Professional Documents
Culture Documents
Roshni Ravi
Josemartin M.J.
INTRODUCTION
are
centered
around
zero
frequency
{
,
,, }.
A. Mathematical Background
The Fourier transform encompasses a vast area of each
and every field of engineering. Fourier transform takes a
function from time/spatial domain and maps it onto
frequency domain [2]. For a continuous time domain
signal, its Fourier transform is given by
n=0, 1, 2, . N, is the frequency sample.
The CDFT is defined as
I.
468
TABLE I.
Time
domain
values
Corresponding
frequency
domain values
Error
Corresponding
difference
time and frequency
between time
values after 71
and frequency
iterations
domain values
0.0898
0.9102
0.4067
-0.0745
2.8912
-0.0745
0.1088
4.0601
0.0989
-0.0601
4.8649
-0.0708
0.1351
6.0435
0.1351
-0.0435
6.8249
-0.0435
0.1751
8.0189
0.1751
-0.0189
8.7575
0.0256
0.2425
9.9726
0.2330
10
0.0274
10.6268
0.1393
11
0.3732
2.0745
0.3732
12
0.1402
11.8598
0.4580
13
0.6931
12.3069
0.9595
14
0.5352
13.4648
1.5672
15
1.9669
13.0331
2.5983
16
3.8288
12.1712
3.8287
17
42.0167
-25.0167
5.1250
for = 1,2..,K
the maximum difference is given by
Now a new signal with a smaller maximum difference
is given by
where is the gain constant of the iteration loop. For fast
convergence should be as large as possible but stability
condition of the iteration limits its value. The iteration
method puts no constraint over the initial input signal
except that the signal should be symmetric.
Flexibility of iterative method lies at its choice for
initial input signal. When ramp signal is the input signal
we obtain a Fourier invariant signal as shown in Fig. 1 and
values depicted in Table I. Similarly when Gaussian signal
469
Bandwidth
(B)
Absolute
Difference (T-B)
0.0584
0.0654
0.0070
0.1369
0.0175
0.1193
TABLE III.
Error
Corresponding time
difference
and frequency
between time
domain values after
and frequency
71 iterations
domain values
Time
domain
values
Corresponding
frequency
domain values
0.0439
0.0061
-0.0378
0.0287
0.0634
-0.0062
-0.0696
-0.0062
0.0895
0.0063
-0.0832
-0.0069
0.1234
-0.0064
-0.1299
0.0066
0.1664
0.0066
-0.1597
-0.0069
0.2191
-0.0069
-0.2261
0.0073
0.2821
0.0073
-0.2748
-0.0040
0.3549
-0.0078
-0.3627
0.0083
0.4363
0.0084
-0.4280
-0.0025
0.5243
-0.0091
-0.5334
0.0101
0.6157
0.0101
-0.6056
0.0252
0.7066
-0.0113
-0.7179
0.0981
0.7926
0.0128
-0.7798
0.2088
0.8688
-0.0140
-0.8828
0.2088
0.9308
0.0423
-0.8885
0.3661
0.9745
0.5193
-0.4552
0.5193
0.9971
2.3611
1.3640
0.6226
Step 1: The values for the elements of centre column and
centre row of X are assigned as follows.
Step 2: Two weights and to be used for assigning
values to an arbitrary coordinate are determined
using as follows.
Step 3: Interpolated value is assigned to the arbitrary point
by using the weights and
The above steps can be repeated until all the elements of
the kernel matrix are interpolated.
2D-DFT can be used to analyze the property of the
generated 2D matrix. 2D-DFT for an
matrix/image is mathematically defined by [3] as
follows.
II.
470
Figure 4. 2D matrix constructed from 1D Fourier invariant signal by iterative method with ramp as the initial signal
Figure 5. Fourier Invariant (FI) 2D matrix from 1D Fourier invariant signal by iterative method with Gaussian as the initial signal
471
III.
TABLE V.
APPLICATIONS
Original image
Kernel
Size
FI 2D
Matrix
5x5
0.5779
(=2.0)
0.5703
Gaussian Matrix
(=2.5)
(=3.0)
(=3.5)
0.5604
0.5542
0.5502
(=4.0)
0.5475
0.5056
7x7
0.5855
0.5550
0.5348
0.5211
0.5119
9X9
0.6000
0.5486
0.5138
0.4884
0.4709
0.4587
11X11
0.5917
0.5364
0.4935
0.4594
0.4345
0.4166
13X13
0.5948
0.5383
0.4869
0.4419
0.4075
0.3823
15X15
0.5849
0.5356
0.4826
0.4333
0.3933
0.3630
17X17
0.5797
0.5368
0.4845
0.4323
0.3872
0.3514
a)
b)
c)
d)
Figure 6. Test image and the image after adding Gaussian noise
TABLE IV.
(=2.0)
(=2.5)
(=3.0)
(=3.5)
(=4.0)
5x5
Fourier
Invariant
2D
Kernel
Matrix
12.4551
14.6221
15.0164
15.2414
15.3803
15.4717
7x7
12.5063
15.9307
16.7581
17.2503
17.5600
17.7657
9X9
12.6149
16.4176
17.7144
18.5480
19.0931
19.4626
11X11
13.2154
16.7788
18.4144
19.5697
20.3672
20.9247
13X13
13.7201
16.8136
18.6440
20.0614
21.1069
21.8690
15X15
14.3134
16.8223
18.7366
20.3279
21.5812
22.5389
17X17
14.6584
16.8199
18.7619
20.4457
21.8466
22.9698
Kernel
size
Gaussian Matrix
e)
f)
Figure 7. Filtered image using a) 5X5 Fourier invariant kernel b) 5X5
Gaussian kernel with standard deviation 2.0 c) 5X5 Gaussian kernel
with standard deviation 2.5 d) 5X5 Gaussian kernel with standard
deviation 3.0 e) 5X5 Gaussian kernel with standard deviation 3.5
f) 5X5 Gaussian kernel with standard deviation 4.0
472
TABLE VI.
Kernel
Size
FI 2D
Matrix
(=2.0)
24.8646
TABLE VII.
(=4.0)
24.3740
3.0303
2.7576
2.7273
2.6364
2.4545
2.6364
1.8485
1.8485
1.6970
1.6667
1.7273
1.8485
1.0606
1.0606
2.6364
2.1515
2.6061
1.8485
1.2727
2.3939
2.1515
5x5
26.2578
7x7
26.2223
24.1201
23.6803
23.4289
23.2743
23.1731
9X9
26.1471
23.8586
23.1983
22.7989
22.5473
22.3808
11X11
25.7432
23.6696
22.8616
22.3331
21.9862
21.7516
2.4545
3.0909
2.8788
1.1515
2.3333
1.5455
1.7576
13X13
25.4176
23.6516
22.7540
22.1176
21.6763
21.3682
2.6061
4.3030
3.1212
3.1818
2.8182
2.8182
3.0303
15X15
25.0500
23.6471
22.7110
22.0029
21.4833
21.1061
3.0303
2.4242
3.7273
1.9091
1.1818
1.1818
1.4242
17X17
24.8430
23.6483
22.6992
21.9528
21.3771
20.9416
4.4242
1.2424
0.6970
3.0000
3.0303
2.9394
2.6667
IV. CONCLUSION
The 2D Fourier invariant kernel matrix constructed using
the discrete Fourier invariant signal generated by the
iterative method can be used as a novel image smoothing
filter which provides optimum spatial localization and high
frequency noise suppression. It can also be used as an
optimal 2D window for spatial-frequency spectral analysis of
images.
REFERENCES
[1]
[2]
[3]
[4]
2D Fourier invariant
kernel matrix X
Test image A
Figure 8. Illustration of 2D windowing on a test image A by using the
Fourier invariant kernel matrix X
473