| Style transfer is an important branch of image fusion.Style transfer is defined as collecting all important information from multiple images and concluding it into fewer images(usually concluded on one image).This image is richer and more accurate than any source image and contains all the necessary information,including content information and style information.For style image,the traditional style transfer algorithm first manually establishes a mathematical or statistical model,and then changes the image or video frame to fit the model better.This method not only spend extra resources,and there are a lot of shortcomings,is basically,an algorithm for a kind of style or a kind of scene,and the style transfer result especially on video style is often unsatisfactory,so the traditional style transfer algorithm is very limited.With the rise of artificial intelligence,especially deep learning,researchers have begun to combine neural networks to explore new style representation and transfer algorithms.Based on the deep study of deep learning and the research results of current image fusion algorithms,this paper proposes several improved style transfer models based on deep learning.The specific research contents are as follows:(1)This paper proposes style transfer model based on Generative Adversarial Networks(GAN).Based on the current mainstream methods used in L1 loss function to constrain the content between generated image and the input image,pre-trained model(VGG)is introduced to constrain deep feature.Based on this,the proposed model reduces the local deformation and artifacts better.Then,the proposed model is further implemented on color facial images to sketch pictures style,Besides,based on spectral normalized constraints generated against network gradient descent.To achieve the purpose of the stability of training,this paper analyzes the different model style for color facial images to sketch images in the process of style transfer,this paper put forward the Generative Adversarial Network with multiple discriminators,and using the guided filter will desperate the characteristics of high frequency and low frequency characters of the color image.Iterative optimization was carried out on high frequency features and low frequency features.(2)Further,this paper uses the depth-separable structure and channel-wise convolution in the process of training to model the lightweight generator,it reduces extra computational costs sharply,and the style transfer on face-to-cartoon has achieved good result.By applying the proposed high and low frequency discriminators,the model in this paper has a Frechet Inception Distance(FID)reduction of approximately 18%on paired images and 6%on unpaired images compared to the most advanced methods in the term of generating sketch images.The proposed method can even generate a sketch that is more pleasing to the eye than the artist.In the term of Structural Similarity(SSIM)and Peak Signal to Noise Ratio(PNSR),the proposed model also obtains good results.In order to accelerate the convergence of the model,the adaptive normalization term is further proposed.A large number of experiments were conducted between the CUHK Student Dataset,CUHK Face Sketch and Yellow-Stylegan2 dataset and the proposed method is verified.To sum up,this paper proposes three network models based on deep learning,and proves their practical application value through a large number of experiments respectively. |
