Design Of A Lightweight Multi Collection Style Transfer Platform

Art and painting have a long history,and oil painting,watercolor painting,and Chinese painting are all important components of human culture.With the rise of style transfer technology,the general public can convert images into artistic works easily,meeting individual creative needs greatly.This paper focuses on the current issues of style transfer and focuses on multi collection style transfer.A lightweight multi collection style transfer platform is designed and constructed,which enables people to convert styles of images and videos easily and quickly.It provides certain application research value in the field of digital media entertainment.Most existing style transfer models can only transfer one style at a time,and in order to transfer new styles,the model must be retrained.This paper focuses on the multi collection style transfer model and proposes a model MCGAN(Multi Collection Image Style Transfer for GAN)that can convert multiple artistic styles at once.This model is based on the CycleGAN model,which optimizes and adjusts its network structure by introducing additional conditional information as input and output for multiple collection in the generator and discriminator,giving the model the ability to transfer multiple styles simultaneously;The loss function is optimized and adjusted,and style category loss and style discretization loss function are added to make the model distinguish the differences between styles when training similar style datasets;Add saliency feature loss to improve the migration effect of significant regions in the image.The research results indicate that MCGAN has achieved the goal of transferring multiple styles from a single model and to some extent solved the problem of similar style homogenization,resulting in vastly different results for multiple styles after MCGAN transfer.The style features of images include both shallow and deep features.Most existing style transfer models overlook shallow features,which can easily lead to the loss of style information.In order to achieve better style transfer results,this paper focuses on optimizing feature extraction networks and adjusting the structure of the CycleGAN network to address this limitation.This paper introduces the CBAM(Convolutional Block Attention Module)hybrid attention mechanism before the style feature downsampling structure,making the network more focused on key features;Introducing FPN feature pyramid architecture before sampling structure on style features to enhance the network’s retention of shallow features and fusion with deep features.The research results indicate that MCGAN can not only generate images with excellent visual effects,but also has an average improvement of 31.593%,14.299%,and 5.992%compared to CycleGAN in FID(Frechet Inception Distance),PSNR(Peak Signal-to-Noise Ratio),and SSIM(Structural Similarity)indicators respectively.After most style transfer models are trained,the weight files are very large,which hinders model deployment on cloud servers or mobile devices.This paper focuses on lightweight deployment of models and optimizes MCGAN for lightweight optimization.In order to generate multi styles in real-time,this paper has optimized the network structure of MCGAN again,replacing some convolutional layers in the structure with deep separable convolutions.While retaining high-quality style transfer effects,the size of the model is also reduced greatly.The research results indicate that the size of the lightweight MCGAN_Light model has been reduced to the original 60.252%,and compared to the MCGAN model,the decrease in FID,PSNR,and SSIM indicators does not exceed 10.159%.