You are on page 1of 4

Region growing based segmentation with automatic

seed selection using threshold techniques on


X-radiography images
Muthukumaran Malarvel1, Purna Chandra Rao Bhagi,
Gopalakrishnan Sethumadhavan2 Saravanan Thangavel,
School of Computing, Arunmuthu Krishnan
SASTRA University, Indira Gandhi Centre for Atomic Research,
Thanjavur, India. Kalpakkam, India.
1
e-mail: mmkmtech@gmail.com,
2
e-mail: sgk@mca.sastra.edu

Abstract In this article, a novel scheme based on region unevenly illuminated radiography images. Rafael Vilar et al.
growing segmentation is presented to detect features. This [4] presented an approach to establish a weld region in
method provides solution to automatic seed selection in region different stages. The weld region is isolated from the rest of
growing segmentation, thus aims to segment the region of interest the elements by separating an object from its background.
in an uneven background in industry X-radiography images. Otsus method was utilized in this approach to select
Otsu and Kittlers thresholds are employed for the automatic threshold; and the connected components in the binary image
seed selection. The performance of the proposed method is were labeled. Subsequently, the weld region was identified by
evaluated by the relative foreground area error metric and
a criterion differentiating between labeled objects, and the
compared with other existing segmentation methods. The
maximum area was established. Wang et al. [5] presented
performance measures show that the proposed method is capable
of providing the required segmentation of features.
histogram based seed selection method, which employs region
growing of an object. In this method, histogram and threshold
Keywords seeded region growing; automatic seed selection; can be identified iteratively, where a histogram is generated
segmentation; X-radiography image for the output image of the previous image. Zahran et al. [6]
presented a method for automatic segmentation of radiography
images to determine the region of interest by using optimal
I. INTRODUCTION threshold of the image. The histogram was useful for the
The main goal of segmentation is to acquire information for automatic determination of the optimal threshold between the
an easy analysis of features. Several segmentation techniques, objects and the background. Panigrahi et al. [7] proposed an
having different performance, have recently been proposed [1]. automatic seed selection method for ultrasound images. This
Homogeneity is a significant property of a region and is used method employed mean filter in preprocessing the image and
as the fundamental segmentation principle in region growing utilized texture features to select seed point.
based segmentation. The key idea of this method is to split an
image into regions of maximum homogeneity. Seeded region The aforementioned studies focused on segmentation of
growing technique is also important for image segmentation, foreground and background in radiography images with
which is based on the information about the semantic image various pre-processing and post-processing techniques. This
components in the seed selection procedure [2]. In this process, study proposes a new region growing based segmentation
the selection of the labeled seeds is critical. technique with automatic seed selection for segmenting
features from uneven background in industry X-radiography
Threshold based techniques are computationally images without using any pre-processing techniques.
inexpensive, fast and simple to implement but have few
drawbacks such as the threshold based algorithms neglect the For automatic seeded region growing process, seeds are
spatial information of the image, may miss edges, and the selected automatically by addressing both Otsu [8] and
selection of the threshold value is difficult. Similarly, edge Kittlers [9] thresholds. In industry X-radiography images, the
based segmentation gives good results in few types of defects objects of interest are the features of different intensities; while
present in the industry X-radiography images, while few others the background intensity substantially overlaps distribution in
are not clearly predictable by this technique. It is very difficult different regions in an image. It is assumed that an X-
to segment the accurate border and closed edges by using edge radiography image contains more than two principal gray scale
based segmentation in low quality images. Thus, it is difficult regions, since an object is present in an uneven background
to obtain the features and dimensions of the features in an contrast instead of a uniform background contrast in the whole
industry X-radiography image [3]. Therefore, the proposed image. When Otsu and Kittlers methods are applied on
method employs region grows based segmentation technique. industry X-radiography images, the background and
foreground are segmented differently. Thus, we used a
In recent years, region grow based methods [46] have combination of these two methods in selection of seeds
been presented, which are suitable for low contrast and automatically for segmentation.

978-1-5090-0612-0/16/$31.00 2016 IEEE


2016 IEEE International Conference on Computational Intelligence and Computing Research

II. PROPOSED METHOD 1 if N(i) of P t1 < T 1 | P t1


if N(i) of P t2 < T 2 | P t
It is assumed that an image contains two different
(uneven) background contrast levels with foreground objects. g (x, y) =
Consider that there are two background contrast regions Br1 (3)
and Br2 and an objects Fo is present in each region of the input

image f. T1 and T2 are the threshold value of f obtained from 0 otherwise

Otsu and Kittlers methods, respectively. Segmentation of an


input image f to an output image g using T1 and T2 gives The proposed algorithm presented here and the
following different results: diagrammatic representation of the proposed seeded region
growing method is shown in Fig. 2.
If Br1(x,y), Br2(x,y) and Fo(x,y) f(x,y)

g ( x, y) = {
1 for Br1 ( x, y ) and Fo ( x, y ) T1
0 for Br 2 ( x, y ) <T1 or
Proposed algorithm
First phase:

{
Step 1: Find threshold values by using Otsus method (T1) and
for Br 2 ( x, y ) and Fo ( x, y ) T1
g ( x, y) = 10 for Br1 ( x, y ) <T1
Kittlers method (T2) of an input image f.
Step 2: Identify the seed pixels, which are equal to T1 and T2,

g ( x, y) = {
1 for Br1 ( x, y ) and Fo ( x, y ) T2
0 for Br 2 ( x, y ) <T2 or
and assign to (set of all unlabeled pixels in an
image). Then label all seed pixels as Pt1 and Pt 2 .

{
for Br 2 ( x, y ) and Fo ( x, y ) T2
g ( x, y) = 10 for Br1 ( x, y ) <T2
Second phase:
Step 3: Scan all the labeled pixels. If label Pt1 and Pt 2
These results are not enough to segment the foreground then assume it is a seed pixel.
objects accurately. Therefore, the proposed method utilizes two Step 4: Apply the criterion (3).
threshold values to select seeds automatically for better Step 5: Check the neighboring pixels, they should be at least
segmentation. The proposed automatic seed selection method 8-connected and add them to the region if they are
has two phases: (1) Seed collection and (2) Region growing. satisfying the criterion.
The schematic of the proposed method is shown in Fig. 1. Step 6: Repeat step 3 for each of the labeled pixels.
Step 7: Stop if no more pixels can be added.

Fig. 1: Schematics of the proposed method

In first phase, is the set of all unlabeled pixels in an


input image.
{ f | is unlabeled and (0 < > 255)} (1)
Both threshold values T1 (Otsu) and T2 (Kittler) of pixels
matching in a set can be considered as a seed pixel. The set Fig. 2: The proposed seeded region growing
of seed pixels denoted as Si then assigns a label of this set of
seed pixels as Pt1 and Pt 2 based on the threshold values T1 and III. RESULTS AND ANALYSIS
T2, respectively. The proposed method was compared with Otsu and
Kittlers threshold methods, and its performance was evaluated
True if S i = T1 or T 2 | S i
Si = { false otherwise
(2)
through Relative Foreground Area error (RAE) metric [10].
RAE is based on the measure of the area and can be calculated
between the segmented results and the ground truth image. The
In second phase, region growing can be established based ground truth image was obtained by manual thresholding
on the immediate 8-connected neighbor pixel N(i) from the technique and expert opinion. A lower value of RAE indicates
labeled seed pixel with the conditions of (3). a better output.
2016 IEEE International Conference on Computational Intelligence and Computing Research

The proposed scheme is being experimented with the X- Table 1: Performance evaluation measures of the segmented image shown
Radiography of an aluminum cylinder with holes drilled in in Fig. 3.
different diameters from the top of the surface. Image Methods RAE
When Otsu and Kittlers methods are applied on images Otsu 0.4026
of varying intensities, it has been found that the holes are Image I Kittler 0.1489
extracted partially. But, the proposed scheme segmented all Proposed 0.0943
the holes precisely even without applying any of the
preprocessing techniques. Otsu 0.4865
Image II Kittler 0.0674
Table 1 illustrates the RAE metric values of the test image Proposed 0.0356
shown in Fig. 3. The measured values ensure the reliability of
the proposed method. Therefore, we conclude the proposed Otsu 0.6397
scheme well segmented the region of interest from uneven Image III Kittler 0.1114
background and low contrast images. Fig. 3 shows a sample Proposed 0.0931
outcome of the experiments.

Fig. 3: Segmentation results; First to third row image I, II and III, respectively (a) original X-radiography image; (b) ground truth; (c) Otsus method; (d)
Kittlers method; (e) The proposed method.
2016 IEEE International Conference on Computational Intelligence and Computing Research

IV. CONCLUSIONS REFERENCES


In this paper, a new segmentation scheme based on region [1] Peng R, Varshney PK, On performance limits of image segmentation
growing method was proposed without implication of algorithms, Comput Vis Image Underst, Vol. 132, pp. 2438, 2015.
preprocessing techniques in the input image. This scheme can [2] Fan J, Zeng G, Body M, Hacid MS, Seeded region growing: An
extensive and comparative study, Pattern Recognit Lett, Vol. 26,
segment foreground from uneven background in industry X- pp.11391156, 2005.
radiography images using automatic seeded region growing [3] Rathod V R, Anand R S, A comparative study of different
technique. The combination of Otsu and Kittlers thresholds is segmentation techniques for detection of flaws in NDE weld images, J
employed to select seeds automatically. The proposed Nondestruct Eval, Vol. 31, pp. 116, 2012.
algorithm was compared with Otsu and Kittlers threshold [4] Vilar R, Zapata J, R Ruiz, An automatic system of classification of
methods, and RAE evaluated its performance. The results weld defects in radiographic images, NDT & E Int, Vol. 42, pp.467
obtained by the proposed method are better than other 476, 2009.
methods. [5] Wang Y, Huang J, Object Detection in X-ray Images Based on Object
Candidate Extraction and Support Vector Machine, IEEE 9th Intl.
Conference on Natural Computation (ICNC), pp.173177, 2013.
ACKNOWLEDGMENT [6] Zahran O, Kasban H, El-Kordy M, El-Samie FEA, Automatic weld
defect identification from radiographic images, NDT E Int, Vol. 57, pp.
The authors would like to acknowledge the support 2635, 2013.
rendered by the managements of SASTRA University and [7] L. Panigrahi, K. Verma and B.K.Singh, An enhancement in automatic
Indira Gandhi Centre for Atomic Research, Kalpakkam by seed selection in breast cancer ultrasound images using texture features,
way of providing necessary infrastructure and image database Intl. Conference on advances in computing, communications and
for this research. This work was a part of generic image informatics (ICACCI), pp. 21-24, 2016.
processing tool named Bhadraloka being developed for the [8] Otsu N, A Threshold Selection Method from Gray-Level Histograms,
IEEE Trans Syst Man Cybern, vol 9, pp.6266, 1979.
research project entitled Development of techniques for
[9] Kittler J, Illingworth J , Minimum error thresholding, Pattern
processing radiographic images for automated detection of Recognit, vol.19, pp.4147, 1986.
defects (No. 2013/36/40-BRNS/2305) in Dot NET platform [10] Sezgin M, Sankur B, Survey over image thresholding techniques and
at SASTRA with the funding assistance from Board of quantitative performance evaluation, J Electron Imaging, vol. 13(1),
Research in Nuclear Science (BRNS), Department of Atomic pp.146165, 2004.
Energy, Government of India.

You might also like