Research On Image Colorization Based On Weight Learning

Image colorization is a computer aided method to add colors for gray image or video.It is one of the important research topic in the field of digital image processing.According to the difference of color clues,the existing image coloring methods can be divided into two parts,scribble-based colorization and image-based colorization.In scribble-based method,user need draw color clues on gray image,and then the method will transfer the color from the color known position to the whole image.The weight information,representing the similar relationship between pixels,is important in the process of transferring color.It determines the energy in color transfer process.In order to get accurate weight,this paper puts forward a weight learning method,by training the relationship between pixels to get the optimal weights.This article firstly introduce the existing colorization methods according to the classification,scribble-based colorization,image-based colorization and machine learning based colorization,and then analysis some classical algorithm in this part.In scribble-based method,we introduce some affinity based algorithms,which inspire our work.The emphasis is laid on the proposed method.We then introduce the steps and experimental process of the method in detail.The experiments show that the proposed method can obtain better results.Specifically,the innovation of this article mainly includes the following aspects:(1)This paper put forward an affinity-based colorization method by learning weight.We establish a weight learning model between gray image and its color images and then get optimal weight by learning.Given a gray target image,we will get the optimal weight by the learning model,and then use the learned weight to transfer color.Compared with Levin's method,the results obtained with our method will get smaller difference from the original color image.(2)We hypothesize that,the relationship between pixels should be close in the graph if they share similar colors.Namely,the weight computed in the color image is the most accurate representation of relation between pixels.We calculate the color difference in the color image and feature difference in its corresponding gray image as the training set.(3)In order to characterize the relationship between adjacent pixels accurately,this method use not only intensity but the gradient feature of every pixel.Then use feature combination to build the relation between the gray image and color image.Experiments show that the feature combination of grayscale and gradient can produce better results.