MSRM Based Object Extraction Method for Image Sequences
Wenting Yu
1
, Jingling Wang
2
, Long Ye
3
1, 2
Key Laboratory of Media Audio & Video, Ministry of Education, Communication University of
China, Chaoyang District, 100024, Beijing, P.R. China
a
email:ywt_happy@sina.com
Keywords: Object extraction, SLIC-super pixel, MSRM
Abstract. Object extraction, which aims to accurately separate a foreground object from its
background in still images, plays an important role in many computer vision applications. An
interactive object extraction method based on MSRM (maximal similarity based region merging) is
presented in this paper. We can manually mark the target and background only one time in any one
image of the image sequence to obtain the object extraction result of the image sequence. Compared
to currently used method based on graph cut algorithm that manually marks the target and
background on all the images one by one to get the object extraction result, our method is more
efficient and the result is as precious as with other methods.
Introduction
Accurately separate a foreground object from its background plays an important role in many
computer vision applications. Image segmentation is to separate the desired objects from the
background, recently various methods have been proposed, For example, in [1, 2], Lietal combined
graph-cut with watershed pre-segmentation for better segmentation outputs, where the segmented
regions by watershed, instead of the pixels in the original image, are regarded as the nodes of graph
cut. Relies on the work of Graph Cut, a lot of methods are proposed, such as Grabcut and Lazy
Snapping. Jifeng Ning and Lei Zhang proposed a novel interactive region merging method based on
the initial segmentation of mean shift which is called MSRM[3].
Grabcut[4] extends graph cut to color image and incomplete trimaps, and consists of two
portions: automatic hard segmentation and border matting portion. The idea of the automatic
segmentation is to build a graph where each node corresponds to a pixel such that a
max-flow/min-cut algorithm solves the segmentation iteratively. The inclusion of color information
by using Gaussian Mixture Models in the Graph Cut algorithm increases its robustness.
Lazy Snapping[1] also separates the object extraction into two tasks: object context specification
and boundary refinement. Graph cut[5] is used in both tasks. To specify an object in a given image,
the user marks a few lines on the image by dragging the mouse cursor while holding a button (left
button indicating the foreground, and right button for the background). As the author admitted, the
object marking step and boundary editing step have not been combined in a unified way.
In the MSRM [3] method, the interactive information is introduced as markers, which are input
by the users to roughly indicate the position and main features of the object and background. Then
the proposed method will calculate the similarity of different regions and merge them based on the
proposed maximal similarity rule with the help of these markers. The object will then be extracted
from the background when the merging process ends.
Based on the above considerations, we proposed a fast computation object extraction method for
image sequences based on the MSRM [3]. In this method, the users manually mark the target and
background only one time and can extract the object from the image sequence successfully.
Compared to currently used method based on graph cut algorithm that manually marks the target
and background on all the images one by one to get the object extraction result, our method is more
efficient and the result is as accurate as with other methods. Experimental results on multiple kinds
of color image sequences show the effectiveness and convenience of the approach.
Advanced Materials Research Vols. 1049-1050 (2014) pp 1675-1680 Submitted: 2014-08-25
© (2014) Trans Tech Publications, Switzerland Accepted: 2014-09-01
doi:10.4028/www.scientific.net/AMR.1049-1050.1675 Online: 2014-10-10
All rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of Trans
Tech Publications, www.ttp.net. (ID: 130.237.29.138, Kungliga Tekniska Hogskolan, Stockholm, Sweden-11/07/15,13:46:35)