Image Matting With Normalized Weight And Semi-supervised Learning

Nowadays,digital matting technology has become a key technology for producing film special effects.Although with good prospects,it still has many problems to be solved.The main stream matting methods combine sampling-based methods and propagation-based methods.But they could not balance the relative effects of sampling and propagation very well.The fineness of the trimap determines the quality of the matting,but the users want to pay as little effort as possible(so provide coarse trimaps)to get better matting.To solve these two problems,this thesis mainly does the following works.(1)A normalized-weight based matting method is proposed,in which,a normalized weighting parameter is used to well control the relative relationship between information from sampling and from propagation.A reasonable value range for this parameter is given based on statistics from a benchmark dataset.(2)In response to the phenomenon that the smaller the unknown area in the trimap is,the better performance the matting methods get,we propose three semi-supervised learning methods for matting,including the pre-processed semi-supervised learning,the iterative semi-supervised learning and the hybridsemi-supervied learning.The pre-processed semi-supervised learning first automatically labels some pixels of unknown regions in the trimap(reducing the unknown region)based on the original image and the trimap,and then does the matting based on the original image and the updated trimap.The iterative semi-supervised learning label spartial unknown regions in the trimap based on previously obtained matting results,and then put the new trimap into matting process.This process is iteratively performed.The hybrid semi-supervised learning combines the pre-processed semi-supervised learning and the iterative semi-supervised learning.These three semi-supervised learning methods are coupled with the closed matting,the learning matting,the comprehensive matting and the weight control matting for testing.The benchmark dataset,alpha dataset are used and the experimental results show that the semi-supervised learning methods can improve matting performance.