Image Style Transfer Processing Based On Deep Learning

In the field of image processing,image style transfer is a kind of technology that processing image information such as the color,silhouette,and line to change image effect by computer.In recent years,with the development of machine learning,image style transfer processing based on neural network obtained good effect.Based on the deep neural network and research of the existing network,this thesis designs two networks to generate highly stylized effect pictures.The main work of this thesis consists of three aspects as follows:Firstly,this thesis designs a style similarity measure model,which can discriminate whether the effect of the generated image was similar to the real type of style image.Generally,the current image style transfer processing results can only use visual effect evaluation,and it's hard to be objective to compare whether the stylized effect of the generated image is similar to the target style.This thesis uses convolutional neural network to build a deep network that can objectively test whether the stylized image is similar to the real style,and get the quantitative result based on the test.Secondly,based on the "Real-Time Style Transfer" model,this thesis joins a classifier to the original network to put forward a new model--RTST model with similarity losses network,which could solve the original network's problem.In dealing style with low contrast or unclear lines,the original generated image appears dark color block.Joining the classifier as the similarity losses network to the convolutional neural network's training process can influence the overall network optimization process,and guides image transfer network's optimization process in the direction of more realistic style,then get a better style transfer network.Thirdly,based on the RTST model with similarity losses network,this thesis imports the idea of generative adversarial networks,which improves network's performance.In this model,the classifier is as a discrimination network,image transfer network is as a generative network.Two kinds of network training in turn to improve their performance and get a better trained generative network to make the generated stylized image has better visual effect.