Research On Image Style Transfer Method Based On Generative Adversarial Network

Image style transfer is a research direction belonging to computer vision,which has a wide range of applications in many fields,such as digital art,defect detection,face recognition,medical diagnosis,etc.However,traditional image style transfer methods have many problems: training requires a large number of paired images,the quality of generated images is low,and the time required to generate images is too long.Generative confrontation network is a popular technology in the field of deep learning in recent years.Because of its confrontational learning method,it has been widely used in the field of image style transfer.The CycleGAN model is an image style transfer model based on the generative confrontation network,which can realize the style transfer between images by learning the mapping relationship between two different style image domains without paired images.Image style transfer models for domains.But there are still some problems to be solved in this model.Improving the CycleGAN model can improve its performance and stability in image style transfer tasks,making it more reliable and practical in practical applications.Therefore,this thesis takes the CycleGAN model as the research focus,aiming to analyze and understand the CycleGAN model and improve it.(1)The generator network of the Res Net structure adopted by the CycleGAN model is prone to background color distortion when generating images,and the generation effect on complex textures is poor.In response to these problems,this thesis proposes a new generator network that combines the Convnext convolutional structure and the U-Net structure.In this new type of generator network,the Convnext convolutional structure replaces the convolutional structure in the feature extraction part of the traditional U-Net structure.The U-Net structure can solve the problem of background color distortion caused by the loss of underlying information in the original generator network.The Convnext convolution structure has stronger feature extraction capabilities and can better extract complex texture features.Combining the two The latter generator network can effectively solve the limitations of the generator network using the Res Net structure.(2)Aiming at the limitations of the receptive domain mechanism in the discriminator network used in the original CycleGAN model,a new discriminator network that introduces the CBAM attention mechanism is proposed.The introduction of the CBAM attention mechanism can enable the discriminator network to give different attention to different regions,helping the entire model to better learn image features.And on the CycleGAN model that introduced the CBAM attention mechanism,the face image mask wearing task was carried out,proving that the CycleGAN model using the new discriminator network has better image style transfer capabilities.(3)Experimental comparisons were carried out on multiple common datasets.Compared with the original CycleGAN model,the migrated images generated by the improved CycleGAN model with a new generator network and a new discriminator network proposed in this thesis have a Frechet Inception distance(FID)Structural Similarity(SSIM)has a better performance.And finally,based on the improved CycleGAN model proposed in this thesis,the daytime traffic map is converted into the corresponding night road map,and the application level of the improved CycleGAN model proposed in this thesis is expanded.