have been developed rapidly.However,the existing arbitrary style transfer methods still have some problems,for example,unremarkable style representation and distorted content structure in the generated images.In this thesis,to solve the above problems and achieve semantic-aware style transfer,we introduce the attention mechanism and Transformer network.The main research works are as follows:1.To address the problems of insufficient style rendering caused by the global adjustment methods and distortion of structure caused by the local replacement methods,we proposed an arbitrary style transfer method with the sketching and refining operations.In the sketching process,we require sketch features by adjusting the statistics of content features globally.In the refining process,we use the sketch features to guide the replacement of local features.The proposed method renders the content image from coarse to fine by the sketching and refining operations.We prove the effectiveness of the proposed method by comparing it with other stateof-the-art arbitrary style transfer methods.Experiments show that the proposed method can better balance the content structure preservation and style stroke performance in the generated images.2.To address the problem of poor style abstraction of existing style transfer methods using statistical information to represent style features,we proposed a Transformer-based arbitrary style transfer network.In the process of building the network,we use the residual structure to bring in the low-level texture information of images.Also,we use Transformer to enhance the richness of feature information.Finally,we introduce a discriminant network to reduce the distance between the generated image and the real style image.The experimental results show that the method proposed in this thesis provides a reasonable abstraction of the content structure,reduces the generation of checkerboard texture in the image,and improves the smoothness of the generated image.3.We design and build an arbitrary style transfer system to explore the feasibility of bringing deep learning models into practical applications.The system uses the model trained by the arbitrary style transfer method with the sketching and refining operations as its back-end support.The system provides a front-end UI interface to simplify the user interaction process,meanwhile,it also offers three different stylization tasks to give users a better feeling of use,for example,stylization,stylization control,and style interpolation.The artistic images generated by the network are displayed directly in the front-end interface,which is convenient for users to observe the effect.