2015 8th International Symposium on Computational Intelligence and Design

A Novel Method for Image Recognition Based on Polynomial Curve Fitting

Chao JI, Xiaodong YANG, Wei WANG

Department of Navigation
Navy Submarine Academy
Qingdao, China
e-mail: [email protected]

Abstract—Invariant feature is desirable in image recognition coefficients reflect the curve data on its own
and computer vision, while feature matching is limited in the characteristics[9,10].
distance or the relative measurement between the image and For that, a novel method is proposed in this paper. With
the model. In this paper, we propose a novel method for image the complex moment invariants as the characteristic vectors
recognition based on polynomial curve fitting. Non-redundant of the image, the fitting polynomial of the characteristic
complex moments are derived, that are invariant to the vector is selected as the objective function, and the similarity
translation, constraint scale and rotation, and complex of fitting polynomial coefficients to describe the similarity
moments below 4th order are selected as the invariant feature. between the images, finally the method is applied to real ship
With the principle of polynomial curve fitting, similarity images to verify its effectiveness.
measurement between the fitting coefficients of the non-
redundant complex invariants is calculated to describe the II. MOMENT INVARIANTS
images. Finally it is applied in real ship images and its
validation is analyzed. Experimental results presented show A. Geometric Moments and Image Normalization
that the described method has good stability and can be used
as an effective method for image recognition. Assuming that image f ( x, y ) is M  N , the definition of
its (p+q)th order original geometric moments and central
Keywords-image recognition; polynomial curve fitting; geometric moments are as follows respectively [1,3]
complex invariant moments; invariant feature; similarity M N
M p , q   x p y q f  x, y 
 x 1 y 1 (1)
I. INTRODUCTION  p, q  0,1,2, 
M N

Image recognition is extensively applied in the field of 

m p , q 

x 1 y 1
 x
x 0  p
 y
y 0 q
f  x , y 
military and automation. There are two keys in image
recognition: one is to extract the invariant image features; the By definition, as long as f ( x, y ) has definite non-zero
other one is to seek the best matching method based on the values and moments of all orders exist. And the set
features to achieve the target classification. Since Hu M.K. in {M p ,q } is uniquely determined by f ( x, y ) ; and conversely,
1962 used algebra theory to construct the Hu invariant
moments that are invariant to translation, rotation and f ( x, y ) is uniquely determined by the set {M p ,q } , where
scale[1,2], they are widely used in the field of image x0 , y0 are the centroids of the image.
recognition.
M1,0 M 0,1
Based on the recognition of image invariant features, x0  , y0  , m0,0  M 0,0 (2)
there are mainly Hu moments as well as their related M 0, 0 M 0, 0
moments and NMI features (normalized moments of inertia) The central moments are defined as
and proportion features[3-6]. At present, research on the m p,q pq
image invariant features is focused on using different I p, q  , l  1, p  q  2,3, (3)
m0l ,0 2
mathematical theories to construct the invariant features,
while feature matching only confined to the calculation of Moments are used for image recognition and matching,
the distance or related measurement between the image and and one of the main work is to compare the descriptor of the
the template in accordance with certain standards to normalized image with the prior standard template to achieve
determine target category [7,8]. the recognition or matching under the condition of certain
Complex moments and their linear combinations meet similarity[7].
the invariance of translation, scale and rotation. Functions of Using geometric moments for the image normalization,
complex moments have been widely applied in object they have the following characteristics
identification, object pose estimation, image reconstruction, M 0,0   , M 2,0  M 0,2 , M 3,0 0

(4)
etc(such as image processing, image recognition and 
computer vision).Curve fitting is a data processing method M1,0  M 0,1  0, M1,1  0

which used continuous curve to depict or compare the Specific steps of image normalization are in literature [8]
functions of discrete points in the plane, and the fitting for details. After the normalization, the feature descriptors
may not be invariant corresponding to translation, scale

978-1-4673-9587-8/15 $31.00 © 2015 IEEE 354

DOI 10.1109/ISCID.2015.38
 p q
change and rotation of the image, but they can be used as the  p  q 
feature vectors. C p,q   k  l i
k 0 l 0
k l

1l I p  q 
k l ,k l  (7)

B. Complex Invariant Moments and Their Generation Theorem 1 TSR invariance theorem[3]
Invariant feature extraction and construction is one of  k 

k k
important technologies for image recognition. To find a  k
C N
 C pNj , q j ,  p j   q j , k  Z   forms the
calculation method that can satisfy the rotation invariance of  j 1 j 1 j 1 
moment is the key to construct invariant moments. Invariant
space base of the invariant polynomials space which is
moments can construct a set of feature collection invariant to
invariant to translation, constraint scale and rotation, and the
all kinds of space transformations (including translation,
base is complete.
scale and rotation), the changing viewpoint and location.
The above theorem further derived, we can find that the
Ideal image invariant features must have the essential
low-order complex moments can be expressed by central
characteristics of the image and insensitive to changes in
complex environment. Translation, scale and rotation geometric moments, and all of them are real numbers. When
invariance are the important characteristics of invariant the order is more than 2, the complex moments are made up
moments. Complex moments with translational invariance of two serial multiplications, three serial multiplications and
and constraints of scale and rotation invariance, cannot be multiple serial multiplications. According to the TSR
directly used to describe the target image characteristics. The invariance theorem, for each order complex moment, the
linear combination and multiplied of the complex central number of its two serial multiplications is
 r 
moments meet the translation and scale invariance, and thus  r
INT   1      1, r is odd (8)
constitute complex invariant moments.   
2    0, r is even
 1  
When using the complex space describing the image, the  
(p+q)th order central complex invariant moments are defined where r is the order of the complex moments.
as The number of the multiple serial multiplications is˖
 $ ! 
 r % "INT r   1      1, r is odd
M N

C p ,q   x
x0   i y
y 0  x
x0 
i y
y 0  f x, y  (5)
p q (9)
 # 2  
 x 1 y 1  1    0, r is even
 p, q  0,1,2,  

where C0,0  M 0,0 .The normalized complex moments are C. Redundance of the Complex Moments
expressed as Redundant analysis of complex invariant moments can
effectively reduce the unnecessary calculation and improve
C p,q pq
C pN, q  , l  1, p  q  2,3, (6) its use, which has a practical significance. According to the
C0l ,0 2 construction of complex invariant moments, we can find that
Theoretical analysis and practice show that if using high the high order moments contain low moments, and multiple
order of geometric moments describing the characteristics of serial multiplications contain two serial multiplications. As a
the image, there exist instability and complicated calculation, result, each order complex invariant moments can only be
but complex moments don’t. Combined with (1)(3)(5)(6), we made up of two serial multiplications to avoid redundancy.
can obtain TABLE ĉ shows the non-redundant complex invariant
moments less than 4th order.

TABLE I. NON-REDUNDANT COMPLEX INVARIANT MOMENTS LESS THAN 4TH ORDER

Composition of
Order Complex moments Expansion of complex moments
complex moments
1
1st C1N   C pNj , q j C1N,1 I 2, 0  I 0, 2
j 1
2
2nd C2N   C pNj , q j C2N,0C0N, 2 I 2, 0
I 0, 2   4I12,1
2

j 1

3 C3N,0C0N,3 I
3I1, 2   3I 2,1
I 0,3 
2 2

C  C
N N 3, 0
3rd
I  I1, 2   I 2,1  I 0,3 
3 p j ,q j 2 2
j 1 C2N,1C1N, 2 3, 0

4
C4N,0C0N, 4 I 4, 0
6 I 2, 2  I 0, 4   16I 3,1
I1,3 
2 2

4th C4N   C pNj , q j C3N,1C1N,3 I 4, 0

I 0, 4   4I 3,1  I1,3 
2 2

j 1
C2N, 2C2N, 2 I 4, 0  I 0, 4  2I 2, 2 
2

355
 m
III. PRINCIPLE OF POLYNOMIAL CURVE FITTING
Fitting is with one or more simple functions to describe a [ P ( x )
y ]
i 0
n i i
2
is the square error of least square curve
set of points in an enclosed region. In practice the commonly
fitting polynomial Pn (x) .
used functions are polynomial and trigonometric functions.
According to the Weierstrass theorem, any continuous In practical applications, it must meet the requirements of
function within enclosed region can be expressed with a n*m or n(m; When n=m the fitting polynomial is called
polynomial function. So any set of points can be fitted with Lagrange or Newton interpolation polynomial.
one or more polynomial functions. In this paper, we give the complex moments below 4th
Assuming given data points ( xi , yi ) (i=0,1&,m)ˈ' are order: y1  C1N,1 , y2  C2N,0C0N,2 , y3  C3N,0C0N,3 , y4  C2N,1C1N,2 ,
the functions made up of polynomials with not more than n y5  C4N,0C0N,4 , y6  C3N,1C1N,3 , y7  C2N,2C2N,2 .All of the above
order(n(m).
n moments are selected as the feature vector, that is
Now we let Pn ( x)  a x k
k
 ' , in order to get the Y  [ y1, y2 , y3 , y4 , y5 , y6 , y7 ] with its corresponding argument
k 0 X  [ x1 , x2 , x3 , 4 , x5 , x6 , x7 ]  [1,2,3,4,5,6,7] .
following result By polynomial curve fitting to calculate the fitting
m m n
I  [ Pn ( xi )
yi ]2  ( ak xik
yi ) 2  min (10) coefficients, we use the similarity of fitting coefficients to
describe the image similarity, so that we can determine the
i 0 i 0 k 0
If the functions were polynomial, it is called polynomial image target attribute. The similarity of fitting coefficients is
measured by the Euclidean distance
fitting. If Pn (x) demands (10), we call it the least square n
Euclidean  A, B   (  a
b
2 1/ 2
curve fitting polynomial. In particular, when n=1, it is called i i ) (15)
linear fitting or line fitting. i 0
m n
where A  (a0 , a1,, an ) and B  (b0 , b1,, bn ) are vectors

Obviously, I  ( ak x k
yi ) 2 is
i 0 k 0
the multivariate
constituted by fitting coefficients.
function of a0 , a1,, an . So the above problem is to get the
IV. EXPERIMENTATION AND DISCUSSION
extreme value of I  I (a0 , a1,, an ) .
Now we prove the method in this paper with two types of
According to the necessary condition of multivariate ship images got from the Internet. Firstly we get the binary
function extreme value, we have image of the ships in the Matlab software, as shown in Fig.
m n
)I 1.The image (a) is the template in the priori standard set, the
)a j 
 2 ( ak x k
yi ) xij  0 , k  0,1,, n
i 0 k 0
(11)
images (b),(c),(d),(e) are respectively the obtained images
under different geometric transformations; the image (f) is
n m m
That is  (
k 0 i 0
xij  k )ak  x y
i 0
i i (12)
the image of another ship. TABLE Ċ shows the results
calculated according to complex invariant moments in
TABLE ĉ , at the same time the Euclidean distance image
Equation (12) is the linear equations about a0 , a1,, an , (b) ~ (f) relative to the image (a) is given based on equation
it can be expressed by matrix as (14).TABLE ċ gives out the coefficients of polynomial
$ curve fitting and Euclidean distance similarity of the image
n ! $ m n !
m m

" m  1  xi   xi  "  yi  (a) ~ (f).

" m i 0 i 0
 $a0 ! " mi 0 
n 1  " a 
m m
" x
"   "  (13)
"   " 
x 2
 x 1 x y
i i i i i

" 
i 0 i 0 i 0
" i 0
  "  "  
"m  a "m 
" xin  xin1   xi2 n  # n
m m
" xin yi  (a) (b) (c)
"# i 0 i 0 i 0  "# i 0 
(12)or(13)is called normal equations.
We can prove that the coefficient matrix of the equations
(13) is a symmetric positive definite matrix, so they have
unique solution. We work out ak , k  0,1,, n from(13), then
n
(d) (e) (f)
Pn ( x)  a x
k 0
k
k
(14)
Figure 1. Images of two ships

We can prove that Pn (x) in (14) demands (10), that is to

say Pn (x) is the desired fitting polynomial. And

356
 TABLE II. RESULTS OF COMPLEX INVARIANT MOMENTS

Complex Invariant Moments Image(a) Image(b) Image(c) Image(d) Image(e) Image(f)

C1N,1 0.0270 0.0360 0.0268 0.0270 0.0296 0.130

C2N,0C0N, 2 0.000653 0.00118 0.000644 0.000653 0.000775 0.0165

C3N,0C0N,3 8.3610-13 9.0110-12 1.5610-12 8.6310-13 1.0610-11 5.0510-11

C2N,1C1N, 2 1.6810-12 2.9410-12 1.3810-11 1.8210-12 9.2110-12 4.8610-11

C4N,0C0N, 4 1.2910-6 4.2310-6 1.2510-6 1.2910-6 1.7910-06 0.000872

C3N,1C1N,3 1.5410-6 4.9410-6 1.5110-6 1.5310-6 2.1810-06 0.000905

N N
C C
2, 2 2, 2 1.6310-6 5.2110-6 1.6210-6 1.6510-6 2.3610-06 0.000916
The Euclidean distance 0 0.0091 0.00017 0.000012 0.0027 0.10
As we can see from the data in TABLEĊ, all orders of the complex invariant moments between different ship
complex invariant moments of the same ship target image ((a) images is larger, significantly greater than the same ship
~ (e)) changed little, even if the image (a) have undergone a images. We can see that complex invariant moments and
variety of geometric transformations, and the calculated their distance similarity measure can be applied to
Euclidean distance is smaller; but the Euclidean distance of distinguish different target images.
TABLE III. RESULTS OF FITTING COEFFICIENTS SIMILARITY
Fitting Coefficients (ai) Image(a) Image(b) Image(c) Image(d) Image(e) Image(f)
a0 0.132 0.173 0.132 0.132 0.144 0.503
a1 -0.173 -0.225 -0.172 -0.173 -0.189 -0.585
a2 0.0860 0.111 0.0855 0.0863 0.0938 0.264
a3 -0.0203 -0.0261 -0.0204 -0.0205 -0.0223 -0.0583
a4 0.0024 0.0032 0.0024 0.0022 0.0024 0.0061
a5 -0.0001 -0.0001 -0.0001 -0.0001 -0.0001 -0.0003
The Euclidean distance 0 0.071 0.0015 0.00014 0.022 0.57

As can be seen from TABLE ċ, using fitting polynomial [2] J. Flusser, T. Suk, B. Zitova, “Moments and moment invariants in
pattern recognition”. Wiley & Sons Ltd, 2009.
coefficients of complex invariant moments given by this
[3] R. S. Kaushal and S. Shweta, “Construction of complex invariants for
paper to describe the target feature, the Euclidean distance classical dynamical systems”. Annals of Physics, vol. 288, 2001,
values correspond to the Euclidean distance of the complex 253–276 .
invariant moments, and the coefficients of the same ship [4] A. Abo-Zaid, O. Hinton, E. Home, “About moment normalization
image is small with smaller value of Euclidean distance, so and complex moment descriptor”, Proc. of the 4th International
the method proposed in this paper has the very good stability Pattern Recognition Conference, Cambridge, U.K. 1988, pp. 399-409.
describing the target feature. The Euclidean distance of [5] Y. M. Wang, “The Moments of Image-Theory, Algorithm and
fitting coefficients between image (f) and image (a) is large, Applications”. Shanghai: East China University of Science-
so they are obviously not the same kind of image. The results Engineering Publishing, 2002㧚
shown in TABLE ċ are consistent with TABLEĊ. The [6] S. Ullman and R. Basri, “Recognition by linear combination of
models”. Pattern Analysis and Machine Intelligence, vol. 13. 1991,
fitting polynomial coefficients of complex invariant pp. 992-1006.
moments can be used to classify the images for recognition. [7] C. Kan and M. D. Srinath, “Invariant character recognition with
Zernike and orthogonal Fourier-Mellin moments”. Pattern
V. CONCLUSION Recognition. 2002, pp. 143-154 .
A novel method combined non-redundant complex [8] A. Y. S. Mostafa, D. Psaltis, “Image normalization by complex
moments with polynomial curve fitting has been proposed in moments”. IEEE Trans. on Pattern Analysis and Machine Intelligence,
vol.7, Jan. 1985, pp. 46-55.
the paper for image recognition. With the similarity
[9] S. Fumiki, S. Akihiro, “Discrete polynomial curve fitting to noisy
measurement between the fitting coefficients to describe real data”. Lecture Notes in Computer Science. 2012, pp. 75-89.
ship images, the method has been proven to have good [10] J. D. Chen, Z. G. Zheng, “Probability and Statistics”. Peking
stability and can be an effective approach in image target University Press, Beijing, 2007.
recognition. [11] B. Zitova, J. Flusser, “Image registration methods: A survey”. Image
and Vision Computing, vol. 21,Nov. 2007, pp. 977-1000.
REFERENCES [12] J. Zhage, R. R. Korfhang, “A distance and angle similarity measure
[1] M. K. Hu, “Visual pattern recognition by moment invariants”. IRE method”. Journal of the American Society for Information Science,
Trans. on Information Theory, vol. 8, Feb. 1962, pp. 179-187. vol. 50, Sep. 1999, pp. 772-778.

357