A Holistic Method to Recognize Characters in Natural Scenes
Muhammad Ali
and Hassan Foroosh
Department of Electrical Engineering & Computer Science, University of Central Florida, Orlando, Florida, U.S.A.
Keywords: Natural Scene Text Recognition, Active Contours, Holistic Character Recognition.
Abstract: Local features like Histogram of Gradients (HoG), Shape Contexts (SC) etc. are normally used by research
community concerned with text recognition in natural scene images. The main issue that comes with this
approach is ad hoc rasterization of feature vector which can disturb global structural and spatial correlations
while constructing feature vector. Moreover, such approaches, in general, don’t take into account rotational
invariance property that often leads to failed recognition in cases where characters occur in rotated positions
in scene images. To address local feature dependency and rotation problems, we propose a novel holistic
feature based on active contour model, aka snakes. Our feature vector is based on two variables, direction and
distance, cumulatively traversed by each point as the initial circular contour evolves under the force field
induced by the image. The initial contour design in conjunction with cross-correlation based similarity metric
enables us to account for rotational variance in the character image. We use various datasets, including
synthetic and natural scene character datasets, like Chars74K-Font, Chars74K-Image, and ICDAR2003 to
compare results of our approach with several baseline methods and show better performance than methods
based on local features (e.g. HoG). Our leave-random-one-out-cross validation yields even better recognition
performance, justifying our approach of using holistic character recognition.
1 INTRODUCTION
Recognition of text in natural scene context is a
challenging problem in computer vision, machine
learning and image processing. The importance to
solve this problem is equally compelling due to
abundant availability of digital cameras, esp. those in
mobile devices e.g. smartphones and wearable
glasses (e.g. Google Glass). There is a need of
applications like assisted navigation for visually
impaired people (e.g. OrCam
device mounted on
glasses, www.orcam.com) on one hand and mining
huge online image data repositories for textual
content (to automatically generate useful information
for marketing, archival, and other purposes etc.) on
the other hand. Other utilities of natural scene text
recognition include and automatic reading of
informational signs for automobile drivers or
driverless cars.
Although the problem of document text
recognition is a solved one, yet its counterpart in
natural scene scenario is far from being solved. This
is primarily due to the difficult and potentially
uncontrolled imaging conditions that occur while
imaging text in the wild. Hence the problem has been
broken down into the following four sub problems:
1. Cropped Character Recognition
2. Cropped Word Recognition
3. Scene Text Detection
4. Full-image Scene Text Recognition
There is another related problem proposed by (Wang
et al., 2011) to recognize words in an image given a
short vocabulary pertaining to the image.
To compare results, several public datasets are in
use, e.g. ICDAR2003 (Lucas et al., 2003), WLM
dataset (Weinman et al., 2009), Chars74K datasets
(de Campos et al., 2009), SVT (Wang et al., 2011),
NEOCR (Nagy et al., 2011), etc.
In this paper we address the sub problem 1 above
and use the two most popular datasets like
ICDAR2003 robust character dataset and the
Chars74K datasets.
Figure 1 shows sample images from Chars74K
and ICDAR datasets. The challenge is obvious from
the extent of shape variation and sample noise due to
distortions, illumination differences, object
occlusions, etc. Hence, we are confronted with all
sorts of structured and/or random noise.
Ali, M. and Foroosh, H.
A Holistic Method to Recognize Characters in Natural Scenes.
In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 449-457
ISBN: 978-989-758-175-5
Copyright
c
2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
449
Figure 1: Sample characters from Chars74K and ICDAR
datasets. Top row shows samples from synthetic Chars74K-
Font dataset.
The paper is organized as follows: Section 2
describes the related research work done on this
problem. Section 3 gives a brief look at active contour
model used in this paper. Section 4 describes steps in
our method to solve the problem. Section 5 presents
experimental setup, results, and discussion. Section 6
gives a recap of this paper along with future research
directions.
2 RELATED WORK
Since the introduction of ICDAR 2003 Robust
Reading Competition and the associated challenge
datasets (Lucas et al., 2003), the area of scene text
recognition has seen an increase in research efforts to
solve the problem. Various solutions have been
proposed for the sub problem of robust character
recognition.
Some researchers used off-the-shelf OCRs to
recognize characters. (Chen and Yuille, 2004) used an
adaptive version of Niblack’s binarization algorithm
(Niblack, 1985) on the detected textual regions and
then employed commercial OCRs for final
recognition. The reported results with ABBYY
(www.abbyy.com) were good for their dataset
(collected from cameras mounted on blind people.)
Later performance of ABBYY reported by (Wang
and Belongie, 2010) and (de Campos et al., 2009)
showed its poor performance on more challenging
ICDAR and Chars74K datasets.
Overall, the literature in natural scene character
recognition is dominated by local feature-based
methods: These methods mainly focus on extracting
a feature vector, e.g. a Histogram of oriented
Gradients (HoG) (Dalal and Triggs, 2005) or some
variant of it, from a character image and then using
some classifier, e.g. Nearest Neighbor, SVM etc., to
recognize the character. (de Campos et al., 2009) used
various feature descriptors including Shape Contexts
(SC), Scale Invariant Feature Transform (SIFT),
Geometric Blur (GB), etc. in combination with bag-
of-visual-words model. The results, however, left a
lot of room for improvement. (Weinman et al., 2009)
used a probabilistic framework wherein they utilized
Gabor filters in their similarity model to recognize
characters in their dataset. (Wang and Belongie,
2010) showed better performance than (de Campos et
al., 2009) by incorporating HoG features. (Neumann
and Matas, 2011) used maximally stable extremal
regions (MSER) to create MSER mask and then got
features along its boundary which they used in SVM
for classification. (Donoser et al., 2008) used MSERs
in conjunction with simple template matching to get
initial character recognition results which are
subsequently improved by exploiting web search
engines to get final recognition results. In addition,
unsupervised feature learning system has been
proposed by (Coates et al., 2011) that utilizes a
variant of K-means clustering to first build a
dictionary then map all character images to a new
representation in the dictionary. A recent holistic
approach based on tensor decomposition has been
proposed in (Ali and Foroosh, 2015) that takes into
account some problems with local feature based
methods but the authors perform ad hoc pre-
processing to account for rotation. Hence their
method relies on making a character image upright to
take care of rotated characters.
In this paper, we propose a holistic approach to
solve the problem and present a novel feature vector
based on active contour model. The use of active
contour models in shape recognition is common but
we are not aware of its application on scene text
recognition. The closest application we found was in
(Yi and Tian, 2014) where the authors establish
boundary points using discrete contour evolution
during the process of finding character polygons as a
first step in getting stroke configurations. Other than
this, their approach quite different from ours.
The results we get show that the proposed method
effectively captures character shape variations
occurring in natural scene images in a holistic manner
thus avoiding the problems associated with ad hoc
rasterisation of local image features. Our
contributions are:
1. Novel feature vector
2. Rotation invariance
Our results show that we perform better than several
popular baseline methods.
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
450
3 SNAKE: AN ACTIVE
CONTOUR MODEL
As described above, we are interested in holistic
recognition of characters extracted from natural scene
images. An active contour is a dynamic object (curve)
which evolves to wrap around an object boundary.
This idea of capturing object shape motivates us to
use it to extract holistic feature for our character
recognition problem.
Active contour models have been used for image
segmentation and shape description (Kass et al.,
1987; Xu and Prince, 1997; 1998) since long time. In
the following section, we first briefly recap the basics
and then move on to give its novel application to
deriving our holistic feature for characters synthetic
or extracted from natural scene images.
3.1 Basics
Consider the situation of a 2D image. An active
contour model can be described as a closed loop of
points or pixels. It is also called a snake for its
movement in image plane. Mathematically, a snake is
a set of points in the image plane. It can be
characterised by the following parametric curve:
,
(1)
where, and are the coordinates of pixels and is
the parameter. As described in (Ivins and Porrill,
2000), we can associate an energy functional with
this curve (note that the curve is a loop in this case).
|
|
|
′′
|
(2)
where,
is the external image energy (derived
mainly from image gradients) and α and β are the
internal curve parameters for tension and stiffness
respectively.
Assuming
= α, and
= β as constants, the
minimization of equation (2) can be done by
satisfying two independent Euler equations in :
(3)
Following (Ivins and Porrill, 2000), the derivatives in
(3) can be approximated by finite differences as
follows:
4
6
4
(4)
2
,
4
6
4
(5)
2
,
where
and
are the components of image force
(computed from gradients in ‘x’ and ‘y’ direction). To
solve (4) and (5), we use the semi-implicit method
discussed in (Ivins and Porrill, 2000) where two sets
of finite difference equations are formed to describe
the x and y coordinates of the entire snake; these
equations can be written in terms of a cyclic
symmetric pentadiagonal banded matrix
incorporating the constants α and β as follows:
.
, ;.
,
(6)
where, and are vectors containing the x and y
coordinates of all the snake elements;
and
are
the corresponding vectors of image forces acting on
contour points. We can solve these equations
iteratively by using a discrete and small time step ‘
’
(we set
= 1 for our experiments). The stiffness and
tension constraints are applied at time t+1 after
adjusting the snake according to the image forces at
time t:
.
,
(7)
.
,
(8)
The matrix inversion in the above equations (7 and 8)
is taken only once because they are composed of
constant terms. Hence, the above iterative
minimization provides for a fast way to solve the
equation (2).
We compute the external image energy,
, by
taking a weighted combination of three factors: lines,
edges, and corners. Thereafter, we compute the
effects of external forces on each point of the contour
by interpolation. The new coordinates of contour
points are computed from equations (7) and (8)
above. To further expand the reach of image forces
and let the snake enter concave regions, we also
utilize gradient vector flow as mentioned in (Xu and
Prince, 1998).
3.2 Feature Vector
As we evolve the snake around character images and
compute new coordinates of each point at each time
step of the above iterative minimization, we compute
the following:
∆
∆
∆
(9)
∆
1
∆
∆
(10)
where, ∆
and ∆
. The
quantities ∆
∆
represent the distance and
angle increments respectively for each point (pixel)
A Holistic Method to Recognize Characters in Natural Scenes
451
Figure 2: Illustration of our framework for feature vector extraction. Outer loop shows the location of pixels in the initial
contour and inner loops show the evolution of the initial contour after some iterations. (a) Contour evolution after 50 iterations
(b) After 125 iterations (note the wrapping of contour around the character (c) Final contour after 200 iterations (d) Enlarged
view of the points and computation of distance and angle increments.
on the contour at time instant . We accumulate the
increments over the course of evolution for all points
and concatenate to form our descriptor as follows:
∑
∆
1
.
.
∑
∆
∑
∆
.
.
∑
∆
(11)
where, is the final feature vector, and is the
number of iterations of the contour evolution. The
size of the above feature vector is equal to twice the
number of contour points (i.e., 2 =1252
250 in our experiments).
The rotational invariance property of the above
feature vector follows from its design (the initial
contour is circular) and its use in conjunction with the
cross-correlation similarity metric.
4 SCENE CHARACTER
RECOGNITION
Our framework for scene character recognition starts
with pre-processing images, followed by training
where we generate feature vector using active contour
evolution, as depicted in Figure 2, for all training
samples, and finally classification by getting the
feature vector for each test image and computing
similarity using cross-correlation.
4.1 Pre-processing
The cropped character images from natural scene
datasets contain a lot of non-character structures and
imperfect cropping artefacts which makes it difficult
to effectively capture typeface and shape variations.
Since the focus of our work is to demonstrate
effectiveness of holistic recognition framework based
on active contour feature, we pre-process each image
in training and testing sets to keep the images as noise
free as possible. To this end we adopt binarization for
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
452
image segmentation to reduce noise and extract,
possibly only, character structures. This somehow
lets us isolate the classification problem from the
binarization problem.
4.1.1 Image Segmentation & Normalization
Segmentation of textual information from natural
scene images is a challenging problem due to noise
and distortions introduced mainly by uncontrolled
imaging conditions. Many researchers have
attempted to tackle it, e.g. see (Chen and Yuille,
2004); (Mishra et al., 2011); (Kita and Wakahara,
2010); (Field and Learned-Miller, 2013).
Since the focus of our paper is character recognition,
we adopt and extend the method in (Ali and Foroosh,
2015) in an effort to segment each image to get the
correct textual foreground (in white). To this end, we
obtain two binary images: one from the output of
Otsu’s method and the other by inverting it. At this
point, we do a simple analysis of the skeleton by
counting number of pixels of both images to get the
correct segmented image. We then perform a
connected component analysis based on the
observation that cropped characters mostly fall in the
middle of the image. We consider any small pixel
group as noise if its size is less than a small fraction
(<5%) of the size of the largest central connected
component. The process is shown as a flow chart in
the Figure 3.
Figure 3: Segmentation of character images.
Figure 4: Sample binarized character images.
After segmentation, we normalize each image to
have the size of 32x32 pixels. To make sure that curve
evolution doesn’t start too close to the character
boundary, we pad the image with a 16x16 frame of
zeros. Hence the final size of the image becomes
64x64.
4.2 Training
For training, we take individual images of each class
and pre-process them. For each image, we envelope
the character with a circular contour (radius of
contour is fixed at 30 pixels) centred on the image and
sampled uniformly with 125 points. The contour is
then evolved towards the character and for each point
on it, we accumulate direction (angle) and distance,
until the character boundary is reached. The two
quantities (direction and distance) are then
concatenated and normalized to form a feature vector
containing 250 elements.
Figure 2 illustrates the process of extracting the
feature vector for a random image in the training
process. The sensitivity of results for different
parameter settings is discussed in (Section 5.3.1).
4.3 Classification
To classify a test image, we first pre-process it and
then compute its feature vector by evolving a circular
contour towards it. We then measure similarity of the
test vector with each of the training vectors using the
cross-correlation metric and recording the maximum.
The final classification is given by taking the
maximum over all classes.
5 EXPERIMENTS
We evaluated our approach using various character
datasets. This includes synthetic dataset Chars 74K-
Font as well as popular natural scene character
datasets Chars74K-Image and ICDAR. We
performed experiments using different settings to
report our results on the datasets. We also compare
our method with several baseline methods.
A Holistic Method to Recognize Characters in Natural Scenes
453
5.1 Datasets
The Chars74K-Font is a synthetic dataset of English
alphabet generated in various typefaces for 62
character classes: ‘A’ to ‘Z’, ‘a’ to ‘z’, and digits ‘0’
to ‘9’. The dataset consists of 62,992 images with
1,016 images per class.
The English subset of Chars74K dataset consists
of 12,503 characters. Characters have been cropped
from 1,922 images of advertisement signs and
products from stores etc. The dataset is not split in
training and testing sets, rather the authors (de
Campos et al., 2009) give their proposed training and
testing splits for comparison with their results. There
is, however, a split between ‘GoodImg’ and ‘BadImg’
and as obvious from the names, the respective splits
contain ‘good’ and less noisy (7705 images) as well
as ‘bad’ more noisy images (4798 images) for a total
of 12,503 images.
The ICDAR2003 robust character dataset
contains 11,615 images of cropped scene characters
and the dataset comes split into training and testing
subsets. Characters have mostly been cropped from
images of books titles, storefronts and signs and
exhibit great variability in terms of resolution,
illumination, colour, etc. The test set has 5340 images
in total but those belonging to 62 classes (A-Z, a-z,
and 0-9) are just 5,379.
5.2 Results
For all experiments, we fixed the parameters alpha
and beta to 0.05, the number of iterations to 200, the
radius of initial contour to 30 pixels. This parameter
setting yield good results across all datasets. Further
discussion on this setting is deferred until Section
5.3.1.
In the first experiment, we used Chars74K-Font
synthetic font dataset. We randomly picked 15
samples each for training and testing to make fair
comparison with the results of de Campos et al. The
results are shown in Table 1. The second column of
Table 1 presents interesting results when training on
synthetic fonts and testing on Chars74K-Image test
split proposed in de Campos et al. Here also, we show
better performance than the reported method.
In our second experiment, we used the whole
ICDAR2003 training set to get features for each class
of characters. The accuracy on the test set was 62%
(see Table 2).
(Wang et al., 2012) reported accuracy of 83.9%
on a modified version of the ICDAR2003 test set, but
they re-cropped all images for their experiments and
their set contains 5198 images, which is less than
those in ICDAR2003 test set. Hence, their results are
not compared here.
In Figure 5, the lines parallel to the main diagonal
of the confusion matrix reflect ambiguities due to
character case, e.g. small case ‘c’ confused with ‘C’,
etc. In Table 2, we also report our results on the
training and test splits proposed by (de Campos et al.,
2009) for Chars74K, viz., Chars74K-15, where the
suffix ‘15’ specifies the number of training and test
samples to be used for the experiment.
Table 1: Character Recognition performance on Chars74K-
Font dataset & Chars74K-15 Test Split.
Method Chars74K-Font
Chars74K-15
Test Split
GB+NN
(de Campos et al., 2009)
69.71% 47.16%
Proposed Method 71% 56%
Figure 5: Confusion matrix for ICDAR2003 test set
Numbers 1-62 show character classes A-Z, a-z,0-9. Lines
parallel to the main diagonal show character confusions.
Table 2: Character Recognition performance on
ICDAR2003 and Chars74K-15 datasets.
Method ICDAR Chars74K-15
GB+NN
(de Campos et al., 2009)
41% 47.09%
HoG+NN
(Wang and Belongie, 2010)
51.5% 58%
SYNTH+FERNS
(Wang et al., 2011)
52% 47%
NATIVE+FERNS
(Wang et al., 2011)
64% 54%
Stroke Config.
(Yi and Tian, 2014)
62.8% 60%
Proposed Method 62% 59%
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
454
Our third experiment was to test the performance
of Chars74K-15 test split on a modified training set
(using all training samples but those in the test split)
as per de Campos et al., The results are reported in
Table 3.
Table 3: Recognition Performance on Chars74K-15 Test
Split.
Method Chars74K-15 Test Split
ABBYY FineReader
(www.abbyy.com)
31%
GB+NN
(de Campos et al., 2009)
54.3%
Proposed Method 61.5%
The difference in results of Table 2 (2
nd
column)
and Table 3 clearly show that the number of training
samples in the former (15 in this case) is not sufficient
to capture the variation in test samples. Hence, we
were prompted to do another experiment with leave-
random-one-out cross-validation (CV) setting. We
show results in Table 4 for both Chars74K and
ICDAR2003.
For ICDAR2003 we combined training and
testing sets to get one big set for CV. For Chars74K,
as mentioned before, the data already comes without
training and testing splits. The results in Table 4 show
median accuracy over 100 trials.
Table 4: Recognition Performance using leave-random-
one-out cross-validation (CV).
Method ICDAR Chars74K
Proposed Method + CV 65% 62%
Figure 6: Some test samples from different datasets that our
approach correctly recognized.
Figure 7: shows the cases where our method failed. Some
images here are not even easy human observers to recognize
correctly due to low contrast, shape ambiguities, noise etc.
5.3 Discussion
5.3.1 Parameter Sensitivity
In Figure 8, we show how the accuracy changes with
respect to the snake’s parameter values. The
parameters α and β are internal smoothness
coefficients and control how the contour behaves as it
evolves over the given generations. .Here we assume
α=β.
To estimate the values of the parameters, we
perform experiment on the Chars74K-Font dataset
and compute accuracy over different values of α and
β spread over an arithmetic scale from 0.001 to 0.01.
We found that the best accuracy occurs at α=β = 0.05.
Hence, we use this value throughout our experiments.
The other factors influencing the results are noise.
Although we segment the text in pre-processing step,
yet we find that in many cases the reason of our
system not performing well is the noise. We see that
performance improvement can be achieved if a more
elaborate algorithm is used for noise removal.
Figure 8: Snake internal parameters’ sensitivity estimated
on Chars74K-Font dataset. We assume parameters α=β in
this case and the best value occurs at α = β = 0.05.
A Holistic Method to Recognize Characters in Natural Scenes
455
5.3.2 Runtime Analysis
The runtime of our approach can be estimated from
the constituent processes namely, preprocessing and
feature extraction. The preprocessing phase depends
on Otsu binarization, and morphological
skeletonizing operators, which are in general
,
where is the size of the image. As regards the
feature extraction phase, we deal with contour
points laid out in a circle. Each movement of the
contour depends on computing external and internal
forces acting on each point. Luckily, the iterative
optimization method used in modelling snake
evolution uses just one computation (involving
matrix inversion) of internal force matrix whose size
depends on . External image forces involve
computation of image gradients in horizontal and
vertical directions and need
operations in one
pass. Finally, we iterate over evolution steps to get
the snake to its terminal shape. Since, and are
fixed prior to running the algorithm and is usually
very small compared with the size of image, the total
cost turns out to be
.
6 CONCLUSIONS
In this paper we put forth a novel feature to
holistically solve natural scene character recognition
problem that avoids dependency on specific features.
Through our results we showed the potential of using
our novel feature to better capture shape and font
variations in scene character images. We got better
results than several baseline methods and achieved
improved recognition performance on the datasets
using leave-random-one-out cross-validation,
showing the importance of feature-independency and
preservation of spatial correlations in recognition.
In future we hope to get state-of-the-art
performance using better image segmentation
methods and also optimizing other parameters of
contour evolution. We also look forward to using
contour evolution on grayscale images directly.
REFERENCES
Ali, M., and Foroosh, H., 2015. Natural Scene Character
Recognition without Dependency on Specific Features.
In Proceedings of the International Conference on
Computer Vision Theory and Applications (VISAPP),
Berlin, Germany, March 2015.
Chen, X., and Yuille, A., 2004. Detecting and reading text
in natural scenes. In Computer Vision and Pattern
Recognition, 2004. CVPR 2004. IEEE 2004. Vol. 2. pp.
II–366.
Coates, A., Carpenter, B., Case, C., Satheesh, S., Suresh,
B., Wang, T., Wu, D., and Ng, A., 2011. Text detection
and character recognition in scene images with
unsupervised feature learning. In International
Conference on Document Analysis and Recognition
(ICDAR), 2011. IEEE 2011, pp. 440–445.
Dalal, N., and Triggs, B., 2005. Histograms of oriented
gradients for human detection. In International
Conference on Computer Vision and Pattern
Recognition (CVPR) 2005. IEEE 2005, pp.886-893.
de Campos, T. E., Babu, B. R., and Varma, M., 2009.
Character recognition in natural images. In Proceedings
of the International Conference on Computer Vision
Theory and Applications (VISAPP), Lisbon, Portugal,
February 2009. pp. 273–280.
Donoser, M., Bischof, H., and Wagner, S., 2008. Using web
search engines to improve text recognition. In 19th
International Conference on Pattern Recognition,
ICPR 2008. Vol. no. 14, pp. 8-11.
Hazan, T., Polak, S., and Shashua, A., 2005. Sparse Image
Coding using a 3D Non-negative Tensor Factorization.
In International Conference on Computer Vision
(ICCV), 2005. IEEE 2005. Vol. 1, pp. 50–57.
Field, J., and Learned-Miller, E., 2013. Improving Open-
Vocabulary Scene Text Recognition. In International
Conference on Document Analysis and Recognition
(ICDAR) 2013. IEEE 2013, pp. 604–608.
Ivins, J., and Porrill J., 2000. Everything you always wanted
to know about snakes. AIVRU Technical Memo 86, July
1993 (Revised June 1995; March 2000).
Kass, M., Witkin, A., and Terzopoulos, D. 1987. Snakes:
Active contour models. International Journal of
Computer Vision. v. 1, n. 4, pp. 321-331.
Kita, K., and Wakahara, T., 2010. Binarization of color
characters in scene images using k-means clustering
and support vector machines. In International
Conference on Pattern Recognition (ICPR), 2010.
IEEE 2010, pp. 3183–3186.
Lucas, S. M., Panaretos, A., Sosa, L., Tang, A., Wong, S.,
and Young, R., 2003. ICDAR 2003 robust reading
competitions. In Proceedings of the Seventh
International Conference on Document Analysis and
Recognition 2003. IEEE 2003. Vol. 2, pp. 682–687.
Mishra, A., Alahari, K., and Jawahar, C., 2011. An MRF
model for binarization of natural scene text. In
International Conference on Document Analysis and
Recognition (ICDAR), 2011. IEEE 2011, pp. 11–16.
Nagy, R., Dicker, A., and Meyer-Wegener, K., 2011.
NEOCR: A Configurable Dataset for Natural Image
Text Recognition.
In CBDAR Workshop, ICDAR 2011,
pp. 53-58.
Neumann, L., and Matas, J., 2011. A method for text
localization and recognition in real-world images. In
Computer Vision–ACCV 2010, pp. 770–783.
Niblack, W., 1985. An introduction to digital image
processing. Strandberg Publishing Company.
Otsu, N., 1979. A Threshold Selection Method from Gray-
Level Histogram. In Trans. System, Man and
VISAPP 2016 - International Conference on Computer Vision Theory and Applications
456
Cybernetics. IEEE 1979. Vol.9, pp.62-69.
Wang, T., Wu, D., Coates, A., and Ng, A., 2012. End-to-
End Text Recognition with Convolutional Neural
Networks. In International Conference on Pattern
Recognition (ICPR), 2012. IEEE 2012, pp. 330.
Wang, K., Babenko, B., and Belongie, S., 2011. End-to-end
scene text recognition. In International Conference
Computer Vision (ICCV), 2011. IEEE 2011, pp. 1457–
1464.
Wang, K., and Belongie, S., 2010. Word spotting in the
wild. In Computer Vision–ECCV 2010, pp. 591–604.
Weinman, J., Learned-Miller, E., and Hanson, A., 2009.
Scene text recognition using similarity and a lexicon
with sparse belief propagation. In Pattern Analysis and
Machine Intelligence TPAMI. IEEE Transactions 2009.
Vol. 31, no. 10, pp. 1733–1746.
Xu, C., and Prince, J. L., 1998. Snakes, Shape, and Gradient
Vector Flow. IEEE Transactions on Image Processing,
1998.
Xu, C., and Prince, J. L., 1997. Gradient Vector Flow: A
New External Force for Snakes. In Proc. IEEE Conf. on
Comp. Vis. Patt. Recog. (CVPR), Los Alamitos: Comp.
Soc. Press, pp. 66-71.
Yi, C., and Tian, Y., 2014. Scene text recognition in mobile
applications by character descriptor and structure
configuration, IEEE Trans. IP, pp 2972-2982, 2014.
A Holistic Method to Recognize Characters in Natural Scenes
457
