Style Transfer Based On GAN

Generative Adversarial Network(GAN)has achieved great performance in several image generation,image falsification and image manipulation tasks.Image-to-image transfer is obviously a trend in the field of computer vision.However,it becomes challenging for some major drawbacks,such as the time-and computation-consuming training process,mode-collapse and lack of paired training data.To handle the above-mentioned limitations,we propose a novel model,called Enhanced Cycle conditional GAN(Enhanced CCGAN).Our model alleviates the problem of the lack of the aligned paired data problem by calculating the cycle consistency loss function.It realizes the representation disentanglement by using content encoder and style encoder with different architecture.In the content encoder,we use the ResNet block to extract the content feature to realize the function of multi-level feature fusion.We propose a semanticIn view of the limitations of the above GAN network,we propose a new model,which we call enhanced cyclic conditional GAN network.(1)To improve the quality of image label training style,a style encoder based on convolutional neural network,a style encoder based on multi-layer perceptron and a content encoder based on convolutional neural network are proposed respectively.Furthermore,the style migration model based on the three proposed encoders is given.Simulation results show that compared with cyclegan,stargan and unit models,the proposed model improves the training efficiency of the model and the quality of style transfer image generation;(2)To improve the accuracy of style tensor extraction in style encoder,a latent semantic loss function based on style tensor space is proposed.Furthermore,a style transfer model based on confrontation generation network using the loss function is given.Simulation results show that compared with cyclegan,stargan and unit models,the proposed model reduces the migration loss of image style from target domain to source domain.(3)Aiming at the problem of strengthening the ability of style encoder to extract the global features of image,a MLP model style encoder training method based on AdaBoost training idea is proposed.Further,the MLP model style encoder using this training method is given.Simulation results show that compared with the style encoder composed of convolutional network,the proposed model reduces the migration loss of image style from target domain to source domain.Experimental results have shown that our model can produce images with high quality and diversity across several data domains and significantly outperform the state-of-the-art models.