Você está na página 1de 5

2010 Second International Conference on Intelligent Human-Machine Systems and Cybernetics

Key Point Detection and High Speed Image Registration using BLoG

Jong-Min Kim, Meang-Kyu Song, Kyoung-Ho Kim Woong-Ki Lee


Computer Science and Statistic Graduate School Dept. of Computer Science and Statistic
Chosun University Chosun University
GwangJu, Korea GwangJu, Korea
mrjjoung@hotmail.com leewk@chosun.ac.kr

Abstract— SIFT is one of the feature region extraction SIFT preprocesses with gray scale image and an image is
methods. It is widely used in various fields including object produced for each Gaussian distribution, and after producing
recognition, mosaic object tracking, etc. However, since SIFT the different size of images with Image Pyramid, DOG
requires lots of arithmetic operations, it is not appropriate for (Difference of Gaussian) is found through the difference of
real-time use. Therefore, many researchers has been tried to Gaussian image produced. When this DOG is used, a lot of
prepare other measures. SIFT using PCA or SURF is one of robust key points to the size and rotation changes are
them. This paper is to study about the real-time CCTV produced. Since these key points include numerous key
combined system using the local partial image and to suggest points due to the noise and errors and use the high
the combination and recognition method for frames. Its
dimensional key point vector, even though it shows a good
purpose is to reduce the combination speed of the suggested
algorithm compared to the existing one and to enhance the
result, it is not appropriate for the real-time image processing
overall recognition rate according to this. Since the existing because of a large amount of arithmetic operators. In
SIFT algorithm method has a patent and a slow processing addition, it even has a patent so that there are a lot of
speed, in this paper the improved local image regenerating limitations on practicality. Therefore, the methods designed
method was used and the speed was raised according to the like SURF (Speeded Up Robust Features) [7] that could find
actual CCTV processing speed. Based on the recognition rate the robust key points faster based on the PCA-SIFT[8] or
and speed by using the localized image, this paper consisted of Hessian[9] which applied the PCA (Principal Component
realization of overall system and introduction of algorithm, Analysis) method that reduced the high dimensional vector
and the comparative evaluation of performance was carried into the lower dimension was suggested and it has been used
out through the actual experiment. By applying the CCTV in various fields until now. However, this one is also not
operating in real time, the low-cost system was built with Inline enough in real-time image matching. So this is one of the
System not with each image and this could reduce the physical fields that the study about the faster matching methods is in
fatigue of watcher, have an economic effect and could be used progress.
by general users easily. Since the real-time image matching method needs fast
arithmetic operations, when finding the similarity between
Keywords-Mosaic, SIFT, BLOG matching images, it is important to find the robust points that
have an effect on the image matching with a small number of
I. INTRODUCTION
0B
points rather than a lot of points. This should detect the
Image matching is used various fields. The first step for unchangeable descriptor not to be effected by the noise and
image matching is to detect the feature region. For detection illumination changes as well as size and rotation changes.
methods of feature region, there are points, lines and edges. Therefore, this paper was to suggest the methods to reduce
In the early days it was detected mainly with points. The first the Image Pyramid, to leave the strong images between
operator to detect the points was started with Moravec[1] matching images and to reduce a lot of matching points.
operator, but since it had a limitation of anisotropy, Harris
II. SUGGESTED BLOG TECHNIQUE
and Stephens[2] suggested the key point detection method
2B

using the autocorrelation function which has a large effect on A. Overall Configuration
the detection but costs a lot on calculation. This Harris
7B

Corner Detection is still used a lot but it has a disadvantage The Figure 1 for much faster image matching is the
of slow speed due to a lot of arithmetic operations. overall configuration.
Therefore, various methods were designed to improve the
speed including Susan[3] and Fast Corner Detection[4].
These methods had difficulty to match images with the key
point detection because they were affected by noise, size,
rotation, etc. Among them, SIFT (Scale Invariant Feature
Transform), the method suggested by David G. Low [5, 6]
lastly, which detects the robust points that do not change
with the size and rotation changes has the most excellent
performance in image matching so that still it is used in
many fields such as robot or similar image recognition, etc.
Figure 1. Image matching overall configuration using BLOG

978-0-7695-4151-8/10 $26.00 © 2010 IEEE 259


245
DOI 10.1109/IHMSC.2010.161
The first step is to extract the key points by using BLOG In here, the first and last Bit Slicing are not included. The
method, the second step is to remove the key points that were reset image will set up the robust key points by applying
produced by the noise, and the third step is to remove the certain number of critical values.
region formed by the error and localize it. The last step is to
match the images by using the RANSAC [10] method.
B. BLOG (BitRate Laplacian Of Gaussian)
8B

The suggesting BLOG(BitRate Laplacian Of Gaussian) is


the method to detect the key points much faster by using the
Laplacian function and Gaussian function.
First, make the image with isotropy using the Laplacian
function. Since the Laplacian function is a quadratic
differential function, it has a response at the points, lines or Figure 4. Result of Robust Key Points
isolated points with changes. With the image obtained from
the Laplacian function, a new image should be produced by
C. Removing the noise and lowering the error
using the Gaussian function. Not like DOG and LOG, in
9B

BLOG as shown in Formula 1 the image is produced by The number of robust key points produced here still is a
applying the Gaussian function to the Laplacian image with lot. All these robust points include the inefficient key points
different critical values and instead Image Pyramid is not set that are not correct due to the noise and errors. Therefore, to
up separately to reduce the arithmetic operations. remove the noise the factors of continuous direction
component were reduced from the bit-planed image through
BLOG( x, y,σ ) = ((L( x, y),σ ) × I ( x, y)) | (25σ ) Formula 3. The Formula 3 only shows the Y-axis direction,
(1)
but all four directions, x-axis, y-axis and diagonal directions
By extracting the image made into the Bit-Plane Slicing were applied.
as shown in Figure 2 the key points will be detected. f ( x, y ) = f ( x − 1, y ) ∩ f ( x + 1, y )
(3)

By removing the noise in each bit unit, the noise around


the bit sliced robust point can be removed. Figure 5 is the
picture showing the difference of before and after removing
the noise.

Figure 2. Bit-Plane Slicing

Figure 3 is the result showing each bit by applying


Formula 1 according to the critical values setting up
differently one another.

Figure 5. Noise Removal (4 directions)

Even though the noise was reduced, there are many


points left in the image due to the remaining robust points
produced by errors. Formula 4 which applies the Gaussian
function was used to extract the complete key points without
errors.
Key Point = (Lx+1, y − Lx−1, y )2 + (Lx, y +1 − Lx, y−1 )2 (4)
Figure 3. Images according to each critical value of BLOG

With each BLOG image produced here, the robust key To apply the Formula 4, first measure the maximum
points will be extracted from Formula 2. value with the boundary of 16 * 16 regions as the center of
6 robust key points and extract the key points. When applying
Robust Po int = ∑ BLOG( x, y,σ ) (σ = 0.4n) (2) the Formula 8, the most robust key points can be made with
n=2 the measured maximum value and many surrounding robust
values. Since the errors can be reduced by resetting the
points formed by surrounding errors with this, as well as it
can reset the most robust key points among many key points

246
260
set up in the image that crowded with various directions, the
number of points needed in the image matching can be
reduced as shown in Figure 6.

Figure 9. Key points extraction by scale changes


Figure 6. Image Error Remove
However, when comparing the key points extracted by the
size changes after removing the noise of these points, the
D. Image Matching
10B
main matching points are still extracted even though the size
To match the images with key points extracted from is changed as shown in Table 1.
BLOG, Random Sample and Consensus (RANSAC) [10]
method which is one of the image matching methods was TABLE I. RESULT BY SCALE CHANGES
used. Figure 7 is the picture that decides the image matching 50% Original 170%
points by using RANSAC method. Figure 8 is the final result Image Image/Noise image/Noise Image/Noise
Robust
Key points
points
of matching three different images with matching points. removed removed removed

1.Pal Sang Jeong


102/57 405/102 1230/271 39 33
(768*576)

2. Mountain 1
57/22 125/63 96/59 17 14
(320*240)

3. Mountain 3
51/21 108/52 51/40 19 17
(320*240)

4. Garden 1
53/47 160/74 128/78 38 31
(256*256)

5. Garden 2
56/36 155/41 154/42 31 25
(256*256)
Figure 7. Matching point decision using RANSAC
From Table 1, it is about 83.3% that when the robust
points extracted by the size changes become the main
matching point. That is, even though the key points have an
effect on the scale change, the robust points that affect on the
image matching are almost maintained.
B. Rotation Change Study Analysis
12B

To check how much the BLOG suggested affected on the


rotation, the rotation transformations were examined in two
dimensional Euclid space and three dimensional space.
Figure 10 is the image that was obtained by rotation changes
Figure 8. Final Image with 10˚ intervals of rotation in two dimensional Euclid
space.
III. RESULTS AND ANALYSIS
3B

A. Size Change Study Analysis


1B

The experiment was done by changing the various sizes.


As a result of extracting the key points according to the
changes, the points were extracted irregularly from the size
changes as shown in Figure 9.

Figure 10. Key point extraction by the rotation changes

Figure 11 is the table that compared the robust points


according to the result above with points that affects on
matching. In this table, the key points were maintained as
81.7%.

247
261
Figure 11. Result by Orientation changes with 10 degrees interval

Among them, the key points that affected on matching was Figure 13. Performance comparative table of BLOG
maintained more than 90%. That is, the robust key points that
affected on the image matching according to the rotation The important things in the real-time image matching are
changes in Euclid space had almost no changes. the correct matching, less arithmetic operations and fast
The study about three dimensional rotation changes matching speed. As a result of comparing, the key point
followed by two dimensional changes was shown in Figure extraction accuracy of suggested BLOG was dropped
12. As a result, in three dimensional studies the robust key slightly. However, it is exceptional in speed of finding the
points were appeared a lot differently but the key points for key points and memory efficiency more than the compared
image matching were only reduced a little and they were object. Also, in case of speed Harris Corner Detection and
maintained enough for conditions needed on image matching. Fast Corner Detection may be much faster, but the speed of
overall image matching is more efficient in the matching
method of finding the robust key points than matching
method that finds every corner.
IV. CONCLUSION AND FUTURE PLAN
4B

To match the image, it is important to find the robust key


points against size, rotation and contrast changes. For perfect
image matching without speed limitation, the key point
extraction method using DOG is most appropriate for
matching. However, it is not suitable for real-time image
matching.
In the real-time image matching that values the speed, the
SURF using LOG can be used but since BLOG does not use
Image Pyramid, it is excellent in memory efficiency and
speed. On the other hand, it is not enough for the key point
extraction because it can’t extract all the key points. But it
can almost maintain the robust key points that are needed in
image matching. Therefore, it is necessary to have a further
study about more accurate matching point extraction while
maintaining the speed and memory efficiency of BLOG.
Figure 12. Result from the orientation changes in three dimensional space REFERENCES
14B

[1] Rodrigo de Luis-Garcia, Carlos Alberola-Lopez, Otman Aghzout and


C. The extraction speed of robust key point and memory
13B
Juan Ruiz-Alzola "Biometric identification systems,J" Signal
efficiency Processing Vol. 83, Issue 12. pp. 2539-2557, Dec 2003.
[2] Z. Zang, M. Lyons, M. Schuster and S.Akamatsu, “ Comparison
Figure 13 is the graph that compared with the key point between Geometry-Based and Gabor Wavelets-Based Facial
rate, robust key point rate used in matching, speed, memory Expression Recognition Using Multi-Layer Perceptron ” ,
efficiency and noise rate with key points of BLOG which Proceedings of Third IEEE International Conference on Automatic
suggests the DOG (Difference of Gaussian) used in SIFT, Face and Gesture Recognition, pp.454-459, 1998.
LOG (Laplacian of Gaussian) used in SURF, Harris Corner [3] Hiroshi Murase and Shree K, Nayar, "Visual Learning and
Detection and Fast Corner Detection that is generally used a Recogntion 3-D object from appearance", international journal of
Computer Vision, Vol,14,1995.
lot.
[4] Juwei Lu; Plataniotis, K.N.; Venetsanopoulos, A.N., "Face
recognition using LDA-based algorithms, “ IEEE Transactions,
Neural Networks, Volume: 14, Issue: 1, pp. 195-200, Jan. 2003
[5] J.Kathunen, E.Oja, L Wang R. Vigario, and J. Joutsensalo, "A class
of neural networks for independent component analysis," IEEE Trans.
Neural Networks, vol. 8, pp. 486 - 500, May 1997.

248
262
[6] Martinez, A,M; Kak, A.C;, "PCA versus LDA, "IEEE Transactions, Transactions, Pattern Analysis and Machine Intelligence, Volime: 19,
Pattern Analsis and Machine Intelligence, Volume: 23, Issue: 2, Issue: 7, pp. 711 - 720, July 1997.
pp.228 - 233, Feb 2001. [9] A. D. Kulkarni, Computer Vision and Fuzzy-Neural Systems,
[7] F. Crow, "Summed-area tables for texture mapping", In Proceedings Prentice-Hall, 2001.
of SIGGRAPH, Vol. 18(3), pp. 207-212, 1984. [10] Jong-Min Kim, Hwan-Seok Yang and Woong-Ki Lee "Network-
[8] Belhumeur, P.N,; Hespanha, J.P.; Kriegman, D.J.;, "Eigenfaces vs. Based Face Recognition System Using Multiple Images", Lecture
Fisherfaces: recognition using calss specific linear projection," IEEE Notes in Artificial Intelligence LNAI 4088 pp626-631, 2006

249
263

Você também pode gostar