WATERSHED SUPERPIXEL
Zhongwen Hu
1
Qin Zou
2
Qingquan Li
?,1
1
Shenzhen Key Laboratory of Spatial Smart Sensing and Service,
Shenzhen University, P.R. China
2
School of Computer Science, Wuhan University, P.R. China
{zwhoo@szu.edu.cn, qzou@whu.edu.cn, liqq@szu.edu.cn}
ABSTRACT
As a pre-processing tool, superpixel algorithms have been
popular used in many computer-vision applications. High
efficiency is a desired property of superpixel algorithms, e-
specially in real-time vision systems. In this paper, a novel
high-efficient superpixel algorithm is developed based on the
watershed algorithm, namely the spatial-constrained water-
shed (SCoW). SCoW performs watersheding in a marker-
controlled manner, with a set of evenly placed markers.
To align superpixel boundaries to image edges, an edge-
preserving scheme is embedded into the SCoW which makes
a balance between the homogeneity and the compactness.
Without any complex computing, the proposed superpixel
algorithm is found to produce high quality superpixels as
traditional superpixel algorithms, while holding much higher
efficiency.
Index Terms— superpixel, watershed, spatial constraint,
image segmentation
1. INTRODUCTION
Superpixel segmentation, which over-segments an image
into a number of superpixel regions, is a widely used pre-
processing step for image- and vision-based systems. A
superpixel region is commonly defined as a perceptually uni-
form and homogenous region in the image [1, 2]. With the
development of computer vision and multimedia techniques,
superpixel algorithms have becoming increasingly popular
for many applications, e.g., object detection and tracking [3],
image segmentation and modeling [4, 5, 6], salience detec-
tion [7, 8, 9, 10] and image and object classification [11], etc.
This study is jointly supported by grants from Postdoctoral Science
Foundation of China under grants No. 2014M562206 and2012M521472,
Geographic National Condition Monitoring Engineering Research Cen-
ter of Sichuan Province under grant No. GC201513, National Natural
Science Foundation of China under grant No.61301277, Shenzhen Sci-
entific Research and Development Funding Program under grants No.
ZDSY20121019111146499 and No. JSGG20121026111056204, Shenzhen
Dedicated Funding of Strategic Emerging Industry Development Program
under grant No. JCYJ20121019111128765 and National Basic Research
Program of China (973 Program) under grant No. 2012CB725303. *Cor-
responding author: Dr. Q.Q. Li.
Fig. 1. Watershed Superpixels. (a) Traditional watershed, (b)
watershed with uniform markers but without spatial constrain,
(c) the proposed SCoW.
One main advantage of using superpixel is to improve the
computational efficiency in the corresponding systems. With
superpixels, the later-stage processing would be speedup s-
ince the number of image primitives has been dramatically
reduced as compared to the original pixel representations.
Another advantage is that, the superpixels provide to investi-
gate the image in a larger scale, which can give better spatial
support to compute regional features.
In the past two decades, important efforts have been
made in the field of image segmentation, e.g., watershed
segmentation [12], Normalized Cut [13], graph-based seg-
mentation [14], statistic region merging [15], etc. Generally,
image segmentation is used as a pre-processing step for vision
computing applications. However, with increasing demand
of high-performance pre-processing, the above segmentation
methods gradually show their limitations, and superpixel al-
gorithms are found to be more suitable as a pre-processing
tool. Exactly, a set of famous superpixel algorithms have
been proposed in the past decade [16, 17, 1], In [16], a
method named Turbopixels introduces spatial constraint to
level-set evolution to get dense over-segmented and compact
superpixels. While in SLIC [1], spatial constraint is embed-
ded into a k-mean clustering method to achieve superpixel
segmentation. These two methods can produce superpixels
effectively, however both requires a complex iterative process
to refine the results. Some newly developed algorithms can
be found in [18, 19, 20]. In [19], superpixels are gained
by a structure sensitive over-segmentation technique which
exploits Lloyd’s k-means algorithm with a geodesic distance.
- 1
- 2
- 3
前往页