Deep Learning Techniques For Image Style Transfer And Destyle Transfer

Style transfer is a computer vision task that aims to apply the style of one image to another one.Thanks to the recent advance in deep learning,convolutional neural networks have been able to succesfully extract style and content features from an input image.However,current techniques do not fully exploit semantic information and results are not satisfying.We first proposed a semantic style transfer algorithm that can process content images with more than two semantic regions.Given segmentation masks,our algorithm allows the user to select which style to apply to the different regions of the input image.Our algorithm solves an iterative process and separates style and content features using a pre-trained VGG.To control the style transfer spatially,we mask the style features with the segmentation masks and we cap the gradient during optimization.We also simplify the parameters tuning problem with an auto tuning scheme.We affect smaller weights to smaller regions and we ensure that each style is equally represented.We finally improve the quality and the speed of style transfer with a multi scale representation.We empirically found that using two scale is optimal.Compared to our baselines,our algorithm is 12% faster and generated images do not present ghosting artifacts.As a second contribution,we developed a face destylisation algorithm.Given stylized faces,our algorithm can recover photorealistic faces.Turning paintings into photo is difficult and current style transfer techniques perform poorly on this task.In parallel,generative adversarial networks,commonly refered as GANs,can generate highly realistic images.Our proposed algorithm combines the benefits of a conditional GANs with perceptual losses.Our generator removes style features using an auto-encoder with skip connections.Unlike style transfer,we use a markovian discriminator to represent the style loss.Our work focus on faces destylisation.To train our model,we build a high quality dataset composed of 108 K pair of stylized and non stylized faces.We outperform previous methods both qualitatively and quantitavely.Our model reduces the FID and LPIPS distance by 38% and 39% respectively.