A FAST AND EFFICIENT METHOD FOR CHECK IMAGE
QUALITY ASSESSMENT
Raju Gupta, Tavneet Batra, Pramod Kumar and Dinesh Ganotra
Newgen Software Technologies Limited, A-6, Satsang Vihar Marg
Qutab Institutional Area, New Delhi–110067 India
Keywords: Check truncation, check quality, check usability, check scanner, image exchange, substitute check.
Abstract: With the enactment of check 21 Act, check image quality has become a critical requirement. Banks
responsible for capturing check images (check truncation) have to warrant their images. With a plethora of
capturing devices and outsourcing of check acquisition process, assurance of check quality becomes
complex. Currently, banks deploy separate subsystem for Image Quality Analysis (IQA), which is based on
defect metrics defined by Financial Services Technology Consortium (FSTC) Phase I project (Image quality
and usability assurance, 2004). The problem with this approach is that IQA cannot match the scanning speed
and has to be deployed as a separate process. Another problem with a predefined defect metrics is that it is
dependent on check content. This paper proposes a fast and efficient method to estimate the quality and
usability of check images. The method is independent of the check content or layout. IQA based on this
algorithm can be deployed at the scanning stage. The checks will have a pre-printed pattern in the form of a
logo. This pattern will be detected and analysed for quality and usability. The results show that our
algorithm is able to sort unusable check images efficiently. In future, we plan to use this pre-printed pattern
as a measure of check security.
1 INTRODUCTION
The Check Clearing for the 21st Century Act (Check
21) (Advetorial, 2005; Check Clearing for the 21st
Century Act, 2003) is designed to promote
innovation in the payments system and to enhance
its efficiency by reducing the legal impediments to
check truncation. The law facilitates check
truncation by creating a new negotiable instrument
called a substitute check, which permits banks to
truncate original checks, to process check
information electronically and to deliver substitute
checks to banks that want to continue receiving
paper checks.
The processing steps involved in a clearing
system before Check 21 Act were tied to possession
of physical checks by financial institutions. This
meant physical checks had to be moved at each step
in the clearing process, taking its toll on resources in
terms of manpower and time (Check clearing, 2003).
While on one hand, Check 21 presents new
opportunities to the banking industry to benefit from
check image exchange, on the other it poses new
challenges in the form of check image quality
assurance and security. Banks responsible for
capturing check images will have to ensure the
quality of images.
FSTC (Image quality and usability assurance,
2004) had taken the check quality and assurance
initiative and launched two projects - first, that
identifies and defines defect metrics for quality
assessment; and second, that defines a framework
for assessing the usability of an image. The check
quality was determined by matching 16 quantifiable
parameters against predetermined thresholds. Check
quality alone cannot assure the usability of a check.
In the second project, it was found that out of 16
parameters the two most significant parameters that
determine check usability are image ‘too light’ and
‘too dark’.
There are two problems using the FSTC
approach (Wang et al., 2002). First, determination of
the 16 parameters for every check is time consuming
and therefore cannot be integrated with scanning.
This warrants the need for a separate IQA phase.
However, this increases check-processing time and
requires more resources to complete the job. Second,
thresholds for different defect parameters are
109
Gupta R., Batra T., Kumar P. and Ganotra D. (2007).
A FAST AND EFFICIENT METHOD FOR CHECK IMAGE QUALITY ASSESSMENT.
In Proceedings of the Second International Conference on Computer Vision Theory and Applications - IFP/IA, pages 109-114
Copyright
c
SciTePress
generic; they do not vary for different kinds of
checks. So, a check with very dark background
might get rejected on the count that the image is too
dark and similarly a check image devoid of
background may get rejected on count of image
being too light.
In this paper, we have proposed how to
overcome the above-mentioned problems and also
suggested a method for providing a security feature
for the checks. It is achieved by designing a pattern
that will be pre printed on the checks. Analysis of
this pattern will be fast owing to its small size, and
since characteristics of this pattern are known in
advance we can define global thresholds for defect
metrics. We have chosen two metrics as proposed by
FSTC Check Usability report (Image quality and
usability assurance, 2004), image ‘too dark’ and ‘too
light’.
Image degradation results due to various factors
like the variable adjustment controls of the
acquisition devices like, brightness, contrast,
focusing, uneven illumination, camera calibration
etc. (Smith, 1998). Figure 1 shows the changes in the
gray levels that take place in the acquired image
with respect to the original document.
Figure 1: Straight line represents the gray levels varying
linearly from 0 to 255. The curved plot represents the
change in gray levels after image acquisition.
2 PROPOSED APPROACH
The proposed method is implemented in two parts.
The first part enfolds the designing of the logo,
which is incorporated at the printing stage. The
second part pertains to the assessment of the quality
parameters, which is integrated with image
acquisition phase.
For an effectual solution, a number of factors
need to be considered while designing the pattern.
Since the overall quality assessment of the image is
benchmarked using the logo, it should be printed at a
convenient location on the check. An image of the
Figure 2: Block diagram of the proposed method.
check is then acquired using an image-
acquisition device. The position of the logo can be
determined using approaches like registration marks
(Tseng and Chen, 1998; Turolla et al., 1997). The
pattern or logo image should be designed in a way
that it can accurately gauge the usability factors like
‘too light’ and ‘too dark’ for the image.
Our method, shown in Figure 2, is used to
correctly evaluate the quality of the acquired image
by leveraging the fact that the improper brightness
and contrast introduced by image acquisition device
will always lead to degraded quality of the image
(Lin and Chang., 1999). Our method, based on the
logo image, gives a consolidated quality factor by
computing different image-quality parameters
(Wang et al., 2002). This quality factor is reliable
(Zhang and Zhang, 2004), as it is computed by a 3-
dimensional geometrical analysis and not using any
semantic approach. The calculated factor is
compared against a pre-calculated threshold, and the
image is categorized for its usability. The value of
the pre-calculated threshold is obtained on the basis
of analysis of data set relative to different user
environments.
Phase - I
Ima
g
e ac
q
uisition
Logo Analyzer
Calculation of Quality factor from Logo
Ima
g
e Characteristics
Comparison with
predetermined
threshold
High Quality
Phase - II
+
Pre-printed checks with logo
Low Quality
VISAPP 2007 - International Conference on Computer Vision Theory and Applications
110
2.1 PHASE–I (Logo Definition)
Two images with concentric circles of varying gray
levels are printed on the check. The region outside
the circumference has a uniform gray scale of 0.5 on
a normalized scale of 0 to 1. The first image (Figure
3(a)) is drawn with non-linearly increasing gray
levels from circumference towards center of the
circle and the second image (Figure 3(b)) is drawn
with non-linearly decreasing gray levels from
circumference towards center of the circle. The non-
linearity is chosen such that the pixel co-ordinates
and the gray levels define a 3-D sphere in cartesian
system. When a hypothetical sphere is drawn with
same center and radius as that of circle, the gray
level corresponds to the surface of the sphere from
the center.
(a)
(b)
Figure 3: (a) Upper hemisphere and (b) Lower
hemisphere.
The two-dimensional circular image on the paper
is treated as three-dimensional hemi-sphere with x-
coordinate as one dimension, y-coordinate as second
dimension and the gray level at that (x, y) location as
the third dimension. The two hemi-spheres when
combined together form a sphere shown in Figure 4.
Figure 4: Three-dimensional-mesh simulation of two
hemispheres.
2.2 PHASE–II (Calculation of Quality
Factor)
H
u
: Upper hemispherical region of the logo pattern
in the check image to be analyzed.
H
l
: Lower hemispherical region of the logo pattern
in the check image to be analyzed.
S
o
: Sphere formed by combining the upper
hemispherical region and lower
hemispherical region embedded in the
original logo image.
+
u
η
: Count of all pixels outside the periphery of
hemisphere H
u
(shown as shaded area in
Figure 5(a))
.
u
η
: Count of all pixels on the periphery of
hemisphere H
u
.
+
u
K
: Set of
+
u
η
pixels in H
u
u
K : Set of
u
η
pixels in H
u
The value of variables
+
l
η
,
l
η
,
+
l
K and
l
K for
the lower hemisphere H
l
can be calculated similarly.
The average gray level between the bases of H
u
and H
l
is determined. On check image it is the
average gray level of the two regions (shown as
shaded region in Figure 5(a) and (b)).
Figure 5: Schematic diagram of (a) Upper and (b) Lower
hemisphere. The shaded area depicts the common base of
two hemispheres, which lies outside the sphere.
Consider ‘r’ as the radius of S
o
and I(x,y) as the
gray level value at any (x, y) .
The value
γ
p
is computed as
(1)
+
=
+
=
u
i
p
u
u
i
ηγβ
η
/)(
1
(2)
I(x, y), if (x, y)
ε
+
u
K
γ
p
=
{
0, otherwise
(b)
(a)
Shaded Region
A FAST AND EFFICIENT METHOD FOR CHECK IMAGE QUALITY ASSESSMENT
111
+
=
+
=
l
i
p
l
l
i
ηγβ
η
/)(
1
(3)
where
u
β
is the average base value of H
u
and
l
β
is
the average base value of H
l
.
The distance
δ
of each pixel coordinate (x, y)
ε
lu
KK
is calculated as.
)),(),(()()(
22
lu
meanyxIryrx
ββδ
++=
lu
KKyx ),(
(4)
The standard deviation σ
of all the
δ
values
about ‘r’ gives the quality factor of the image
(Ivkovic and Sankar, 2004). σ is inversely
proportional to the quality factor of the image.
Figure 6: (a) Deviation of gray level I(x,y) of each pixel of
H
u
from ‘r’. (b) Deviation of gray level I(x,y) of each pixel
of H
l
from ‘r’.
The value of σ when compared with the pre-
calculated threshold correctly identifies the image
quality. The threshold is calculated by
experimenting with different sets of image data. The
data set should contain check documents with
different backgrounds, contents and layout, and the
images should be acquired using different scanners,
photocopiers, etc. The calculated threshold can serve
as a reference for users who do not want to re-train
the system for the specific environment.
It has been observed by experimenting on
different quality sets of images that σ
obtained for
poor-quality images is high as compared to good
quality images. (Figure 6 and Table 1).
3 BENEFITS
The proposed approach offers a potential gain in the
efficiency of the entire system. By integrating the
proposed solution at the point of capture, a
significant reduction in time can be achieved. This is
shown in Figure 7.
Figure 7: (a) A schematic diagram indicating the total time
taken by the existing image acquisition and quality
analysis phase. (b) A schematic diagram indicating the
deployment of the proposed solution at the point of
capture and reduction in time.
In the experiments conducted it was found that
the average time to scan a check is 1 second. When
run on batch of 612 check images the IQA solution
takes 0.6 seconds average to analyze the quality of
one check image. For a case where 80% of checks
are accepted during the first pass and 20% checks re-
scanned and accepted during the second pass the
total time taken to analyze the batch is {612 * (1 +
0.6) + 122 * (1 + 0.6) =} 1174.4 seconds.
(a)
(b)
(b)
T1(n)
Checks with
rinted lo
o
Rejected checks re-scanned
+
Proposed Solution
Sorting
Checks scanned
at 60 ppm
(a)
Checks with
rinted lo
o
Checks
scanned at
60 ppm
IQA
Solution
Sorting
Rejected checks re-scanned
T1(n)
T2(n)
+
T2 > T1
VISAPP 2007 - International Conference on Computer Vision Theory and Applications
112
Figure 8: Logo pattern of acquired check images and the corresponding 3–D simulation of embedded object with
calculated σ. As indicated above σ for good-quality images is lesser as compared to poor - quality images.
On the other hand, the proposed solution (Figure
7(b)) takes average 0.18 seconds to determine the
quality of a check. Hence for a batch size of 612
images and the rejection statistics as above the total
time taken to analyze the batch is {612 * (1 + 0.18)
+ 122 * (1 + 0.18) =} 866.12 seconds. Although this
time is dependent on factors like scan resolution,
pixel depth and machine configuration, however, as
the rejection rate during the first pass decreases the
gain in the performance increases.
By analyzing just a part of the check image, the
model is more efficient than other IQA solutions that
analyze the entire check content. Also, the method
being independent of check contents is not limited
by variation in check templates.
4 EXPERIMENTAL RESULTS
In this section we present the experimental results of
testing our algorithm on different check images. The
data consists of checks having different content,
layout and image statistics. Also the images are
acquired using different image acquisition devices.
Table 1 shows 15 observations from the data set of
612 images. Figure 9 shows the plot of σ calculated
from images of the data set. Figure 8 shows a set of
2 logo images detected from different check images.
As depicted in the corresponding 3-D simulations of
the embedded pattern, the shape of the object
deteriorates for bad quality images. To assess the
tolerance in the change of shape of object threshold
is calculated by providing supervised learning to the
system on different sets of sample data. We were
successfully able to classify the quality of 86% of
the images in the data set based on the calculated
threshold.
Table 1: Results of analysis on different image sets.
Image
Index
Standard
deviation, σ
Computation
time (in secs.)
I
a
3.15 0.17
I
b
3.01 0.22
I
c
3.38 0.25
I
d
2.82 0.23
I
e
12.18 0.25
I
f
14.07 0.25
I
g
10.35 0.26
I
h
8.58 0.22
I
i
3.40 0.25
I
j
2.81 0.22
I
k
19.11 0.26
I
l
3.52 0.20
I
m
8.40 0.25
I
n
7.89 0.20
I
o
2.88 0.23
(a)
σ = 4.12
(b)
σ =10.59
A FAST AND EFFICIENT METHOD FOR CHECK IMAGE QUALITY ASSESSMENT
113
Figure 9: σ plotted for the data set of 612 images. Images
with σ above the threshold have been rejected on the count
of bad quality.
5 DISCUSSION AND FUTURE
SCOPE
In order to perform a robust analysis of the quality
level of the check image, the initial exercise to
calculate the predetermined threshold should include
data specific to the settings of the check scanner.
Some applications may require the check documents
to be scanned using a particular scanner or may
require documents to be printed on a particular
printer. If the model is trained in a specific
environment, this will enable us to assess the quality
more accurately. The more extensive the initial
exercise done to find the predetermined threshold,
more accurate the result is.
Proposed assessment method can be further
improved by adding more features to the logo image
that are sensitive to the acquisition process. Features
can also be incorporated that can serve as security
marks for the authorization of instruments (Wang et
al., 2006). Authentication process that helps to
distinguish a bona-fide copy from a forged one can
be constructed around the said pattern where in a
number of parameters can be extracted from
multiple images of the authentic document.
Individual thresholds can be calculated for each
parameter and these thresholds can be compared
with the parameters obtained from the document
presented for certification.
REFERENCES
Advetorial, 2005. Banking Frontiers Magazine, Cheque
Truncation, Published by Glocal Strategies and
Services. Vol. 4, No. 6, Page 9, September 2005.
Check Clearing for the 21st Century Act, October 28,
2003. Retrieved November 1, 2006 from
http://www.federalreserve.gov/paymentsystems/trunca
tion/default.htm.
Image quality and usability assurance initiative phase 1,
July 26, 2004. Financial Services Technology
Consortium. Retrieved November 1, 2006, from
http://www.fstc.org/projects/image_quality_phase_1/i
ndex.php
Ivkovic, G., Sankar, R., 2004. An algorithm for image
quality assessment. In proceedings of ICASSP '04,
IEEE International Conference on Acoustics, Speech,
and Signal Processing., Vol: 3, pp. 713-716.
Lin, C.Y., Chang, S.F., 1999. Distortion Modeling and
Invariant Extraction for Digital Image Print-and-Scan
Process. In ISMIP’99, Intl. Symposium on Multimedia
Information Processing, Taipei, Taiwan, Dec. 1999.
Smith, E. H. B., 1998. Characterization of Image
Degradation Caused by Scanning. In Pattern
Recognition Letters, Vol. 19, No. 13, 1998, pp. 1191-
1197.
Tseng, L.Y., Chen, R.C., 1998. Recognition and Data
Extraction of Form Documents Based on Three Types
of Line Segments. In Pattern Recognition, Vol. 31,
No.10, pp. 1525-1540.
Turolla, E., Belaid, Y., Belaid, A., 1997. Form Item
Extraction Based on Line Searching. In Graphics
Recognition--Methods and Applications (Lecture
Notes in Computer Science), 1997.
Wang, Z., Bovik, A.C, 2002. A universal image quality
index. In Signal Processing Letters, IEEE, Vol.9,
No.3, pp. 81-84.
Wang, Z., Bovik, A. C., Lu, L., 2002. Why is image
quality assessment so difficult? In Proc. IEEE
International Conference on Acoust., Speech, and
Signal Processing, vol. 4, (Orlando), pp. 3313-3316,
May 2002.
Wang, Z., Wu, G., Sheikh, H. R., Simoncelli, E.P., Yang,
E., Bovik, A.C., 2006. Quality-aware images. In IEEE
Transactions on Image Processing, Volume: 15 ,
Issue: 6, pp. 1680-1689.
Zhang, F., Zhang, H., 2004. Digital Watermarking
Capacity and Reliability, In IEEE International
Conference on E-Commerce Technology (CEC'04),
July 2004, pp. 295-298.
VISAPP 2007 - International Conference on Computer Vision Theory and Applications
114