Analysis and Comparison of various Methods for Text Detection from Images using MSER Algorithm

International Journal on Recent and Innovation Trends in Computing and Communication ISSN: 2321-8169
Volume: 5 Issue: 7 759 – 763
_______________________________________________________________________________________________
759
IJRITCC | July 2017, Available @ http://www.ijritcc.org
_______________________________________________________________________________________
Analysis and Comparison of various Methods for Text Detection from Images
using MSER Algorithm
Dr. Dilip Sharma
Ujjain Engineering College, Ujjain (Mp)
Email: drdilipsharma72@gmail.com
Amit Kumar Pandey
Ujjain Engineering College, Ujjain (Mp)
Email: amitrwa@gmail.com
Abstract — In this paper analysis and comparison of various methods for text detection is carried by using canny edge detection algorithm and
MSER based method along with the image enhancement which results in the improved performance in terms of text detection. In addition, we
improve current MSERs by developing a contrast enhancement mechanism that enhances region stability of text patterns to remove the blurring
caused during the capture of image Lucy Richardson de blurring Algorithm is used.
Keywords- MSER, CC,
__________________________________________________*****_________________________________________________
I. INTRODUCTION
In present daily life text plays an important role in daily life
because of its rich information that is why automatic text
detection in natural images has many applications [1-4]. But
detecting the text from natural image is always a challenging
problem. Due to the presence of variation in the background
and as the size of the text also not fixed in case of natural
images it is very difficult to identify the text accurately.
Through tremendous efforts have recently been devoted in this
research but still reading texts in unconstrained environment is
still challenging and remain a problem [4-6]. Today text
detection finds many applications in various fields, including
visual impairment assistance, tourist assistance, content based
image retrieval and unmanned ground vehicle navigation.
Today most of the images are taken from the camera and other
handhold devices which is not fixed and sometimes due to
movement of the object the problem of blurring is observed
which makes it even more difficult to detect the text from
natural images [7-9]. Here in this thesis idea is proposed to
detect and recognize the text contains in the image as the main
problem in computer vision is to separate the text from the
background components [9-12]. There are many methods
which are still used to detect the text from the natural scene
such as text detection using edge analysis, robust text
detection, Real time text tracking, but none of them is
promising [13].
II. DIFFERENT METHODS FOR TEXT DETECTION
2.1 Texture based method: - Surface based techniques
look at nearby composition highlights inside little districts of a
picture. The content present in the pictures displays some
unmistakable textural properties, which might be utilized to
recognize it from the foundation. [3]
Gabor channels, Wavelets, Fast Fourier change, and so forth
are generally used to remove the textural properties of a
content district in a picture. In the event that the composition
35 elements are steady with the attributes of the content, all
pixels in the locale are set apart as content.[8]
2.2 Region based technique: - Area based strategies use
properties of the shading or dim scale in a content locale or
their disparities with the relating properties of the foundation.
Area based strategies can be further separated into two classes
1. Connected segment (CC) based
2. Edge-based
These techniques are otherwise called base up methodologies,
because of the way they work; i.e. by first recognizing
rudimentary (little) sub-structures, for example, CCs or edges,
and after that blending these sub-structures progressively into
bigger structures, until all the content territories are identified
[14].
2.2.1 Connected part (CC)
In CC-based strategies, the fundamental components
are made utilizing the 31 likeness of neighbor pixels in
grayscale or shading levels, while the edge construct
techniques center in light of the high differentiation between
the content and the foundation, distinguishing first the edges
brought on by the content shapes and afterward gathering
them, if conceivable CC-based techniques utilize a base up
methodology by gathering little segments into progressively
bigger parts until every one of the areas are recognized in the
picture. A geometrical investigation is expected to combine

Volume: 5 Issue: 7 759 – 763
_______________________________________________________________________________________________
760
_______________________________________________________________________________________
the content segments utilizing their spatial course of action in
order to sift through non-content segments and check the
limits of the content districts. In the CC approach, little
districts speaking to the content and non-content are to be
distinguished. With this in perspective, shading lessening by
bit dropping and shading bunching quantization is endeavored
and a short time later a multi-esteem picture deterioration
calculations utilized to decay the information picture into
various closer view and foundation pictures [Jain and Yu
1998].
2.2.2 Edge based methodology-
In the edge based methodology is was endeavored to
get high complexity edges for the continuous content hues y
utilizing the red casing of the RGB shading space (Agnihotri
and Dimitrova 1999). By method for the convolution
procedure with various veils, first the picture is improved, and
after that the edges are identified. This edge picture is further
prepared by gathering the neighboring edge pixels to single
associated part structures [7].
III. ANALYSIS OF EDGE BASED METHODOLOGY
3.1 Edge detection: - Edge identification is an operation in
PC vision framework which recognizes the sharp change in the
picture pixel. By recognizing the edges present in the picture
we can extraordinarily diminish the measure of information to
be handled .There are a few diverse edge identification
calculation exists yet here we are concentrating for the most
part on the calculation created by john F. Vigilant in 1986
[16]. In spite of one of the most seasoned technique for edge
identification it is one of the standard edge recognition
strategies and still utilized by the specialists.
3.2 CANNY EDGE DETECTION ALGORITHM:
The vigilant edge recognition is the most generally
utilized edge identification calculation to find sharp force
changes which is utilized to distinguish object limit in any
image. In shrewd edge location technique the calculation
characterize the pixel as an edge if the angle greatness of the
Pixel is bigger than those of the pixel at both its sides in the
direction of maximum power change
The calculation keeps running in 5 steps:
1. Smoothing: Blurring of the picture to expel clamor.
2. Discovering angles: The edges ought to be checked where
the inclinations of the picture has huge extents.
3. Non-most extreme concealment: Only neighborhood
maxima ought to be set apart as edges.
4. Twofold thresholding : Potential edges are controlled by
thresholding.
5. Edge following by hysteresis: Final edges are controlled by
stifling all edges that are not associated with an extremely
certain edge.
3.3 Image Enhancement-It is a process in which the
quality of image (poor illumination, coarse quantization) is
enhanced .In case of image enhancement the quality of the
image need to be improved without the availability of the
reference image. The idea behind the image enhancement is to
produce certain changes in the image which make the vision
system to easily understand the idea behind the image.
(a) (b) (c)
Fig.1: three different backgrounds with same grayscale
3.4 Contrast stretching
Low-contrast images can result from poor
illumination, lack of dynamic range in the image sensor, or
even wrong setting of a lens aperture during image acquisition.
The idea behind contrast stretching is to increase the dynamic
range of the gray levels in the image being processed [10].
Fig. (2). Image and its histogram before and after contrast
enhancement
3.5 Smoothing filter
Smoothing channels are utilized for obscuring and for
clamor diminishment. Obscuring is utilized as a part of
preprocessing steps, for example, expulsion of little points of
interest from a Picture before item extraction and spanning of
little crevices in lines or bends. Commotion diminishment can
finish by obscuring with a straight channel furthermore by
nonlinear sifting [12].

Volume: 5 Issue: 7 759 – 763
_______________________________________________________________________________________________
761
_______________________________________________________________________________________
3.6 Maximally Stable Extremal Regions
MSER regions are connected areas characterized by
almost uniform intensity, surrounded by contrasting
background. They are constructed through a process of trying
multiple thresholds.
The selected regions are those that maintain
unchanged shapes over a large set of thresholds. For color
images MSER algorithm replaced thresholding of the intensity
function with Agglomerative clustering, which is based on the
color gradients [3].
FIGURE(3) : EXAMPLES OF MSER REGION
3.7 MSER algorithm
MSER is a technique for blob location in pictures.
The MSER calculation separates from a picture various co-
variation locales, called MSERs: a MSER is a stable
associated part of some dark level arrangements of the picture.
• MSER depends on taking areas which stay almost the same
through extensive variety of limits. – All the pixels underneath
a given edge are white and every one of those above or
equivalent is dark. – If we are demonstrated a grouping
ofthresholded images with casing t relating to limit t, we
would see initial a dark picture, then white spots comparing to
nearby power minima will seem then become bigger.
These white spots will in the long run converge, until
the entire picture is white. The arrangement of every
associated segment in the succession is the arrangement of all
extremal locales. Optionally, circular edges are appended to
MSERs by fitting ovals to the districts. Those areas descriptors
are kept as elements. The word extremal alludes to the
property that all pixels inside the MSER have either higher
(brilliant extremal districts) or lower (dim extremal locales)
power than every one of the pixels on its external limit.
3.8 Methodology:
Text Recognition Phase
Step 1: Load Image
In this step firstly load the input image in which we
have to detect text. Before preceding towards next step first of
all the algorithm crop that portion of image that contains text,
Further the text can be rotated in plane, if required.
Step 2: Noise Removal and De-blurring Image
Because of defects in the imaging and catching
procedure, be that as it may, the recorded picture constantly
speaks to a degraded adaptation of the first scene. The
corruption results in picture blur, affecting identification and
extraction of the helpful data in the pictures. It can be brought
about by relative movement between the camera and the first
scene, by an out of center of optical framework, environmental
turbulences and deviations in the optical framework.
Lucy Richardson (LR) calculation is an iterative
nonlinear restoration method.The L-R calculation emerges
from most extreme probability plan in which picture is
displayed with toxic substance measurements. Its execution
within the sight of commotion is observed to be better than
that of other deconvolution calculations.
We can use other deblurring methods also like wiener
filtering.
Step 3: Contrast Adjustment and Conversion RGB image
to Binary Image
Picture upgrade strategies are utilized to enhance a
picture, where "enhance" is now and again characterized
dispassionately (e.g., build the sign to-commotion proportion),
and once in a while subjectively (e.g., make certain elements
less demanding to see by altering the hues or intensities)
Further in this progression RGB Image is changed over into
dim scale Image
Step 4: Edge Enhancement
In this progression, canny edge identification
calculation is utilized for picture edge discovery. The
calculation keeps running in 5 separate strides: Smoothing:
Blurring of the picture to evacuate clamor. Discovering slopes:
The edges ought to be checked where the inclinations of the
picture has extensive extents.
Non-most extreme concealment: Only nearby maxima ought
to be set apart as edges. Twofold thresholding: Potential edges
are controlled by thresholding. Edge following by hysteresis:

Volume: 5 Issue: 7 759 – 763
_______________________________________________________________________________________________
762
_______________________________________________________________________________________
Final edges are dictated by smothering all edges that are not
associated with an exceptionally certain (solid) edge.
To adapt to obscured pictures the propose calculation utilized
the properties of Canny edges.
Step 5: MSER region detection
As the power complexity of content to its experience
is regularly critical and a uniform force or shading inside each
letter can be expected, MSER is a characteristic decision for
content recognition. While MSER has been distinguished as
one of the best area identifiers because of its vigor against
perspective point, scale, and lighting transforms, it is delicate
to picture obscure. Along these lines, little letters can't be
recognized or recognized in the event of movement or defocus
obscure by applying plain MSER to pictures of constrained
determination.
3.9 Text Extraction Phase
Step 1 and 2: Geometric Filtering and Character
Connecting
With the extraction of edge-improved MSER, we get
a paired picture where the forefront CCs are considered as
letter hopefuls. As in most best in class content identification
frameworks, we play out an arrangement of basic and
adaptable geometric minds every CC to sift through non-
content items. As a matter of first importance, substantial and
little protests are rejected.
At that point, subsequent to most letters have angle proportion
being near 1, we dismiss CCs with extensive and little
viewpoint proportion. A moderate limit on the angle
proportion is chosen to ensure that some extended letters, for
example, "i" and "l" are not disposed of.
Step 3 & 4: Text line formation and Word separation
Content lines are imperative signs for the presence of content,
as content quite often show up as straight lines or slight bends.
To detect these lines, we first pair wise bunch the letter
competitors utilizing the accompanying principles. The
following phase of the calculation finds lines of content inside
the distinguished competitor districts. This permits the
aggregate number of CCs to be lessened, evacuating non-
character CCs and thus enhancing the odds for higher
exactness.
IV. COMPARISON
Connected component based method fails in some
natural scene images which have very poor contrast text and
strong illumination.
Table No. 1
Methods Accuracy
Advantage/Disadvantage
Texture based
Method
88.52% Inefficient when
background in the image
is more complex like
trees, vehicles.
Edge-based
method,
94.66%
Works on complex
backgro und. Fails for
small slanted/curved
text.
Morphology
operators,
Histogram
Projection ( X
and Y
histogram)
84.66% Fail in case of touching
characters and over-
lapping lines.
Maximum Color
Difference
(MCD),
Boundary
Growing
Method (BGM),
89.67%
Insensitive to contrast
Texture based techniques usually give better results
in complex backgrounds than region based techniques but
have computationally very heavy hence not suitable for
retrieval systems for hefty databases. Therefore, there is a need
to improve the detection results of region-based techniques to
be used for retrieval and indexing of large multimedia data.
V. CONCLUSION
This paper presents review on existing methods for text
detection, and recognition with their feature. Also this paper
summarizes the key ideas, advantages, disadvantages and
applications of text detection technique. Detecting and
recognizing text from natural scene image is more difficult
task than all other types of images. It has various affecting
factors like light effects, orientation, font styles, blur, etc.
Even though there are many algorithms, no single unified
approach can fits for all the applications. So there is lot of
scope to work with the text detection, extraction, segmentation
and recognition from natural scene images. Also there is scope
for detecting text from various.
REFERENCES
[1] Gao, Jiang, and Yang, Jie. An adaptive algorithm for text
detection from natural scenes. In Proceedings of the, “IEEE
Conference on Computer Vision and Pattern Recognition”
(December 2001)
[2] Robust Text Detection in Natural Scene Images Xu-Cheng
Yin, Member, IEEE, Xuwang Yin, Kaizhu Huang, and Hong-
Wei Hao, "IEEE Transactions on Systems”, june,(2008)

Volume: 5 Issue: 7 759 – 763
_______________________________________________________________________________________________
763
_______________________________________________________________________________________
[3] Gao, Jiang, Yang, Jie, Zhang, Ying, andWaibel, Alex. Text
detection and translation from natural scenes.Tech. Rep.
CMU-CS-01-139, ―Carnegie Mellon University, School of
Computer Science, Carnegie Mellon University, Pittsburgh”,
PA 15213, 2001.
[4] Text-Attentional Convolutional Neural Network for Scene
Text Detection Tong He, Weilin Huang, Member, ―IEEE”,
Yu Qiao, Senior Member, IEEE, and Jian Yao, Senior
Member, IEEE, (2007)
[5] B. Shiva Kumar Reddy, Lakshmi Boppana and Ashok
Agarwal. ‖BER Analysis of CVSD Vocoder for WiMAX
using GNU Radio‖. IEEE Region 10 Symposium, 2014 IEEE.
[6] Real-time text tracking in natural scenesbyCarlos Merino-
Gracia1,2, Majid Mirmehdi2 Neurochemistry and
Neuroimaging Laboratory, ―University of La Laguna‖ , La
Laguna, Spain, (2009)
[7] Gao, Jiang, and Yang, Jie. An adaptive algorithm for text
detection from natural scenes. In Proceedings of the , ―IEEE
Conference on Computer Vision and Pattern Recognition‖
(December2001)
[8] Gao, Jiang, Yang, Jie, Zhang, Ying, andWaibel, Alex. Text
detection and translation from natural scenes.Tech. Rep.
CMU-CS-01-139, ―Carnegie Mellon University, School of
Computer Science, Carnegie Mellon University, Pittsburgh‖,
PA15213,2001
[9] Zhang, Dong-Qing, and Chang, Shih-Fu. Learning to detect
scene text using a higher-order MRF with belief
propagation.In IEEE Workshop on Learning in Computer
Vision and Pattern Recognition vol. 06, pp. 101–108 (2004).
[10] Y.-F. Pan, X. Hou, and C.-L.Liu.Text localization in natural
scene images based on conditional random _eld. In ICDAR,
pages 6-10, IEEE Computer Society, (2009).
[11] Jain, A.K., and Bhattacharjee, S. Text segmentation using
Gabor filters for automatic document processing. Machine
Vision Applications 5 pages 169– 184, (1992).
[12] Wu, Victor, Manmatha, R., and Riseman, Edward M. Finding
text in images. In ―Proc. Intl. Conf. on Digital Libraries‖
(1997).
[13] Wu, Victor, Manmatha, R., and Riseman, Edward M. Finding
text in images. In ―Proc. Intl. Conf. on Digital Libraries‖
(1997).
[14] Thillou, Céline, Ferreira, Silvio, and Gosselin, Bernard. An
―embedded application for degraded text recognition.Eurasip
Journal on Applied Signal Processing‖ 13 (2005).
[15] Garcia, C., and Apostolidis, X. Text detection and
segmentation in complex color images. In ―Proc. Intl. Conf.
on Acoustics, Speech, and Signal Processing‖ vol. 4, pp.
2326–2330, (June 2000).
[16] Sergei Azernikov. Sweeping solids on manifolds. In
―Symposium on Solid and Physical Modeling‖, pages 249–
255,2008.
[17] John Canny. A computational approach to edge
detection.Pattern Analysis and Machine Intelligence, ―IEEE
Transactions on, PAMI‖-8(6):679–698, Nov. (1986).
[18] F. Mai, Y. Hung, H. Zhong, and W. Sze.―A hierarchical
approach for fast and robust ellipse extraction. Pattern
Recognition‖, 41(8):2512–2524, August (2008).
[19] David Gerónimo, Antonio López, and Angel D. Sappa
Computer Vision Center, ―University at Autònoma de
Barcelona Edifici O, 08193 Bellaterra, Barcelona‖, Spain,june
(2007)
[20] Piotr Doll ´ ar, Christian Wojek, Bernt Schiele, and Pietro
Perona, Submission to ―IEEE Transaction on pattern analysis
and machine intelligence‖, vol .1, page no. 1-19, (2012)
[21] X. Lin. Reliable OCR solution for digital content re-
mastering. In Society of ―Photo-Optical Instrumentation
Engineers (SPIE) Conference Series‖, (Dec. 2001).

Analysis and Comparison of various Methods for Text Detection from Images using MSER Algorithm

Recommended

More Related Content

What's hot (16)

Similar to Analysis and Comparison of various Methods for Text Detection from Images using MSER Algorithm (20)

More from rahulmonikasharma (20)

Recently uploaded (20)

Analysis and Comparison of various Methods for Text Detection from Images using MSER Algorithm