Chinese Journal of Electronics
Vol.20, No.1, Jan. 2011
Separation of Objects with Unclear Edges from
the Nonuniform Background
∗
LI Huimin
1,2
, LIU Shufen
1
and WANG Honghua
3
(1.College of Computer Science and Technology, Jilin University, Changchun 130012, China)
(2.Department of Electronic Warfare, System Engineering Research Institute, Beijing 100036, China)
(3.Institute of Sciences, PLA University of Science and Technology, Nanjing 211101, China)
Abstract — This paper presents a method for separat-
ing objects with unclear edges from the nonuniform illu-
mination background. The paper uses an iterative lowpass
filter to estimate the rough background first. And then,
objects are located by removing the estimated background
from the original image, meanwhile the rough segmenta-
tion is worked out too. Next, the primary segmentation
output is dilated by means of the iterative lowpass filter
again to connect most of the objects together as a larger
one. After that, some background points are selected from
the dilated image as the control points for image back-
ground reconstruction based on B-spline. Finally, the re-
constructed background is removed from the original image
directly to make the separation. A quantitative evaluation
of the performance is also presented in this paper.
Key words — Image segmentation, Unclear edge,
Nonuniform background, B-spline, Iterative lowpass filter.
I. Introduction
To separate objects from background in an image pictured
by a camera under conditions of poor and nonuniform illumi-
nation is a common requirement in a variety of applications
involving visual inspection
[1]
. For example, Fig.1(a)showsan
image taken by a camera for the quantitative evaluation of
Thin layer chromatography (TLC) where all of the 5 spots
need to be precisely separated from the background.
Thresholding is a simple but effective method to solve this
kind of problem. Among various global thresholding methods,
Otsu’s method
[2]
was tested to be the most efficient one
[3]
.
But this technique doesn’t work well when images contain un-
even background or objects with unclear edges, as shown in
Fig.1(a). Because they only consider the gray-scale statistic
relationship between pixels, other relationships such as spa-
tial position are not involved. Therefore, local thresholding
becomes the choice.
Yanowitz and Bruckstein
[1]
took the points with the high
gradient as the key points to make the thresholding surface.
Their method works very well for images of the object with
a clear edge, but it is not suitable for images of objects hav-
ing unclear edges, see Fig.1. Liu, et al.
[4]
, made a modifica-
tion to Yanowitz’s method by setting a surface parallel to the
background as the thresholding surface. Their consideration
achieved a more concordant thresholding surface to the origi-
nal image background. But the way of obtaining the support
points was as the same as Yanowitz’s method, it doesn’t per-
form well when dealing with images whose object’s edge is
unclear.
Fig. 1. A TLC image containing objects with unclear edges
against an uneven background. (a)Original;(b)Sobel
gradient; (c) Profile of one line of (a); (d) Its corre-
sponding line of (b)
Gatos et al.
[5]
presented a new adaptive approach for the
binarization of degraded document images. They used Sauvola
and Pietikainen’s method
[6]
, which made a hypothesis on the
gray values of text having values of 0 and background pix-
els with values near 255, to estimate the rough foreground
regions, and then obtain the background surface by interpo-
lating neighboring background intensities.
Zhang
[7]
developed another method to use the concave dis-
tribution feature of illumination to make the segmentation.
The performance of the method is quite good, but the trade-
off is that it has to find the illumination distribution feature
of the light sources first.
∗
Manuscript Received Jan. 2010; Accepted Aug. 2010.