Research On Unsupervised Image Style Transfer Based On Generative Adversarial Network

With the advent of the mobile Internet,people’s demand for image content diversification is getting higher and higher.As an important image content production technology,style transfer has been greatly developed in recent years.Unsupervised style transfer has a stronger application prospect because it does not require the characteristics of paired datasets,and has become the mainstream of current style transfer research.But in general,unsupervised style transfer is lacking in the quality of generated maps compared to supervised style transfer.During the research process,it was found that the realization of the unsupervised characteristic of the classic unsupervised style transfer network is done through the cycle consistency loss.However,Cycle GAN,Dual GAN,and Disco GAN based on cycle consistency loss generally have problems such as texture disorder,migration matching errors,and poor style and artistic conception simulation when dealing with style transfer tasks,and cannot simulate the texture and strokes of artistic paintings well.ACL-GAN network,which focuses on unilateral transfer,has the problem of color distortion and blurred images.To solve the above problems,the main work of this thesis has the following four aspects:(1)Aiming at the problem of cyclical consistency loss that results in messy textures and poor simulation of style and artistic conception caused by equal mapping of two data distribution domains,this paper proposes the CCME-GAN model based on the idea of strengthening the effect of unilateral transfer.First of all,in the design of the network architecture,a multi-level evaluation network architecture is proposed based on the threelayer semantic information of the image,and the migration effect from the source domain to the target domain is strengthened by reconstructing the organizational structure relationship between the networks;Secondly,in the improvement of the loss function,a multi-scale adversarial loss and a cycle correction loss are proposed to guide the iterative optimization direction of the model with more stringent goals and generate pictures with better visual quality;Thirdly,from the perspective of strengthening the ability to extract important features,a mixed attention mechanism was introduced in the downsampling stage of the generative network,which realized the enhanced extraction of important features and the suppression of secondary features;Finally,the ACON activation function is introduced in each stage of the network to improve the nonlinear expression ability of the network and realize the enhancement of the network expression ability.(2)In order to further fit the distribution distance between the generated data domain and the real data domain and achieve a better style simulation effect,this paper proposes the CCME-WGAN model from the perspective of improving the distribution distance metric.Firstly,the objective function is reconstructed by using the Wasserstein distance instead of the original JS divergence distance measure,and based on the reconstructed objective function,a joint Loss function suitable for unsupervised dual domain mapping is further proposed,which improves the upper limit of network to fit the data distribution ability;Secondly,in order to further enhance the migration effect from the source domain to the target domain and improve the quality of the generated graph,from the perspective of differential generation network expression ability,the multi-scale feature extraction ability of the source domain to target domain generation network is strengthened through skip connection;Finally,in order to improve the comprehensive judgment ability of the discriminant network,the receptive field of the probability matrix is expanded by using dilated convolution,so that it can refer to more contextual information during the iteration process,and realize the probability judgment carrying more receptive field information.