Image Matting Based On Active Learning

At present,the image matting algorithm has more or less some problems,which are mainly reflected in following points.First of all,it is difficult to manually mark the correct foreground and background information for the hair and edge regions in the image;Secondly,it takes a lot of interaction to achieve satisfactory results;Lastly,it takes a long time for the results of image matting to be fed back to the user.Spectral matting algorithm divides the picture into several matting components for labeling,which better overcomes the difficulty of manually labeling hair and edge areas.However,when the number of matting components becomes larger,the cost of manual labeling is still large.The goal of active learning is to reduce the overall cost of labeling by selecting the most valuable samples for labeling.In this thesis,active learning is introduced into image matting for the first time.It is expected that the desired matting result can be obtained only by selecting a small number of region(matting components)with a large amount of information in the image for labeling.This thesis mainly does the following work:(1)In order to alleviate the problem of large amount of annotation work and the difficulty in labeling specific areas,the active learning method is combined with image matting algorithm for the first time,and a image matting algorithm based on active learning is proposed.(2)Introduce image significance detection into image matting.Comparatively analyze the effects of five classical significance detection algorithms on foreground detection,and choose the optimized method to calculate the significance of image matting components in this thesis.Propose automatic pre-labeling of matting components based on their significances,can quickly improve matting effect without increasing user burden.(3)For the active learning image matting,multiple sample selection strategies are proposed and analyzed,which are classified into basic sample selection strategy and improved sample selection strategy.The basic sample selection strategies include: component significance strategy,component center prior strategy,component background prior strategy,component uncertainty strategy and component representativeness strategy.The improved sample selection strategy can be improved simultaneously from two aspects: combined basic selection strategy and automatic pre-labeling based on component significance.Using the matting field standard test database,a large number of experiments were conducted on images in the alpha database.The experimental results show that,while ensuring the accuracy of image matting results,the use of active learning matting algorithm can greatly reduce the degree of difficulty and the number of marking samples(matting components in this thesis).