IJCSI International Journal of Computer Science Issues, Vol.

7, Issue 5,September 2010

ISSN (Online): 1694-0784
ISSN (Print): 1694-0814

A Study Of Image Segmentation Algorithms For

Different Types Of Images
Krishna Kant Singh1 , Akansha Singh2

1
Deptt. Of Electronics & Communication
Hermes Engineering College
Roorkee , India

2
Deptt. Of Information Technology , AKGEC
Ghaziabad, India

Abstract otherwise it is zero. Segmentation is the

operation at the threshold between low-level
In computer vision, segmentation refers to the process image processing and image analysis. After
of partitioning a digital image into multiple segments segmentation, it is known that which pixel
(sets of pixels, also known as superpixels).Image belongs to which object. The image is parted into
segmentation is typically used to locate objects and regions and we know the discontinuities as the
boundaries (lines, curves, etc.) in images. More boundaries between the regions. The different
precisely, image segmentation is the process of types of segmentations are
assigning a label to every pixel in an image such that
Pixel-Based Segmentation: Point-based or
pixels with the same label share certain visual
characteristics.The result of image segmentation is a pixel-based segmentation is conceptually the
set of segments that collectively cover the entire simplest approach used for segmentation.
image, or a set of contours extracted from the image . Edge-Based Segmentation: Even with perfect
Each of the pixels in a region are similar with respect illumination, pixel based segmentation results in
to some characteristic or computed property, such as a bias of the size of segmented objects when the
color, intensity, or texture.Due to the importance of objects show variations in their gray values
image segmentation a number of algorithms have been .Darker objects will become too small, brighter
proposed but based on the image that is inputted the objects too large. The size variations result from
algorithm should be chosen to get the best results. In
the fact that the gray values at the edge of an
this paper the author gives a study of the various
algorithms that are available for color images,text and object change only gradually from the
gray scale images. background to the object value. No bias in the
size occurs if we take the mean of the object and
the background gray values as the threshold.
Keywords: image segmentation , region growing,
marker However, this approach is only possible if all
objects show the same gray value or if we apply
1. Introduction different thresholds for each objects. An edge-
All image processing operations generally aim at based segmentation approach can be used to
a better recognition of objects of interest, i. e., at avoid a bias in the size of the segmented object
finding suitable local features that can be without using a complex thresholding scheme.
distinguished from other objects and from the Edge-based segmentation is based on the fact
background. The next step is to check each that the position of an edge is given by an
individual pixel to see whether it belongs to an extreme of the first-order derivative or a zero
object of interest or not. This operation is called crossing in the second-order derivative.
segmentation and produces a binary image. A
pixel has the value one if it belongs to the object; Region-based methods focus attention on an
important aspect of the segmentation process
IJCSI International Journal of Computer Science Issues, Vol. 7, Issue 5,September 2010
www.IJCSI.org

missed with point-based techniques. There a only utilize gray-Level information can not
pixel is classified as an object pixel judging extract the target from background; we must by
solely on its gray value independently of the means of color information. Accordingly, with
context. This meant that isolated points or small the rapidly improvement of computer processing
areas could be classified as object pixels, capabilities, the color image processing is being
disregarding the fact that an important more and more concerned by people. The color
characteristic of an object is its connectivity.If image segmentation is also widely used in many
we use not the original image but a feature image multimedia applications, for example; in order to
for the segmentation process, the features effectively scan large numbers of images and
represent not a single pixel but a small video data in digital libraries, they all need to be
neighborhood, depending on the mask sizes of compiled directory, sorting and storage, the color
the operators used. At the edges of the objects, and texture are two most important features of
however, where the mask includes pixels from information retrieval based on its content in the
both the object and the background, any feature images and video. Therefore, the color and
that could be useful cannot be computed. The texture segmentation often used for indexing and
correct procedure would be to limit the mask size management of data; another example of
at the edge to points of either the object or the multimedia applications is the dissemination of
background. But how can this be achieved if we information in the network. Today, a large
can only distinguish the object and the number of multimedia data streams sent on the
background after computation of the feature? Internet, However, due to the bandwidth
Obviously, this problem cannot be solved in one limitations; we need to compress the data, and
step, but only iteratively using a procedure in therefore it calls for image and video
which feature computation and segmentation are segmentation.
performed alternately. In the first step, the
features are computed disregarding any object 2.1 Seed Region Growing Algorithm And
boundaries. Then a preliminary segmentation is Watershed Algorithm
performed and the features are computed again,
now using the segmentation results to limit the The basic idea of region growing method is a
masks of the neighborhood operations at the collection of pixels with similar properties to
object edges to either the object or the form a region. The steps are as follows:
background pixels, depending on the location of
the center pixel. To improve the results, feature (i) find a seed pixel as a starting point for
computation and segmentation can be repeated each of needed segmentation.
until the procedure converges into a stable result. (ii) Merge the same or similar property of
pixel (Based on a pre-determined
Model-Based Segmentation : All segmentation growing or similar formula to
techniques discussed so far utilize only local determine) with the seed pixel around
information. The human vision system has the the seed pixel domain into the domain
ability to recognize objects even if they are not of seed pixel.
completely represented. It is obvious that the (iii) These new pixels act as a new seed
information that can be gathered from local pixel to continue the above process until
neighborhood operators is not sufficient to no more pixels that satisfy the condition
perform this task. Instead specific knowledge can be included.
about the geometrical shape of the objects is
required, which can then be compared with the The seed region growing algorithm is proposed
local information. This train of thought leads to by Adams and Bischof, Metmert and
model-based segmentation. It can be applied if Jackway[10] further described the dependency
we know the exact shape of the objects contained relationship between pixels in the seed growth:
in the image. (i) The first order of dependence occurs
when the number of pixels has the
2. Color Image Segmentation algorithm same difference ratio as their vicinity.
(ii) The second order of dependence
The human eyes have adjustability for the occurs when a pixels has the same
brightness, which we can only identified dozens difference ratio as their vicinity
of gray-scale at any point of complex image, but Frank and Shouxian Cheng applied the automatic
can identify thousands of colors. In many cases, seed selection method, they selected seed which
IJCSI International Journal of Computer Science Issues, Vol. 7, Issue 5,September 2010
www.IJCSI.org

can represents needed segmentation region based up-table is prepared, that reveals the connectivity
on certain similarity criteria and proposed a of pixels that have different labels. For most
strategy to solve the above two pixels purposes, the labeling of the pixels found the
dependence [1]. The watershed algorithm is first pass, combined with the look-up-table,
more representative in the application of provides a complete segmentation, but to
mathematical morphology theory for image actually output an image with unique pixel labels
segmentation. Watershed algorithm is a region- in each region, a pass through the image using
based segmentation techniques image that uses the look-up table is required. A new parallel
image morphology [2]. Watershed algorithm is region segmenting and labeling algorithm is
an iterative adaptive threshold algorithm. The available, that is applicable to gray-scale images,
idea of watershed algorithm is from geography, and is appropriate to coarse scale parallel
it see gradient magnitude image as a topographic programming. The key feature of this algorithm
map, the gradient magnitude in correspond with is the geometric splitting of the image into
altitude, the different gradient in correspond with rectangular blocks, with one pixel overlap at
the peak and basin in valley in the image. It sees joins. Then using Cohen's one pass algorithm,
every object of image (including background) as each region is separately labeled. Then by
a separate part and requested there must have examination of the overlap regions, the
one tag at least in the each object (or seed connectivity of different region labels is
points). Marker is knowledge about the object determined, through connectivity tables, and
based on application-oriented; it is selected by finally the overall image is completely
the operator manually or by automatic process. segmented into connected domains of common
The objects can use watershed algorithm to gray-scale. The parallelizable algorithm for the
transform and develop regional growth after segmentation of gray-scale images involves
maker. performing the one-pass algorithm on
rectangular sub-images, with a single row or
3. Gray-scale Image Segmentation column overlap. What is thus produced is a label
The segmentation of image raster data into for each image pixel, together with a
connected regions of common gray-scale has connectivity LUT for each region. Then, from
long been seen as a basic operation in image the overlap rows and columns produce overlap
analysis. In texture analysis, just this type of tables, showing how labels in each region are
segmentation is possible after individual pixels related.
in an image have been labeled with a numeric
classifier. In preparing images for used in 4. Text Segmentation : It is well known that
geographic information systems (GIs) this text extraction, including text detection,
segmentation is usually followed by the localization, segmentation and recognition is
production of a vector representation for each very important for video auto-understanding. In
region. The original algorithm for segmentation, this paper, we only discuss text segmentation,
developed by Rosenfeld-pfaltz[3], described a which is to separate text pixels from complex
two pass 'sequential algorithm' for the background in the sub-images from videos. Text
segmentation of binary images. The key feature segmentation in video images is much more
of the Rosenfeld-pfaltz algorithm is that the difficult than that in scanning images. Scanning
image is raster-scanned, first the forward images generally has clean and white
direction, from top left to bottom right, then background, while video images often have very
backwards. During the forward pass, each pixel complex background without prior knowledge
is located a region label, based on information about the text color. Although there have been
scanned through; the regions so demarcated may several successful systems of video text
have pixels with more than one label therein. extraction, few researchers specially study text
During the backwards pass, a unique label is segmentation in video images deeply. The used
assigned to each pixel. Hence this classic strategies could be classified into two main
algorithm can be described as a two pass categories: (1) difference (or top-down) and (2)
algorithm. In a previous paper Cohen [4] similarity based (or bottom-up) methods. The
presented a one-pass algorithm was presented for first methods are based on the foreground-
the segmentation of gray-scale images. Cohen's background contrast. For example, fixed
single pass algorithm proceeds to label pixels on threshold value method [2], Otsus adaptive
a forward pass, exactly as in Rosenfeld-Phltz, thresholding method [5], global & local
except that during the same forward pass, a look- thresholding method [2], Niblacks method [6].
IJCSI International Journal of Computer Science Issues, Vol. 7, Issue 5,September 2010
www.IJCSI.org

In general, they are simple and fast, but fail when [4] Harvey A. Cohen , Parallel Algorithm for
foreground and background are similar. Gray-scale Image Segmentation, Proceedings
Alternatively, the similarity based methods Australian New Zealand Conf. on Intelligent
cluster pixels with similar intensities together. Information Systems,1996.
[5] N. Otsu, A threshold selection method from
For example, Lienhart uses the split & merge
grey level histogram, IEEE Trans. Syst. Man
algorithm [7], Wang et al. combine edge Cybern., vol. 9 no. 1,1979, pp. 62-66.
detection, watershed transform and clustering [6] W. Niblack, An Introduction to Digital Image
[8]. However, these methods are unstable, since Processing, pp. 115-116, Prentice Hall, 1986.
they exploit many intuitive rules about text [7] R. Lienhart, Automatic Text Recognition in
shape. As an alternative, Chen et al. convert text Digital Videos, Proceedings SPIE, Image and
pixel clustering to labeling problem using Video Processing IV, pp. 2666-2675, 1996.
Gibbsian EM algorithm [9]. This method is [8] K. Wang, et al., Character Segmentation of
effective but too time consuming. The main Color Images from Digital Camera, Intl. Conf. on
Document Analysis and Recognition, pp.210-214,
problem of most existing methods is that they are
2001
sensitive to text color, size, font and background [9] D. Chen, et al., Text Detection and
clutter, since they simply exploit either general Recognition in Images and Video
segmentation method or some prior knowledge. Frames, Pattern Recognition, vol. 37, pp. 595-
To overcome these problems a new algorithm 608, 2004.
based on stroke filter is available which [10] Peck, M.A., Linda G.Shapiro.George
discovers the intrinsic characteristics of text and C.Stockman, Computer Vision, vol. 103, pp.
design a robust algorithm specially for text 2005-2024, 2000.
segmentation. In this algorithm stroke filter (SF),
About The Authors
describes the intrinsic characteristics of text in
terms of scale, orientation and response, a stroke K.K.Singh is currently working as Assistant
feature based text polarity determination method Professor in ECE department of Hermes
is available and local region growing method for Engineering college. He holds B.Tech (EI) and
segmentation refinement based on stroke features M.Tech (VLSI) degree. He has more than five
and global & local statistic similarities. years of teaching experience. He has published a
number of popular text books including Digital
5. Conclusion Communications, VLSI Technology, Automatic
Control Systems, Mobile Computing, Digital
Image segmentation has become a very Image Processing, Network Analysis and
important task in todays scenario. In the present Synthesis and IT Infrastructure management. He
day world computer vision has become an has also presented and published a number of
interdisciplinary field and its applications can be papers in national/international
found in any area be it medical , remote sensing , journals/conferences. His areas of specialization
electronics and so on . Thus, to find a appropriate are Digital Communications, VLSI Technology,
segmentation algorithm based on your Automatic Control Systems , Networks and
application and the type of inputted image is many more.
very important. In this paper the author has
explained and suggested a few application Akansha Singh is currently working as
specific segmentation algorithm which also take Assistant Professor in IT department of AKGEC,
into consideration the type of image inputted like Ghaziabad. She holds B.Tech (CS) and M.Tech
color, gray scale and text. (CS) honors degree. She has several years of
teaching experience. She has published a number
References
of popular text books including Web
[1] Frank and Shouxian Cheng , Automatic Technology, IT Infrastructure Management,
seeded region growing for color image Mobile Computing, Digital Image Processing ,
segmentation,Image and vision computing Network Analysis and synthesis . She has also
Journal,vol.23 issue 10,pp.877-886. presented and published a number of papers in
[2] R.C. Gonzalez, R.E. Woods, Digital Image national/international journals/conferences. Her
Processing, Prentice-Hall, Englewood Cliffs, NJ, areas of specialization are Web Technology,
2002. Mobile Computing, Digital Image Processing
[3]A.Rosenfeld and D.pfaltz, Sequential and many more.
operatons in Digital Picture Processing J.A.C.M.
(1966) EN0 4 pp 471-494.