Research And Application Of Image Migration Algorithm Based On Improved PatchMatch

In recent years,the computer hardware has developed rapidly,and its computing power has also been greatly improved.The major companies have provided open source deep learning platform for users.So computer graphics has made great progress in this context.Therefore,it is very promising to automatically convert pictures to artistic style pictures by using computers.The artistic style of the picture is also the direction of many scholars in the field of computer image for several decades.In 2017 there has been a lot of picture style algorithm.Scholars from Microsoft Asia Research Institute put forward the concept of Deep Image Analogy.They converted visual attributes to two different but semantically related images.Visual attributes include color,tone,texture,and style.For example,we need to transform the visual attributes of A and B' two pictures.A can be sketching or oil painting,B' can be a picture of real scenes,but A and B 'must describe the same semantic scenery.The core of the Patchmatch algorithm is to calculate the mapping between blocks in a given two picture,which is faster and better than some of the previous methods.But the basic algorithm can only find a nearest neighbour,and only in the same angle.In order to make the results better,the thesis extends the algorithm and applies it to the Deep Image Analogy.There are 3 main aspects of the main improvement:First,we hope to find multiple matching blocks in every block search.So this thesis decided to extend the core matching algorithm to find K neighbors instead of nearest neighbors.Second,Feature maps in different scale states may be needed during feature matching.The thesis extends the matching algorithm to find various angles of features.Third,the efficiency of the traditional two-dimensional search is low,and the matching of the features is low,so the thesis changes the mode of communication.The original way of propagation is to move the selected area in the direction of XY axis.After changing the way of propagation,its nodes connect to the k nearest neighbor list,and block the nearest neighbor list as the candidate block for the next area.The experimental results show that our method can better match the features and the image is more natural.