Unsupervised Image Style Transfer Based On Generative Adversarial Networks

A painting can express the artist’s intention through color,line,shape and other aspects,but it is impossible to ask medieval artists such as Van Gogh and Monet to paint a modern city.However,with the powerful development of artificial intelligence,the image style transfer technology is used to synthesize stylized images that is similar in style to a target image and in semantic content to a source content image.The convolutional neural network of deep learning can be used to extract that the content and style features of the input image,and the corresponding stylized image is synthesized.The input of labeled or unlabeled data is treated as the difference between supervised and unsupervised neural networks.The pre-training of unsupervised network does not require the input of label data,and only analyzes the regularity of dataset.Based on the unsupervised Cycle GAN network,this thesis proposes the art style image migration model to reduce the noise texture,transition migration,image distortion and other problems in the generated image.In order to verify the validity of the proposed model,the proposed style transfer models were tested on relevant data sets.The work of this paper is as follows:(1)The Cycle-DPN-GAN model was proposed.Firstly,the positional normalization and dynamic moment shortcut(PONO-DMS)were introduced into the generator to better preserve and transmit the feature information extracted from the input image.Secondly,the loss function of the Multi-Scale-Structural Similarity Index(MS-SSIM)and the L1 regularization were combined to strengthen the constraints of image brightness,color contrast and structure of the reconstructed image.The results of Cycle-DPN-GAN model can better maintain the integrity of the internal structure information and strengthen the constraints on the color contrast of the image.(2)The AMS-Cycle GAN model was proposed.Firstly,the positional normalization and moment shortcut(PONO-MS)modules were inserted in the middle of the encoder and decoder to improve the stability of image feature information transmission.Secondly,MS-SSIM loss and the L1 regularization loss functions,which can effectively optimize the quality of the generated image.Finally,the attention mechanism module and spectral normalization were designed for the discriminator.The former improves the network performance and learns relevant information through the interdependence between the feature channels,thus effectively assisting the fine-tuning of the model,while the latter to improve the stability of training.The results generated by the AMS-Cycle GAN model significantly reduce the problems of noise texture,color distortion and boundary blurring.(3)In the improved model of Conv Ne Xt-Cycle GAN.Firstly,by replacing Res Net residuals in the generator with Conv Ne Xt residuals.The inverted bottleneck structure and the idea of grouping convolution were adopted in this model,which can reduce the parameter scale of the model and improve the overall performance of the model when the accuracy is improved slightly.Secondly,the same channel-based attention mechanism module as the AMS-Cycle GAN model was introduced into the discriminator,so that the model can pay more attention to useful features.The AMS-Cycle GAN model effectively preserves the contents of artistic images when they were transformed into natural images.