Research And Implementation Of Gans In Image Style Conversion

With the development of the Internet,the need for image diversification processing is increasing.At the same time,image style conversion has gradually become a research hotspot which has a wide range of research significance and application value.In recent years,traditional image style conversion techniques based on deep convolutional neural networks have encountered huge challenges.But,the proposal of generative confrontation network provides a new idea for solving the problem in image style conversion.In addition,in terms of the resolution of image generation and the authenticity of samples,generative confrontation network has achieved better results.Therefore,image style conversion based on Generative Adversarial Networks has become a current research hotspot.In the paper,from the perspective of unsupervised image style conversion,analyzed and studied the CycleGAN model proposed by other researchers.And using attention to improve CycleGAN conversion model.The paper mainly includes the following work:1.The biggest problem with CycleGAN model is that it not suitable for image style conversion tasks with large differences in shape and texture.In this paper,the attention is combined with the original network.At the same time,classifier is used to focus on the source domain and the target domain of the image,the attention is used to help the model control where to carry out the larger transformation.In the end,experiments show that the effect of style conversion in this article is better than the original CycleGAN model on the horse2zebra dataset and selfie2anime dataset.2.In this paper,adaptive instance normalization method is used in the Dncoding block of the generator.By using this method,the ratio between instance normalization(IN)and layer normalization(LN)can depending on the importance of detailed and global information.3.This paper designs and implements a prototype system based on the improved CycleGAN model.The image style conversion system is built using the idea of micro-service and Docker container technology.The system adopts the design of front and rear end separation.In terms of function,it is mainly divided into three modules:screenshot and upload module,background service module and algorithm module.The screenshot and upload module is responsible for obtaining pictures and uploading pictures to the server.The background services module is mainly used for processing the requests from front end.The algorithm module is for image style conversion.In the thesis,the main work mainly consists of the following.On the one hand,the CycleGAN network is improved and optimized.By using the attention to improve the CycleGAN.Experiments show that the improved network has better style conversion effect than the original network.On the other hand,an image style conversion system based on the improved algorithm is designed and implemented.The system adopts the idea of micro-service and containerization technology,which greatly improves the flexibility and expansibility of the system.