IEEE International Symposium on Industrial Electronics (ISlE 2009)

Seoul Olympic Parktel, Seoul, Korea July 5-8, 2009--

Fast Image Blurring Using Lookup Table for Real
Time Feature Extraction
Jegoon RYU
Graduate School of Information,
Production and Systems
Waseda University
Fukuoka, Japan 808-0135
Email: [email protected]
Abstract- In this paper, a novel fast Gaussian blurring method
that uses Look-Up Table (LUT) is presented for creating a
real time Difference of Gaussian (DoG) pyramid. The LUT for
fast and accurate Gaussian blur is obtained from the values
calculated from one-dimensional(lD) Gaussian elements and
pixel values of image. To evaluate the method, computation time
and similarity with a conventional two-dimensional(2D) Gaussian
blur are measured, and compared to other blurring methods.
Based on the computational results, the proposed method shows
good performance than other blurring methods such as box filter,
Stack-blur, Romain Guy's method, and recursive Gaussian blur.
The proposed method could be effectively applied to create the
DoG pyramid for feature extraction.
I. INTRODUCTION
Image blurring (or smoothing) is one of the most funda-
mentally used operations in image processing. There are well-
known linear and non-linear blurring methods such as conven-
tional 2D Gaussian filter and median filter. They are used as
parts of more complex algorithms including feature extraction,
noise reduction, and segmentation. Lowe [1] applied the
conventional Gaussian filter for creating DoG pyramid in scale
invariant feature transform (SIFT). DoG kernel (D(x, y, u))
and Gaussian filter (G(x, y, u)) are represented as follows :
D(x, y, o ) == L(x, y, ku) - L(x, y, o ) (1)
L(x, y, o ) == G(x, y, o ) * I(x, y) (2)
In equation 1, sigma value is increased by a certain step
to obtain the DoG pyramid. Also, for calculating accurate
local extrema, a floating-point computation is performed in
Gaussian convolution. This process requires relatively much
computation cost. Bay [2] applied the box filter for fast
implementation of feature extraction, called Speed Up Robust
Feature (SURF). SURF approximates second order Gaussian
derivatives with box filters. Although the method has fast
performance, the accuracy of feature point is that SIFT is
superior to SURF, because of a floating-point computation of
Gaussian filter.
There are several approaches for fast blurring. Young [3]
and van Vliet [4] have designed computationally efficient
methods for approximating Gaussian-based convolutions by
978-1-4244-4349-9/09/$25.00 2009 IEEE
1864
Toshi Hiro NISHIMURA
Graduate School of Information,
Production and Systems
Waseda University
Fukuoka, Japan 808-0135
Email: [email protected]
running a recursive infinite-impulse-response (IIR) filter for-
ward over the input signal, then running a second IIR filter
backward over the first filter's output. To transition between
the two filters, they use a suboptimal heuristic that produces
significant amplitude and phase distortion for all points within
about three standard deviations of the right-hand boundary.
Perri [5] introduced a SIMD 2-D convolver for fast FPGA-
based image and video processor. The authors utilized the
re-configurability nature of FPGA to introduce a new SIMD
image and video processor. This processor can dynamically
adapt itself to perform 2-D convolutions between variable bit
resolution. The results showed efficient solution for image con-
volution when compared to traditional DSP. Alex [6] presented
a few tricks which can help make real-time blurring possible.
Alex used the box filter for implementation of fast blurring.
The computation cost of the method is independent of kernel
size, meaning that it takes the same time to blur the image no
matter how much blurring is required. Mario Klingemann [7]
used a compromise method between Gaussian blur and box
blur. The method is called Stack-blur, since it uses a kind of
moving stack of colors whilst scanning through the image.
Romain Guy [8] and Durand Emmanuel [9] introduced fast
motion blurring for motion blur on web.
In this paper, we propose the fast Gaussian blurring method
using look-up table. The purpose of this paper is to highlight
more fast and accurate optimizations of blurring algorithms.
Also, this fast and accurate blurring method is required for real
time feature extraction. Look-Up Table technique has utilized
for fast computation, and leads to low computation cost for
image processing [10]-[14]. In the design of LUT, ID Gaus-
sian coefficients and pixel value of image are used to obtain
the LUT. And, for performing two directional convolutions,
the method uses integer coefficients LUT instead of floating-
point. This approach achieves low computational cost, and can
significantly outperform straightforward implementation by
using a Very Large Scale Integrated circuit (VLSI). Evaluation
of the method is measured by computation time and similarity
with a conventional 2D Gaussian blur, and compared to other
blurring methods such as box filter, Stack-blur, Romain Guy's
method, and recursive Gaussian blur. Based on the results, the
proposed method has good performance at computation cost
(a) (b) (c) (d)
Fig. I. Two-dimensional convolution filters for blurring image: (a) Box filter, (b) Tent filter, (c) Quadratic filter, and (d) Gaussian filter.
(5)
(4)
+r
Q[y ,x] = L C[y,x + k]
k = - r
where,
b
o
1.57825 +2.44413q + 1.4281q2 +0.422205 q3
b
1
2.44413q +2.85619q2 + 1.26661q3
b
2
- 1.4281q2 - 1.26661q3
b
3
0.422205l
and the relation between the desired a and q is given by:
{
0.98711a - 0.96330, a ?: 2.5
q = 3.97156 - 4.14554V1 - 0.26891a, 0.5 :::; a :::; 2.5.
A relative accuracy of this approximation increases as filter
radius a increases, but even with small a the accuracy is
good [3].
3) Box blur: A box blur, also known as moving average,
is an simple linear filter in which each pixel in the resulting
image has a value equal to the average value of its neighboring
pixels in the input image. The property of using all equal
coefficients allows implementation with a simple accumulation
algorithm. It is efficiently faster than using a sliding window
algorithm, but it lacks smoothness of a Gaussian blur. The
algorithm is based on a fact that sum Q of elements in
the rectangular window can be decomposed into sums C of
columns of this window:
The one-dimensional recursive filter has the normalization
constant B, and the filter coefficients can be calculated as
follows:
Thus, the per-pixel complexity of Gaussian blur becomes
O(logr). However, the constant is quite large, and for many
practical purposes Gaussian blur can be successfully approxi-
mated with simpler filters.
2) Recursive Gaussian filter: Using higher orders of recur-
sive filter allows a good approximation of Gaussian filtering.
This filter should be computed in two passes : forward and
backward directions, and it has 12 multiplications per pixel in
two-dimensional case, independently of the value of a in the
Gaussian kernel. The following one-dimensional forward (or
backward) difference equation takes an input (or intermediate
result) x[n] and produces an intermediate output result (or the
output) y[n] [3] :
y[n] = Bx [n] +(b1y [n- 1] + b
2
y[n- 2]+ b
3
y[n- 3])jb
o
(3)
A. Linear blurring
J) Gaussian blur: A Gaussian blur is created by applying
a Gaussian function, which is used in statistical for normal
distribution. The Gaussian blur enables a wide range of
photorealistic effects. It is also used as pre-processing step
of reducing image noise, reducing detail, and scale-space
construction.
A Gaussian blur effect is mathematically generated by con-
volving an image with a kernel of Gaussian values. Gaussian
filter on a square support is linearly separable. That is, two-
dimensional Gaussian filter can be decomposed into a series
of one-dimensional filtering for rows and columns. When
the filter radius is relatively small (less than few dozens),
the fastest way to calculate the filtering result is direct one-
dimensional convolution. When a filter radius is large, di-
rect convolution becomes expensive, and FFT-based overlap-
add(OLA) convolution is often used to perform an efficient
Gaussian blur for large blur radius . A common mistake here
is to transform the whole image row (or column) with FFT,
do the same with a zero-padded Gaussian kernel, multiply
complex spectra and do the inverse transform. First of all, it
should be considered that the result of convolution has a length
N + M - 1, where N is the signal size and M is a filter kernel
size (equal to 2r + 1), i.e. the output signal is longer than
the input signal. Without proper padding (extension) of data,
regular convolution will turn to a circular convolution, leading
to problems near image boundaries. Secondly, calculating FFT
of the complete image row is not optimal , since the complexity
of FFT is O(N log N) . The complexity can be reduced by
breaking the kernel into sections with an approximate length
M and performing OLA convolution section-wise. The FFT
size should be selected so that circular convolution is excluded.
Usually optimal performance is achieved when FFT size F is
selected as the smallest power of 2 larger than 2M, and signal
section size is selected as F - M +1 for full utilization of FFT
block. This reduces the overall complexity of one-dimensional
convolution to O(N log M).
and similarity compared to other blurring methods.
This paper is organized as follows. Section II shows
the blurring methodology. In Section III, the proposed fast
Gaussian is represented in detail. Computational results are
summarized in Section IV. Finally, the conclusion is presented
in Section V.
II. BLURRING METHODOLOGY
1865
where C[y,x] is sum of (2r + I)-pixel columns centered at
[y, x], and Q[y, x] is sum of (2r + 1)(2r + I)-pixel square
window centered at [y, x]. This allows a simple update rule
when window Q is moving from left to right:
Q[y, x +1] == Q[y, x] +Cry, x +r + 1] - Cry, x - r] (6)
Column sums C can be, in turn, effectively updated when
window Q is moving down to the next row:
Cry + 1,x] == C[y,x] + I[y +r + 1,x] - I[y - r,x] (7)
where I [y, x] is image pixel value. The result of the box filter
is equal to
complexity on i.i.d. data) is also introduced in [16]. It has
a complexity of O(logr). This running max algorithm can
be used for adding pixels to a 2D window of a 2D minimax
filter with a worst-case complexity of 0 (r log r) operations
per pixel.
3) Kuwahara filter: A Kuwahara filter a edge-preserving
smoothing filter, which smoothes . It takes a kernel around
the objective pixel the input image of size J == K == 4L + 1,
where L is a integer. This kernel is split into four overlapping
regions. In each of the four regions, the mean brightness and
variance are measured. The output value is the mean value of
that region has the smallest variance.
(12)


o
a.; [n] = {
The number of samples (N) for Gaussian convolution using
one-dimensional kernel is finite. It is common to choose N ==
3u or Su.
g2D (x, y)
Table 1 and 2 represent the coefficients of one-dimensional
Gaussian kernel from sigma 1 to sigma 2)2, and the sample
number of Gaussian kernel is 1 x 7 and 1 x 9.
III. THE PROPOSED METHODOLOGY
A Gaussian blurring effect on image is typically generated
by convolving an image with Gaussian filter as changing
sigma. In practice, it is best to take advantage of the Gaussian
Blur's linearly separable property by dividing the process into
two passes. In the first pass, a one-dimensional kernel is
used to convolve the image in only the horizontal or vertical
direction. In the second pass, another one-dimensional kernel
is used to convolve in the remaining direction. The result has
the same effect as convolving with a two-dimensional kernel
in a single pass, but requires fewer computations.
where a is referred to as both the standard deviation and scale.
The multi -dimensional Gaussian filter can be separated into
one-dimensional Gaussian filters along its main axes. Equation
10 can be rewritten as follows:
A. Separable Gaussian filter
The use of the Gaussian kernel for blurring has become ex-
tremely popular. The point spread function (PSF) and transfer
function for the space Gaussian are given by:
(8)
(9)
Ixl < T
else
{
1- E1
T '
o
Bartlett(x, T) ==
1
B[y, x] = (2r +1)2 Q[y, x]
and it requires one multiplication per pixel after Q[y, x] is
calculated. So, the overall per-pixel complexity of the box filter
is 4 additions and 1 multiplication per pixel. Some additional
over-head is required to calculate initial values of C and Q at
the image boundaries.
4) Triangular blur: The blurring effect of a box filter is
often not sufficient. An iteration of box filter can improve the
blurring effect. That is, as the number of iterations increases,
the superposition filter approaches a Gaussian blur. Triangular
filter is simply a Bartlett function ( or tent function ). it can
be constructed as constructed as a 2 box filters with the same
radius. The computational complexity of the filter is twice of
a box filter, and visual effect is similar to Gaussian blur.
B. Non-Linear blurring
1) Median filter: Median is a non-linear spatial filter for
the smoothing of signals. The output value is the middle
value of neighboring pixel values which is calculated by first
sorting all the pixel from the surrounding neighborhood into
numerical order. The kernel is usually square but can be any
shape. The filter is often used to remove noise from image
or other signals, and is particulary useful for salt and pepper
noise, speckle noise, and impulse noise. But, One of the main
problems with the median filter is relatively expensive and
complex to compute. The straightforward implementation of
median filter requires 0 (r
2
log r) operations per pixel to sort
the array of (2r +1)(2r +1) pixels in a window. Fortunately,
a fast algorithm is presented, which reduces the complexity to
O(r) [15].
2) Min/Max filters: A max(or min) filter outputs a max-
imal(or minimal) pixel value from its rectangular window.
A straightforward implementation requires O(r
2
) operations
per pixel. Fast implementations of ID running max filter
are described in [16]. A simple and fast algorithm called
MAXLINE2 is using a circular buffer of delayed input ele-
ments. This algorithm works very fast on i.i.d. (independent
identically distributed) data, but has a worst-case complexity
of O(r) for a monotonically decreasing data. An algorithm
with a better worst-case complexity (although with a worse
1866
TABLE I
GAUSSIAN COEFFICIENTS OF a == 1 AND a == V2.
sigma
1D Gaussian coefficients
Caef fl Caef f2 Caef f3 Caef f4 Caef f5 CaefIe Caef f7
0- = 1 0.004468 0.054028 0.242011 0.398985 0.242011 0.054028 0.004468
0- = V2 0.031373 0.105418 0.221339 0.283739 0.221339 0.105418 0.031373
TABLE II
GAUSSIAN COEFFICIENTS OF a == 2AND a == 2V2.
sigma
1D Gaussian coefficients
Caef fl Caef f2 Coe j j-; Caef f4 Caef f5 Caef [e Caef f7 Caef f8 Caef f9
0-=2 0.029548 0.067312 0.123539 0.178587 0.202026 0.178587 0.123539 0.067312 0.029548
0- = 2V2 0.064081 0.092560 0.122042 0.144696 0.153242 0.144696 0.122042 0.092560 0.064081
Using equation 15 and 16, Gaussian coefficients of equa-
tion 14 is corrected. One-dimensional Gaussian kernel has
a symmetric property. Thus, in according to the order of
Ex..Coe]f s( a, k), integer coefficients can be normalized by
adding 1 at each term in the kernel symmetrically until eo- is
O. Equation 17 represents a look-up table for convolution of
two passes using a coefficient values.
B. Design of Look-Up Table
Gaussian blur are used in the computation of high-pass and
band-pass filters. High-pass filters are obtained by subtracting
a blurred result from the original image, and band-pass filters
are obtained by subtracting the blurred images of two different
sigma. We can obtain the Difference of Gaussian as calculating
these process. Lowe [1] in equation 1 and 2 introduced
that DoG pyramid was obtained by the difference between
Gaussian convolution with initial sigma value and Gaussian
convolution with k scaled sigma. Gaussian blur computation is
slowed depending on the sigma's growing. Thus, for reducing
computation time, equation 13 is written as follows:
N
eo- ==p- LLuTCaeffs(eY,k)
k=O
(16)
When eYk is increased like Sk={ -j2, 2, 2-j2, 4, 4-j2...}, eYe
will be like Se={l, -j2, 2, 2-j2, ...}. In this study, we create
the coefficients LUT using Se.
where coef f .peakoalue denotes the max integer coefficient
after scaling, pixelcpeaknoluc is the max pixel value of im-
age, and p is a scaling factor for converting floating coefficients
to integer value. Two step passes are required for applying the
proposed method using two LUTs. Two LUTs are used in the
horizontal and vertical direction as follows:
== *
G( *G( == G( +
== +
(To = J - (TLI
(13)
LuTcanv(n, m) n x m (17)
{
0::; n < coef f .peakcolue
o::; m ::; pixei.peakualsu: x p
Ex.Coej j sio, k) == LuTcaeffs(eY, k) - p X [k] (15)
(19)
(18)
L LuTcanv (LuTcaef fs (eYk' p), I(p, j))
p=-r
r
( L LuT
c anv
(LuT
c aef
fs( eYk, p), B
x
( i, p))) / p2
p=-r
==
r
where B
x
and B
x y
denote the results of applying the LUT
in the horizontal and vertical direction, respectively. Using
equation 18 and 19, floating-point computation for image
blur is not needed. These computation process could be also
usefully applied to the VLSI architecture for image blur
algorithm.
(14)
LuTCaeffs(eY, k) round(p x [k])
{
Ikl::; No
a == {1,-j2,2, ...}
where a denotes the standard deviation of one-dimensional
Gaussian filter, and (2N
o
+1) denotes the sampling number. p
is the scale factor to convert gaussian coefficients to integer
values. When converting the scaled Gaussian's floating-point
values into the integer values needed for the LUT, the sum
of the coefficients will be different from 1 x p. This causes a
darkening or brightening of the image. To remedy this, the
normalization process is required. Equation 15 denotes the
difference between floating-point coefficient and integer value
scaled, and equation 16 is the difference between 1 x p and
sum of integer coefficients scaled.
1867
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Fig. 2. Blurring experiment using conventional two-dimensional Gaussian filter and the proposed method: (a) Original image. (b) shows the result by
conventional two-dimensional Gaussian filter with (J" = I. (c) the proposed method with (J" = I. (d) conventional two-dimensional Gaussian filter with (J" = V2.
(e) the proposed method with (J" = V2. (f) conventional two-dimensional Gaussian filter with (J" = 2. (g) the proposed method with (J" = 2. (h) conventional
two-dimensional Gaussian filter with (J" = 2V2. and (i) the proposed method with (J" = 2V2.
TABLE III
EVALUATION OF BL URR ING MET HODS ACCORDING TO COMPUTING TIME(t
m s
) AND ERROR RATE (e) .
where k is the standard deviation. M and N denote a image
size, and (i, j) is the pixel coordinate in a image. 'Pij and 'r]ij
denote the pixel values of the blurred image by conventional
two-dimensional Gaussian blur and other blurring methods,
respectively.
(20)
MxN
Blurring 0- =1 0- = V2 0- =2 0- = 2V2
methods
t-, e t -, e t-, e t rn s e
Conventional 20 Gaussian filter 47 0 78 0 172 0 297 0
Recursive Gaussian filter 12 14.9 12 13.1 12 11.08 12 9.0
Stack-blur 16(r = 3) 0.81 18(r = 4) 1.02 21(r = 6) 0.75 31(r = 9) 0.65
Romain Guy's Method 20(r = 1, x2) 1.5 30(r = V2, x 3) 1.04 20(r = 2, x2) 0.92 30(r = 2V2, x 3) 1.49
Box filter (x 6) 42(r = 3) 0.9 46(r = 5) 0.9 52(r = 7) 1.0 60(r = 9) 1.2
The proposed method 7 0.2 7 0.3 8 0.5 8 0.8
IV. C OMPUTATI ONAL R ESULTS AND DISCUSSION
The approach of the proposed method were tested using
320 x240 gray images on a Workstation computer of a 3.2Gh
Xeon(TM) CPU and IGBytes main memory. The evaluation of
blurring methods is performed according to the computation
time and similarity with the blurred image by conventional
two-dimensional Gaussian filter.
1868
In the Figure 2, we show the blurred image by conventional
Gaussian filter and the proposed method as changing the
sigma from 1 to 2-j2. And, table 3 represents the comparison
of the computing time and similarity using other blurring
methods. As expected, implementation the conventional two-
dimensional Gaussian filter was relatively slow. Recursive
Gaussian filter was about 2 or 3 times as fast. Computation
time of Stack-blur and box filter is independent of filter
radius(r). Romain Guy's method and box filter also require
repetition accomplishment for similarity with Gaussian blur.
The best results were obtained with the proposed result ap-
proach, which was almost 7 or 30 times as fast as conventional
Gaussian blur. Also, according to the similarity of the blurred
image, the proposed method is superior to other methods. It is
clear that the proposed method performs very well compared
to other methods.
v. CONCLUSION
In this paper, we have introduced a fast and accurate method
for image blurring. The method used a look-up table which
was created from coefficients one-dimensional Gaussian for
two-dimensional Gaussian convolution. The LUT played key
role to omit the floating- point computation for gaussian
convolution. We demonstrated that the proposed method had a
good performance for image blurring according to computing
time and similarity as comparing to other blurring methods
such as conventional Gaussian filter, recursive gaussian filter,
Stack-blur method, box filter, and Romain Guy's Method.
Based on these computation results, the proposed method can
be also effectively utilized to create the DoG pyramid for
feature extraction or other image processing techniques.
REFERENCES
[1] D. G. Lowe, "Distinctive image features from scale invariant features,"
International Journal of Computer Vision, vol. 60, no. 2, pp. 91-110,
2004.
[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool, "SURF:
speeded up robust features," Computer Vision and Image Understanding,
vol. 110, no. 3, pp. 346-359, 2008.
[3] I. T. Youngand L. J. van Vliet, "Recursive implementation of the gaussian
filter," Signal Processing, vol. 44, pp. 139-151, 1995.
[4] L. J. van Vliet, I. T. Young and P. W. Verbeek, "Recursive gaussian
derivative filters," Proc. of the Int. Conf. on Pattern Recognition, vol. 1,
pp. 509-514, 1998.
[5] S. Perri, M. Lanuzza, P. Corsonello and G. Cocornullo, "SIMD 2-D
convolver for fast FPGA-based image and processors," Proc. of the
Military and Aerospace Programmable Logic Devices (MAPLD) Int.
Conf., Washington, 2003.
[6] E. Alex, "Four tricks for fast blurring in software and hardware", [Online]
Available : http://www.gamasutra.com/features/20010209/evans_01.htm.
2001.
[7] M. Klingemann, Stack-blur, [Online] Available: http://incubator.quasi
mondo.com, 2004.
[8] Romain Guy, [Online] Available: http://www.progx.org, 2005.
[9] Duran Emmanuel, [Online] Available: http://www.koders.com/delphi/
fid5AC677DA9310032F9D061DC4D148CFAC158EB37D.aspx?s=
algorithm, 2004.
[10] P. L. Chu, "Fast Gaussian Noise Generator," IEEE Trans. on Acoustics.
Speech. and Signal Processing, vol. 37, no. 10, pp. 1593-1597, 1989.
[11] M. S. Mohammed, S. L. Matilia, L. Nozal, "Fast 2D convolution filter
based on look up table FFT," Proceedings of the IEEE Int. Symp. on
Industrial Electronics, pp. 446-449, 1992.
[12] M. Mese, P. P. Vaidyanathan, "Look up table (LUT) inverse halftoning,"
IEEE Trans. on Image Processing, vol. 10, pp. 1566-1578, 2001.
1869
[13] X. P. Zhang, K. Li, and X. F. W, "A Novel Look-Up Table Design
Method for Data Hiding With Reduction Distortion," IEEE Trans. on
Curcuits and Systems for Video Technology, vol. 18, no. 6, pp. 769-776,
2008.
[14] Y. M. Koo, B. S. Shin, "An Efficient Point Rendering Using Octree and
Texture Lookup," Lecture Note on Computer Science(LNCS 3482), vol.
3482, pp. 1187-1196, 2005.
[15] B. Weiss, "Fast Median and Bilateral Filtering," ACM Transactions of
Graphics (proceedings of the ACM SIG-GRAPH'06 Conference), 2006.
[16] M. Brookes, "Algorithms for max and min filters with improved worst-
case performance," IEEE Trans. on Circuits and Systems, vol. 47, no. 9,
pp. 930-935, Sep 2000.
[17] R. Deriche, "Fast algorithms for low-level vision," IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 12, no. 1, pp 78-87, 1990.