Research On Multi-Domain Image Style Migration Algorithm Based On Improved Generative Adversarial Networks

The field of image style transfer technology has gained popularity in recent years,becoming a prominent area of research.Image style transfer,as the name suggests,utilizes computer-related techniques to extract content and style-related semantics from images and combines them to generate a new image.This technology has a wide range of applications,including film special effects,post-production synthesis,artistic creation,and animation rendering,while also providing ideas and data sources for image generation and dataset creation.Research on style transfer technology is beneficial for the integration of technology and art and has broad market value and application scenarios.In the field of computer vision,there are two main types of image style transfer techniques: traditional methods and deep learning-based methods.Each method has its unique characteristics and advantages.Traditional methods are relatively simple to implement,but deep learning-based methods have superior transfer effects.Traditional image style transfer techniques focus on extracting shallow semantic information,resulting in generated images that fall short of the desired state and are difficult to use.To meet the increasing demand for improved image transfer technology,it is crucial to optimize and improve style transfer techniques.Therefore,in this thesis,we optimized the image style transfer algorithm based on Cycle GAN using deep learning methods to enhance its performance.The main contributions of this thesis are as follows:1.To more finely control the image style transfer process,we introduce a new method that incorporates the generator encoding part into the content and style encoders.The content encoder,based on the variational autoencoder network structure,aims to extract the content information of the original image.Meanwhile,the style encoder introduces self-attention mechanism,which can focus more on the style information of the original image,thereby improving feature extraction accuracy.2.By introducing adaptive instance normalization(Ada IN),the extracted content encoding and style encoding are combined through this module,to control the weight ratio of content features and style features and thus control the decoding effect of image generation,speeding up the stylization effect of the image.3.We implement image reconstruction operations using bilinear interpolation,which can significantly enhance the accuracy and precision of image style transfer while effectively reducing or eliminating the checkerboard effect caused by image discretization.The discriminator uses the Markov discriminator(Patch GAN)to maintain image details and performs comparative experiments on sketch and comicstyle datasets.4.To address the issue of unclear image style transfer,we further optimized the model on the basis of the improved model.We deepen the network layers,replace traditional convolution with dilated convolution,introduce a residual attention mechanism fusion module to the style encoder,and use a multiscale discriminator to extract image features at multiple scales.We also introduce a mapping model U,which makes the generated image and the source domain image have some substantial connection.We conduct comparative experiments on the artistic style dataset and obtain images with a more pronounced target style,verifying the effectiveness and extensibility of the model in style transfer tasks.