Multi Resolution, Multi Sensor Image Fusion

Techniques
Presentation
By
Ch. Ramesh (Regd. No: 0603PH0645)

Research Scholar
Under the Supervision of
Supervisor

Co-Supervisor

Dr. N.B. Venkateswarlu

Professor,
AITAM, Tekkali

Dr. J.V.R. Murthy

Professor,
JNTUK, Kakinada

MOTIVATION
In this era of Information Technology, advancements in sensor technology,
communication technology, computers are churning the unveiling of new
applications which uses images extensively. As a result, the ability to store, access
and transmit image, video information in an efficient manner has become very
crucial. A large no. of bits are typically required to represent or store even for a
simple single band digital image. Moreover, with the rapid advances in sensor
technology such as increased pixel, band, gray level resolutions, image sizes are
growing exponentially and requires huge amount of storage and transmission
bandwidth. Thus uncompressed Multi resolution, Multi sensor images are generally
large in size and require huge amount of storage and transmission band width.
Despite rapid progress in mass-storage density, processor speeds, and digital
communication system performance, demand for data storage capacity and datatransmission bandwidth continues to outstrip the capabilities of available
technologies. At present one of the prominent solutions that is widely used is to
compress data before its storage and transmission, and decompress it at the receiver
or in the application as and when needed. In this study we compare the classification
accuracy of remotely sensed original images and proposed compressed remotely
sensed images so as to investigate and recommend satellite designers to employ onboard compression and broadcast only compressed images to land which saves
bandwidth, space, increases satellite life etc.

OBJECTIVES
If satellite communicates compressed images directly from
satellite it certainly saves bandwidth requirements of a satellite
mission.
Storage requirement reduces by many folds as we will be

storing only compressed images.

Power requirement needs of the storage system reduces
indirectly, giving path to green computing in related applications.
Positive impact on satellite life.
Loading and storing of images take less time compared to

original images, thus response times of imaging analysis systems

increases.

Basic Architecture of JPEG Compression

8 X 8 Blocks

Level
offset

FDCT

Quanti
zation

Zig-Zag
Pattern

Encod
ing

Compressed
Image Data

Source Image Data

JPEG Encoding Steps

Compressed
Image Data

Decoding

Zig-Zag
Pattern

Dequanti
zation

IDCT

Level
Adding
Reconstructed
Image Data

JPEG Decoding Steps

Encoding
8 X 8 Blocks

Level
offset

FDCT

Quanti
zation

Zig-Zag
Pattern

Encod
ing

Compressed
Image Data

Source Image Data

1. Divide the image into 8 X 8 subimages and process them in a

left-to-right and from top-to-bottom fashion.
2. Level Shifting.
3. Calculate the two-dimensional DCT of each subimage.
4. Quantize each DCT by using standard JPEG Quantization matrix.
5. Construct a one-dimensional array from these quantized
coefficients, using a zig-zag pattern.
6. Code this array, using JPEG Huffman tables.

8 X 8 Subimage
52
63
62
63
67
79
85
87

55
59
59
58
61
65
71
79

61
66
68
71
68
60
64
69

66
90
113
122
104
70
59
68

70 61 64
109 85 69
144 104 66
154 106 70
126 88 68
77 68 58
55 61 65
65 76 78

Level shift
73
72
73
69
70
75
83
94

-76
-65
-66
-65
-61
-49
-43
-41

-73
-69
-69
-70
-67
-63
-57
-49

-67
-62
-60
-57
-60
-68
-64
-59

-62
-38
-15
-6
-24
-58
-69
-60

-58
-19
16
26
-2
-51
-73
-63

-67
-43
-24
-22
-40
-65
-67
-52

-64
-59
-62
-58
-60
-70
-63
-50

-55
-56
-55
-59
-58
-53
-45
-34

DCT

Level shift
x
-76
-65
-66
-65
-61
-49
-43
-41

-73
-69
-69
-70
-67
-63
-57
-49

-67
-62
-60
-57
-60
-68
-64
-59

-62
-38
-15
-6
-24
-58
-69
-60

-58
-19
16
26
-2
-51
-73
-63

-67
-43
-24
-22
-40
-65
-67
-52

u
-64
-59
-62
-58
-60
-70
-63
-50

-55
-56
-55
-59
-58
-53
-45
-34

-415 -29 -62 25

7
-21 -62
9
-46
8 77 -25
-50
13 35 -15
11
-8 -13 -2
-10
1 3 -3
-4
-1 2 -1
-1
-1 -1 -2

55 -20 -1 3
11
-7 -6 6
-30 10
7 -5
-9
6
0 3
-1
1 -4 1
-1
0
2 -1
2
-3 1 -2
-1
1
0 -1

(2 x 1)u
(2 y 1)v
C (u, v) (u) (v) f ( x, y) cos
cos

2N
2N

x 0 y 0
N 1 N 1

for u, v = 0, 1, 2, . . . . , N -1

(u )

1/ N
2/ N

for
for

u 0
u 0

1/ N
2/ N

for
for

v 0
v 0

(v )

JPEG Quantization Matrix

16
12
14
14
18
24
49
72

11
12
13
17
22
35
64
92

10
14
16
22
37
55
78
95

DCT
-415 -29 -62 25
7
-21 -62
9
-46
8 77 -25
-50
13 35 -15
11
-8 -13 -2
-10
1 3 -3
-4
-1 2 -1
-1
-1 -1 -2

55 -20 -1 3
11
-7 -6 6
-30 10
7 -5
-9
6
0 3
-1
1 -4 1
-1
0
2 -1
2
-3 1 -2
-1
1
0 -1

16
19
24
29
56
64
87
98

24
26
40
51
68
81
103
112

40 51 61
58 60 55
57 69 56
87 80 62
109 103 77
104 113 92
121 120 101
100 103 99

Quantized DCT
-26
1
-3
-4
1
0
0
0

-3
-2
1
1
0
0
0
0

-6
-4
5
2
0
0
0
0

2
0
-1
-1
0
0
0
0

2
0
-1
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

zig-zag pattern
DCT after quantization
-26
1
-3
-4
1
0
0
0

-3
-2
1
1
0
0
0
0

-6
-4
5
2
0
0
0
0

2
0
-1
-1
0
0
0
0

2
0
-1
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

1-D coefficient sequence

[-26 -3 1 -3 -2 -6 2 -4 1 -4 1 1 5 0 2 0 0 -1 2 0 0 0 0 0 -1 -1 EOB]

1-D coefficient sequence

[-26 -3 1 -3 -2 -6 2 -4 1 -4 1 1 5 0 2 0 0 -1 2 0 0 0 0 0 -1 -1 EOB]

Encoding
1010110

0100 001 0100 0101 100001 0110 100011 001 100011 001

001 100101 11100110 110110 0110 11110100 000 1010 (92 bits)

Compression Ratio = 512/92 = 5.565

Decoding

Compressed
Image Data

Decoding

Zig-Zag
Pattern

Dequanti
zation

IDCT

Level
Adding
Reconstructed
Image Data

1.
2.
3.
4.
5.

Decode the array and Zig-Zag back

Multiply with normalization matrix
Calculate the two-dimensional IDCT
Level shift back
Place the subimages in their correct positions

1010110

0100 001 0100 0101 100001 0110 100011 001 100011 001

001 100101 11100110 110110 0110 11110100 000 1010

Decoding
1-D coefficient sequence
[-26 -3 1 -3 -2 -6 2 -4 1 -4 1 1 5 0 2 0 0 -1 2 0 0 0 0 0 -1 -1 EOB]

1-D coefficient sequence

[-26 -3 1 -3 -2 -6 2 -4 1 -4 1 1 5 0 2 0 0 -1 2 0 0 0 0 0 -1 -1 EOB]

zig-zag pattern

Regenerated array of quantized coefficients

-26
1
-3
-4
1
0
0
0

-3
-2
1
1
0
0
0
0

-6 2
-4 0
5 -1
2 -1
0
0
0
0
0
0
0
0

2
0
-1
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

JPEG Quantization Matrix

16
12
14
14
18
24
49
72

11
12
13
17
22
35
64
92

10
14
16
22
37
55
78
95

16
19
24
29
56
64
87
98

24
26
40
51
68
81
103
112

40
58
57
87
109
104
121
100

Regenerated array of quantized

coefficients
-26
1
-3
-4
1
0
0
0

-3
-2
1
1
0
0
0
0

-6 2
-4 0
5 -1
2 -1
0 0
0 0
0 0
0 0

2
0
-1
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

51
60
69
80
103
113
120
103

61
55
56
62
77
92
101
99

Denormalization
-416
12
-42
-56
18
0
0
0

-33
-24
13
17
0
0
0
0

-60 32 48
-56 0 0
80 -24 -40
44 -29 0
0
0 0
0
0 0
0
0 0
0
0 0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

Denormalization

After Inverse DCT

u
-416
12
-42
-56
18
0
0
0

-33
-24
13
17
0
0
0
0

-60 32 48
-56 0 0
80 -24 -40
44 -29 0
0
0 0
0
0 0
0
0 0
0
0 0

x
0
0
0
0
0
0
0
0

0
0
0
0
0
0
0
0

0
0
0
0
0
0 v
0
0

-70
-72
-68
-59
-54
-52
-45
-35

-64
-73
-78
-77
-75
-71
-59
-47

-61 -64 -69

-61 -39 -30
-58
-9 13
-57
0 22
-64 23 -13
-72 -54 -54
-70 -68 -67
-61 -66 -60

-66 -58
-40 -54
-12 -48
-13 -51
-44 -63
-71 -71
-67 -61
-48 -44

-50
-59
-64
-60
-56
-54 y
-50
-44

(2 x 1)u
(2 y 1)v
f ( x, y)
(u) (v)C (u, v) cos
cos

2N
2N

u 0 v0
N 1 N 1

After Inverse DCT

-70
-72
-68
-59
-54
-52
-45
-35

-64
-73
-78
-77
-75
-71
-59
-47

-61 -64 -69

-61 -39 -30
-58
-9 13
-57
0 22
-64 23 -13
-72 -54 -54
-70 -68 -67
-61 -66 -60

-66 -58
-40 -54
-12 -48
-13 -51
-44 -63
-71 -71
-67 -61
-48 -44

Level Adding
-50
-59
-64
-60
-56
-54
-50
-44

58
56
60
69
74
76
83
93

64
55
50
51
53
57
69
81

67
67
70
71
64
56
59
67

64
89
119
128
105
74
60
62

59
98
141
149
115
75
61
69

62
88
116
115
84
57
61
80

70
74
80
77
65
57
67
84

78
69
64
68
72
74
78
84

Input Subimage
52
63
62
63
67
79
85
87

55
59
59
58
61
65
71
79

Output Subimage
58
56
60
69
74
76
83
93

61 66 70 61 64 73
66 90 109 85 69 72
68 113 144 104 66 73
71 122 154 106 70 69
68 104 126 88 68 70
60 70 77 68 58 75
64 59 55 61 65 83
69 68 65 76 78 94

64
55
50
51
53
57
69
81

67
67
70
71
64
56
59
67

Error Subimage
-6
7
2
-6
-7
3
2
-6

-9
4
9
7
8
8
2
-2

-6
-1
-2
0
4
4
5
2

2
1
-6
-4
-1
-4
-1
6

11
11
-3
-5
11
2
-6
-4

-1
-3
-12
-9
4
11
0
-4

-6
-5
-5
3
-14 9
-7
1
3
-2
1
1
-2
5
-6 10

64
89
119
128
105
74
60
62

59
98
141
149
115
75
61
69

62
88
116
115
84
57
61
80

70
74
80
77
65
57
67
84

78
69
64
68
72
74
78
84

Input Subimage
52
63
62
63
67
79
85
87

55
59
59
58
61
65
71
79

61 66 70 61 64 73
66 90 109 85 69 72
68 113 144 104 66 73
71 122 154 106 70 69
68 104 126 88 68 70
60 70 77 68 58 75
64 59 55 61 65 83
69 68 65 76 78 94

Output Subimage
58
56
60
69
74
76
83
93

PSNR = 10 log10 (L2/MSE) = 32.647 dB

m 1 n 1

MSE

64
55
50
51
53
57
69
81

67
67
70
71
64
56
59
67

64
89
119
128
105
74
60
62

f ( x, y ) f ( x, y ) 2

x 0 y 0

mXn

35.343

59
98
141
149
115
75
61
69

62
88
116
115
84
57
61
80

70
74
80
77
65
57
67
84

78
69
64
68
72
74
78
84

Bolivia

Monolake

Saltlake

Owensvalley

Santacruzvalley

Test Images for Experiments (www.usgs.gov)

Experimental Results

Image : Bolivia

Size: 1015 X 877

Image : Monolake

Size: 1040 X 1040

Band

No. of
Bits
required

Compression
Ratio

PSNR

39.785

Band1

410486

21.079

39.871

25.833

37.939

Band2

306006

28.276

42.237

394036

18.072

34.539

Band3

399635

21.651

40.435

Band4

623322

11.424

30.959

Band4

366063

23.637

41.097

Band5

694304

10.256

29.986

Band5

578965

14.945

37.724

Band6

321611

22.142

37.221

Band6

249582

34.669

46.608

Band7

559107

12.736

31.400

Band7

439079

19.706

39.828

Total

3083120

18.478

34.547

Total

2746816

25.200

41.114

Band

No. of
Bits
required

Compression
Ratio

PSNR

Band1

215085

33.108

Band2

275655

Band3

Experimental Results

Image : Owensvalley

Size: 1047 X 917

Image : Santacruz

Size: 1000 X 1000

Band

No. of Bits
required

Compression
Ratio

PSNR

Band

No. of Bits
required

Compression
Ratio

PSNR

Band1

537518

14.289

32.756

Band1

864451

9.254

33.580

Band2

409426

18.759

35.207

Band2

949586

8.424

33.017

Band3

558702

13.747

32.590

Band3

1069931

7.477

32.202

Band4

565589

13.580

32.645

Band4

1046979

7.641

32.305

Band5

991131

7.749

28.087

Band5

1117245

7.160

32.103

Band6

478658

16.046

35.711

Band6

505166

15.836

40.282

Band7

708596

10.839

30.718

Band7

1076416

7.432

32.515

Total

4249620

14.459

32.530

Total

6629774

9.653

33.714

Experimental Results

Image : Saltlake

Size: 1035 X 920

Band

No. of Bits
required

Compression
Ratio

PSNR

Band1

378973

20.100

35.772

Band2

431445

17.656

34.713

Band3

562593

13.540

32.400

Band4

532688

14.300

32.960

Band5

683441

11.145

30.630

Band6

534180

14.260

33.508

Band7

616314

12.359

31.514

Total

3739634

16.295

33.071

Original Image

Band1

Band2

Band3

Band4

JPEG Compressed Image

Error Image

Original Image

Band5

Band6

Band7

JPEG Compressed Image

Error Image

Original Image

JPEG Compressed Image

Modified JPEG System

1. By K-Means Clustering
2. By Filters
3. By Zooming-Shrinking Technique

K-means Algorithm
Step1: Select K random cluster centers z1, z2, ..zk from the training
data and call it as code vectors
Step2: The training vectors x1, x2, ..xN are classified into the cluster
from which their Euclidean distance is minimum.
Step3: Compute the sum vector for every cluster by adding the
corresponding components of all the training vectors that
belong to the same cluster.
Step4: Compute the centroid for each cluster z1 , z2 ...... zk by
dividing the individual components of the sum vector by the
cluster strength.
Step5: If z1 z1 ; z2 z2 .......... .zk zk then stop else take
z1 as z1 and z2 as z2 ........ zk as zk then go to step 3.

Modified JPEG System by K-Means Clustering

Modified JPEG System by K-Means Clustering

1) The image is first subdivided into pixel blocks of size 8x8 which are
processed from left to right, top to bottom.
2) For each block its 64 pixels are level shifted by subtracting the quantity
L/2 where L is the gray level resolution.
3) The 2D-DCT of each block is computed.
4) Quantize the DCT blocks by standard quantization matrix.
5) The quantized blocks are grouped into clusters by using K-means
Clustering .
6) Encode the cluster centers and indexes and send to the receiver.
7) The receiver decodes the cluster centers and indexes.
8) Based on the cluster centers and indexes the receiver form the output
image.
9) Compute the error between the input image and the output image.

Experimental Results
Image : Bolivia

Size: 1015 X 877

Conventional Approach: No. of bits: 3083120, PSNR: 34.547

No. of Clusters

256

512

1024

2048

4096

8192

No. of Bits for

Proposed Approach

213155

255615

448934

563490

707058

2316670

% of Saving

93.086

91.709

85.438

81.723

77.066

24.859

PSNR for Proposed

Approach

28.369

29.196

30.762

31.332

31.8330

33.243

% of Loss

17.882

15.489

10.956

9.306

7.864

3.774

4096

8192

% of saving

% of Loss

100
80
60
40
20
0
256

512

1024

2048

Experimental Results
Image : Monolake

Size: 1040 X 1040

Conventional Approach: No. of bits: 2749816, PSNR: 41.114

No. of Clusters

256

512

1024

2048

4096

8192

No. of Bits for Proposed

Approach

694972

802534

964972

1216133

1676544

2305239

% of Saving

74.726

70.814

64.907

41.114

39.03

16.167

PSNR for Proposed

Approach

35.133

35.675

36.309

37.080

37.788

38.978

% of Loss

14.547

13.229

11.687

9.811

8.089

5.195

% of Saving

% of Loss

80
70
60
50
40
30
20
10
0
256

512

1024

2048

4096

8192

Experimental Results
Image : Owensvally

Size: 1047 X 917

Conventional Approach: No. of bits: 4249620, PSNR: 32.530

No. of Clusters

256

512

1024

2048

4096

8192

No. of Bits for Proposed

Approach

296741

379188

548239

813606

1428936

2952130

% of Saving

93.017

91.077

87.099

80.854

66.37

30.531

PSNR for Proposed

Approach

26.398

27.145

28.014

28.584

29.992

30.392

% of Loss

18.850

16.553

13.882

12.130

7.802

6.572

% of Saving

% of Loss

100
80
60
40
20
0
256

512

1024

2048

4096

8192

Experimental Results
Image : Saltlake

Size: 1035 X 920

Conventional Approach: No. of bits: 3739634, PSNR: 33.071

No. of Clusters

256

512

1024

2048

4096

8192

No. of Bits for Proposed

Approach

233014

281613

448183

711553

1394363

2700837

% of Saving

93.769

92.469

88.015

80.972

62.713

27.778

PSNR for Proposed

Approach

26.954

27.742

28.997

29.671

30.485

31.647

% of Loss

18.496

16.113

12.318

10.280

7.819

4.305

% of Saving

% of Loss

100
80
60
40
20
0
256

512

1024

2048

4096

8192

Experimental Results
Image : Santacruz

Size: 1000 X 1000

Conventional Approach: No. of bits: 6629774, PSNR: 33.714

No. of Clusters

256

512

1024

2048

4096

8192

No. of Bits for Proposed

Approach

868045

1030519

1268406

1714102

2646135

4692478

% of Saving

86.906

84.456

80.868

74.145

60.087

29.221

PSNR for Proposed

Approach

25.675

26.097

26.513

27.121

28.232

30.224

% of Loss

23.844

22.592

21.359

19.555

16.260

10.351

% of Saving

% of Loss

100
80
60
40
20
0
256

512

1024

2048

4096

8192

Original Image

Band1

Band2

Band3

Band4

JPEG Compressed Image

256 Clusters

Error Image

Original Image

Band5

Band6

Band7

JPEG Compressed Image

256 Clusters

Error Image

JPEG Compressed Image (Band 1 ) with 256 Clusters

Original Image

Band1

Band2

Band3

Band4

JPEG Compressed Image

8192 Clusters

Error Image

Original Image

Band5

Band6

Band7

JPEG Compressed Image

8192 Clusters

Error Image

JPEG Compressed Image (Band 1) with 8192 Clusters

Original Image

JPEG Compression
with 256 Clusters

JPEG Compression
with 8192 Clusters

Modified JPEG System by Filters

Why Filtering ?

Modified JPEG System by Filters

8 X 8 Subimage
52
63
62
63
67
79
85
87

55
59
59
58
61
65
71
79

61
66
68
71
68
60
64
69

66
90
113
122
104
70
59
68

70 61 64
109 85 69
144 104 66
154 106 70
126 88 68
77 68 58
55 61 65
65 76 78

8 X 8 Subimage
73
72
73
69
70
75
83
94

55
54
55
54
54
54
55
55

54
54
55
54
55
54
55
55

DCT
-415 -29 -62 25
7
-21 -62
9
-46
8 77 -25
-50
13 35 -15
11
-8 -13 -2
-10
1 3 -3
-4
-1 2 -1
-1
-1 -1 -2

55 -20 -1 3
11
-7 -6 6
-30 10
7 -5
-9
6
0 3
-1
1 -4 1
-1
0
2 -1
2
-3 1 -2
-1
1
0 -1

54
54
55
54
55
54
55
54

55
54
54
54
55
55
54
54

55
54
55
55
54
55
54
54

55
54
54
55
54
55
54
54

55
54
55
55
54
54
54
54

54
55
55
55
54
55
55
55

0
0
0
0
0
0
0
0

0
0
0
-1
0
1
0
0

DCT
-588
0
0
0
0
1
1
0

0
-1
0
0
0
0
0
2

0
0
-1
0
1 0
-2 0
-1 0
0 1
-1 1
0 -1

0
0
0
0
0
0
0
-1

-1
0
0
0
0
0
0
0

Modified JPEG System by Filters

8 X 8 Subimage
52
63
62
63
67
79
85
87

55
59
59
58
61
65
71
79

61
66
68
71
68
60
64
69

66
90
113
122
104
70
59
68

70 61 64
109 85 69
144 104 66
154 106 70
126 88 68
77 68 58
55 61 65
65 76 78

8 X 8 Filtered image
73
72
73
69
70
75
83
94

52
63
62
63
67
79
85
87

55
61
63
64
66
69
73
79

61
71
78
80
75
69
67
69

DCT
-415 -29 -62 25
7
-21 -62
9
-46
8 77 -25
-50
13 35 -15
11
-8 -13 -2
-10
1 3 -3
-4
-1 2 -1
-1
-1 -1 -2

55 -20 -1 3
11
-7 -6 6
-30 10
7 -5
-9
6
0 3
-1
1 -4 1
-1
0
2 -1
2
-3 1 -2
-1
1
0 -1

66
87
97
108
95
76
65
68

70 61 64
94 86 74
114 101 79
118 103 79
101 91 75
79 74 71
67 67 73
65 76 78

73
72
73
69
70
75
83
94

DCT
-413 -28 -40
-7
-17 -47
-38
5 54
-35
9 17
-1
-4 -5
-7
0
1
-2
-2 3
-3
-2 2

15 23
2 2
-15 -5
-8
1
-4 -1
-2
1
-1 -1
-1 -1

-4
-2
-1
2
0
-1
-2
-2

4 -1
-2
4
1 0
-3
2
-1 -3
-1
3
1
1
0
2

Modified JPEG System by Filters

1. Mean Filtering
2. Median Filtering
3. Outlier Mean Filtering
4. Outlier Median Filtering

Mean Filtering Algorithm

Step 1: A Two dimensional window of size 3x3 is selected and

centered around the processed pixel in the image.
Step 2: Calculate the mean of the pixels in the selected window.
Step 3: Replace the processed pixel by mean pixel value.

Step 4: Repeat the steps 1 to 3 until the processing is completed

for the entire image.

Median Filtering Algorithm

Step 1: A Two dimensional window of size 3x3 is selected and

centered around the processed pixel in the image.
Step 2: Sort the pixels in the selected window according to the
ascending order and find the median pixel value.
Step 3: Replace the processed pixel by median pixel value.
Step 4: Repeat the steps 1 to 3 until the processing is completed
for the entire image.

Outlier
An outlier is an observation that is numerically distant from
the rest of the data

Confidence
Level

80%

90%

95%

98%

99%

99.8%

99.9%

Critical
Values

1.28

1.645

1.96

2.33

2.58

3.08

3.27

Outlier Mean Filtering Algorithm

Step 1: A Two dimensional window of size 3x3 is selected and
centered around the processed pixel in the image.
Step 2: For each pixel, calculate Mean and standard deviation of
its neighboring 3x3 pixels.
Step 3: If a pixels value is observed to be outlier (not in the range
of Mean C*)
then its value is taken as itself
else
Mean is taken as its filtered value.
Step 4: Repeat the steps 1 to 3 until the processing is completed
for the entire image.

Outlier Median Filtering Algorithm

Step 1: A Two dimensional window of size 3x3 is selected and

centered around the processed pixel in the image.
Step 2: For each pixel, calculate Median and standard deviation
of its neighboring 3x3 pixels.
Step 3: If a pixels value is observed to be outlier (not in the range
of Median C*)
then its value is taken as itself
else
Median is taken as its filtered value.
Step 4: Repeat the steps 1 to 3 until the processing is completed
for the entire image.

Modified JPEG System by Mean Filtering

8 X 8 Blocks

Mean
Filter
ing
Source Image Data

Compressed
Image Data

FDCT

Quanti
zation

Zig-Zag
Pattern

Encod
ing

Compressed
Image Data

JPEG Encoding Steps

Decoding

Zig-Zag
Pattern

Dequanti
zation

IDCT

Level
Adding
Reconstructed
Image Data

JPEG Decoding Steps

Modified JPEG System by Median Filtering

8 X 8 Blocks

Median
Filtering
Source Image Data

Compressed
Image Data

FDCT

Quanti
zation

Zig-Zag
Pattern

Encod
ing

Compressed
Image Data

JPEG Encoding Steps

Decoding

Zig-Zag
Pattern

Dequanti
zation

IDCT

Level
Adding
Reconstructed
Image Data

JPEG Decoding Steps

Modified JPEG System by Outlier Mean Filtering

8 X 8 Blocks

Outlier
Mean
Filtering
Source Image Data

Compressed
Image Data

FDCT

Quanti
zation

Zig-Zag
Pattern

Encod
ing

Compressed
Image Data

JPEG Encoding Steps

Decoding

Zig-Zag
Pattern

Dequanti
zation

IDCT

Level
Adding
Reconstructed
Image Data

JPEG Decoding Steps

Modified JPEG System by Outlier Median Filtering

8 X 8 Blocks

Outlier
Median
Filtering
Source Image Data

Compressed
Image Data

FDCT

Quanti
zation

Zig-Zag
Pattern

Encod
ing

Compressed
Image Data

JPEG Encoding Steps

Decoding

Zig-Zag
Pattern

Dequanti
zation

IDCT

Level
Adding
Reconstructed
Image Data

JPEG Decoding Steps

Experimental Results
Image : Bolivia

Size: 1015 X 877

Conventional Approach: No. of bits: 3083120, PSNR: 34.547

Outlier Mean Filter
Filter

Mean Filter
C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

2077249

2077242

2077250

2124088

2181853

2306572

2464900

2682210

% of Saving

32.625

32.625

32.625

31.105

29.232

25.187

20.051

13.003

PSNR for
Proposed
Approach

32.625

32.625

32.625

32.633

32.646

32.680

32.755

32.932

% of Loss

5.563

5.563

5.563

5.540

5.502

5.404

5.187

4.674

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

35
30
25
20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Monolake

Size: 1040 X 1040

Conventional Approach: No. of bits: 2749816, PSNR: 41.114

Outlier Mean Filter
Filter

Mean Filter
C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

2248277

2248254

2248355

2250035

2255578

2279061

2324609

2421179

% of Saving

18.239

18.239

18.236

18.175

17.973

17.119

15.463

11.951

PSNR for
Proposed
Approach

38.186

38.185

38.185

38.185

38.190

38.211

38.266

38.408

% of Loss

7.121

7.124

7.124

7.124

7.111

7.060

6.927

6.581

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Owensvally

Size: 1047 X 917

Conventional Approach: No. of bits: 4249620, PSNR: 32.530

Outlier Mean Filter
Filter

Mean Filter
C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

2777127

2775122

2775119

2815187

2881656

3040273

3252354

3584143

% of Saving

34.649

34.649

34.697

33.754

32.190

28.457

23.467

15.659

PSNR for
Proposed
Approach

29.740

29.740

29.740

29.755

29.778

29.826

29.934

30.275

% of Loss

8.576

8.576

8.576

8.530

8.459

8.312

7.980

6.932

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

40
35
30
25
20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Saltlake

Size: 1035 X 920

Conventional Approach: No. of bits: 3739634, PSNR: 33.071

Outlier Mean Filter
Filter

Mean Filter
C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

2488597

2488616

2488595

2519675

2570047

2699906

2888193

3180244

% of Saving

33.453

33.452

33.453

32.622

31.275

27.802

22.768

14.958

PSNR for
Proposed
Approach

31.244

31.244

31.244

31.252

31.263

31.289

31.365

31.589

% of Loss

5.524

5.524

5.524

5.500

5.467

5.388

5.158

4.481

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

40
35
30
25
20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Santacruz

Size: 1000 X 1000

Conventional Approach: No. of bits: 6629774, PSNR: 33.714

Outlier Mean Filter
Filter

Mean Filter
C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

4861536

4861534

4861601

4864143

4874122

4935041

5100924

5521025

% of Saving

26.671

26.671

26.670

26.631

26.481

25.562

23.060

16.723

PSNR for
Proposed
Approach

31.002

31.002

31.002

31.004

31.011

31.061

31.209

31.706

% of Loss

8.044

8.044

8.004

8.038

8.017

7.869

7.430

5.955

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

30
25
20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Bolivia

Size: 1015 X 877

Conventional Approach: No. of bits: 3083120, PSNR: 34.547

Filter

Outlier Median Filter

Median
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

2212011

2212009

2212013

2258018

2312490

2429264

2579173

2790591

% of Saving

28.254

28.254

28.254

26.761

24.995

21.207

16.345

9.488

PSNR for
Proposed
Approach

32.903

32.903

32.903

32.948

33

33.123

33.345

33.737

% of Loss

4.758

4.758

4.758

4.628

4.477

4.121

3.479

2.344

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

30
25
20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Monolake

Size: 1040 X 1040

Conventional Approach: No. of bits: 2749816, PSNR: 41.114

Filter

Outlier Median Filter

Median
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

2388899

2388936

2388983

2390644

2395909

2419040

2464756

2556379

% of Saving

13.125

13.123

13.122

13.122

12.870

12.029

10.366

7.034

PSNR for
Proposed
Approach

40.44

40.441

40.44

40.442

40.451

40.485

40.567

40.746

% of Loss

1.639

1.636

1.639

1.634

1.612

1.529

1.330

0.895

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

14
12
10
8
6
4
2
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Owensvally

Size: 1047 X 917

Conventional Approach: No. of bits: 4249620, PSNR: 32.530

Filter

Median
Filter

Outlier Median Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

3032712

3032716

3032711

3069953

3130758

3276024

3475985

3793053

% of Saving

28.635

28.635

28.635

27.759

26.328

22.910

18.204

10.743

PSNR for
Proposed
Approach

29.877

29.877

29.877

29.922

29.978

30.094

30.334

31.004

% of Loss

8.155

8.155

8.155

8.017

7.845

7.488

6.75

4.691

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

35
30
25
20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Saltlake

Size: 1035 X 920

Conventional Approach: No. of bits: 3739634, PSNR: 33.071

Filter

Outlier Median Filter

Median
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

2727405

2727399

2727390

2755637

2798599

2914734

3081383

3348312

% of Saving

27.067

27.067

27.067

26.312

25.163

22.058

17.602

10.464

PSNR for
Proposed
Approach

31.314

31.314

31.314

31.341

31.378

31.466

31.653

32.074

% of Loss

5.312

5.312

5.312

5.231

5.119

4.853

4.287

3.014

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

30
25
20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Experimental Results
Image : Santacruz

Size: 1000 X 1000

Conventional Approach: No. of bits: 6629774, PSNR: 33.714

Filter

Outlier Median Filter

Median
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

5268265

5268290

5268315

5270953

5281047

5339990

5496374

5871691

% of Saving

20.536

20.535

20.535

20.495

20.343

19.454

17.095

11.434

PSNR for
Proposed
Approach

31.429

31.429

31.429

31.433

31.445

31.518

31.718

32.332

% of Loss

6.777

6.777

6.777

6.765

6.730

6.513

5.920

4.099

No. of Bits for

Proposed
Approach

% of Saving

% of Loss

25
20
15
10
5
0
Mean
Filter

C=3.27

C=3.08

C=2.58

C=2.33

C=1.96

C=1.645

C=1.28

Modified JPEG System By Zooming-Shrinking Technique

Interpolation: Interpolation is a method of constructing new data
Points within the range of a discrete set of known data points
1. Nearest Neighborhood Interpolation
2. Bilinear Interpolation
3. Bicubic Interpolation

Modified JPEG System By Zooming-Shrinking Technique

Stage-1:
Shrink the
Image M times

Input Image

Input
Image

Zoomed
Image

Zoom the Image

M times

Difference
Image

1. Shrink the input image M (Shrinking factor) times, the resulting image is called
shrinked image.
2. Zoom the shrinked image M (Zooming factor) times, the resulting image is called
Zoomed image.
3. Find the difference image between the Zoomed image and the input image, the
resulting image is called the Difference image.

Modified JPEG System By Zooming-Shrinking Technique

Stage -2:
Shrinked
Image

Divide into
8x8 blocks

Level
Offset

DeQuantiza
Compressed
Decoding

Shrinked Image
tion

IDCT

FDCT

Level
Adding

Quantization

8x8 block
Merger

Encoding

Compressed
Shrinked image

Zoom by
Reconstructed

Zoomed Image
M times

1. The shrinked image is first subdivided into pixel blocks of size 8x8 which are processed
from left to right, top to bottom.
2. For each block its 64 pixels are level shifted by subtracting the quantity L/2 where L is the
gray level resolution.
3. The 2D-DCT of each block is computed.
4. Quantize the DCT blocks by standard quantization matrix.
5. Form a 1-D sequence of Quantized Coefficients by using Zigzag pattern
6. Coding the coefficients using JPEG Huffman tables
7. The receiver decodes the received codes and forms the reconstructed shrinked image
8. Zoom the reconstructed shrinked image M times

Modified JPEG System By Zooming-Shrinking Technique

Stage-3:
Difference
Image

Divide
into 8x8
blocks

Level
Offset

DeQuantiza
Compressed

Decoding
Difference
tion
Image

FDCT

IDCT

Quantization

Level
Adding

Encoding

8x8 block
Merger

Compressed
Difference image

Reconstructed
Difference
Image

1. The difference image is first subdivided into pixel blocks of size 8x8 which are
processed from left to right, top to bottom.
2. For each block its 64 pixels are level shifted by subtracting the quantity L/2 where L is
the gray level resolution.
3. The 2D-DCT of each block is computed.
4. Quantize the DCT blocks by standard quantization matrix.
5. Form a 1-D sequence of Quantized Coefficients by using Zigzag pattern
6. Coding the coefficients using JPEG Huffman tables
7. The receiver decodes the received codes and forms the reconstructed difference Image.

Modified JPEG System By Zooming-Shrinking Technique

Stage- 4:

Reconstructed
Zoomed Image

Input Image

Reconstructed
Difference Image

Output Image

Output Image

Error

1. Add the reconstructed zoomed image obtained at stage 2 and the reconstructed
difference image obtained at stage 3 and the resulting image is called output image
2. Compute the error (in terms of PSNR) between the input image and output image

Experimental Results
Image : Bolivia

Size: 1015 X 877

Conventional Approach: No. of bits: 3083120, PSNR: 34.547

Interpolation

Nearest

Bilinear

Bicubic

No. of Bits for

Proposed Approach

2929956

2300183

2141462

% of Saving

4.967

25.39

30.542

PSNR for Proposed

Approach

30.431

31.801

31.797

% of Loss

11.914

7.948

7.960

% of Saving

% of Loss

35
30
25
20
15
10
5
0
Nearest

Bilinear

Bicubic

Experimental Results
Image : Monolake

Size: 1040 X 1040

Conventional Approach: No. of bits: 2749816, PSNR: 41.114

Interpolation

Nearest

Bilinear

Bicubic

No. of Bits for

Proposed Approach

2115854

1910807

1803052

% of Saving

23.054

30.511

34.430

PSNR for Proposed

Approach

36.644

36.777

37.047

% of Loss

10.872

10.548

9.892

% of Saving

% of Loss

40
35
30
25
20
15
10
5
0
Nearest

Bilinear

Bicubic

Experimental Results
Image : Owensvally

Size: 1047 X 917

Conventional Approach: No. of bits: 4249620, PSNR: 32.530

Interpolation

Nearest

Bilinear

Bicubic

No. of Bits for

Proposed Approach

4021501

3052130

2838005

% of Saving

5.367

28.178

33.217

PSNR for Proposed

Approach

27.532

29.135

29.203

% of Loss

15.364

10.436

10.227

% of Saving

% of Loss

35
30
25
20
15
10
5
0
Nearest

Bilinear

Bicubic

Experimental Results
Image : Saltlake

Size: 1035 X 920

Conventional Approach: No. of bits: 3739634, PSNR: 33.071

Interpolation

Nearest

Bilinear

Bicubic

No. of Bits for

Proposed Approach

3596665

2639599

2431475

% of Saving

3.823

29.415

34.98

PSNR for Proposed

Approach

28.794

30.307

30.313

% of Loss

12.932

8.357

8.339

% of Saving

% of Loss

40
35
30
25
20
15
10
5
0
Nearest

Bilinear

Bicubic

Experimental Results
Image : Santacruz

Size: 1000 X 1000

Conventional Approach: No. of bits: 6629774, PSNR: 33.714

Interpolation

Nearest

Bilinear

Bicubic

No. of Bits for

Proposed Approach

5746830

4621088

3918882

% of Saving

13.317

30.297

40.88

PSNR for Proposed

Approach

27.150

29.138

29.294

% of Loss

18.40

13.572

13.110

% of Saving

% of Loss

50
40
30
20
10
0
Nearest

Bilinear

Bicubic

Computational Complexity of DCT by Conventional Method

(2 x 1)u (2 y 1)v for u, v = 0, 1, 2 .N-1
Cu, v (u) (v) f ( x, y) cos
cos

2N
2N

x 0 y 0
N 1 N 1

1/ N
(u )
2/ N

for u 0
for u 0

Cos function requires ----64 cos functions requires

1/ N
(v )
2/ N

4m,
256m,

For C(u, v) inside loop requires ---- 128m,

For C(u, v) outside the loop requires --Each C(u, v) requires --64 C(u, v)s requires -- (u ) (v )requires -- Totally

----

1a,
64a,
63a
2m
130m, 63a
8320m,
2d

8576m, 4096a,

for v 0
for v 0

1d
64d

4032a

66d

Computational Complexity of IDCT by Conventional Method

(2 x 1)u (2 y 1)v
f x, y (u) (v)C (u, v) cos
cos

for
2
N

2N
u 0 v 0
N 1 N 1

1/ N
(u )
2/ N

for u 0
for u 0

1/ N
(v )
2/ N

Cos function requires ----- 4m,

64 cos functions requires
Each f (x, y) requires -- 64 f (x, y) requires -- (u ) (v )requires -- Totally
----

x, y = 0, 1, 2 .N-1

1a,
1d
256m, 64a,

256m,
16384m,
2d
16640m,

for v 0
for v 0

64d

63a
4032a
4096a,

66d

Winograds Matrix Multiplication

A and B are the two N X N matrices, C is the product of A X B.
Pairwise row product in ith row

pairwise col product in jth col

N /2

Ai ai , 2 K 1.ai , 2 K

N /2

B j b2 K 1, j .b2 K , j

K 1

K 1

Compute Once, Use many times

n/ 2

ith row jth col element of C: Ci , j (ai , 2 K 1 b2 K , j )(ai , 2 K b2 K 1, j ) Ai B j

K 1

Requires
Requires

1 3
N N 2 Multiplications
2

3
( ) N 3 2 N 2 2 N additions and subtractions
2

8 X 8 Matrix Multiplication by Winograds

N /2

Ai ai , 2 K 1.ai , 2 K
K 1

N /2

B j b2 K 1, j .b2 K , j
K 1

N /2

Ci , j (ai , 2 K 1 b2 K , j )(ai , 2 K b2 K 1, j ) Ai B j
K 1

Each pairwise row product requires 4m, 3a

For 8 rows
32m, 24a
Each pairwise col product requires
For 8 cols

4m, 3a
32m, 24a

Cij requires
64 elements requires

4m, 11a, 2s
256m, 704a, 128s

Totally

320m, 752a, 128s

Computational Complexity of DCT by Winograds Method

(2 x 1)u (2 y 1)v
Cu, v (u) (v) f ( x, y) cos
cos

2N
2N

x 0 y 0
N 1 N 1

C u, v (u ) (v) * (c1 * f ( x, y ) * c1 )
T

c1 requires
c1 * f(x, y) requires
T

[c1 * f ( x, y )] * c1 requires

256m, 64a, 64d

320m, 752a, 128s
(320-32)m, (752-24)a, 128s

288

728

(u ) (v) *[c1 * f ( x, y ) * c1T ] requires 128m

(u ) (v ) requires 2d
Totally

992m, 1544a, 66d,

256s

Computational Complexity of IDCT by Winograds Method

(2 x 1)u
(2 y 1)v
f x, y (u ) (v)C (u, v) cos
cos

2
N
2
N

u 0 v 0
N 1 N 1

f f ( x, y) c1 * ( (u) (v) * c(u, v)) * c1 )

T

C1 requires
(u ) (v ) requires
(u ) (v) * c(u, v) requires
[ (u) (v) * c(u, v)]* c1 requires

256m,
2d
128m

64a,

64d

320m, 752a, 128s

T
c1 *[ (u ) (v) * c(u, v) * c1 ] requires (320-32)m, (752-24)a, 128s

288
Totally

728

992m, 1544a, 66d,

256s

Conventional Vs Winograd
Conventional
DCT
8576 multiplications
4096 additions
66 divisions
IDCT

16640 multiplications
4096 additions
66 divisions

Winograd
DCT
992multiplications
1544 additions
66 divisions
256 subtractions
IDCT
992multiplications
1544 additions
66 divisions
256 subtractions

Computational Complexity of DCT and IDCT by MATLAB

dct2()

idct2()

Experimental Results
Image : Bolivia

Size: 1015 X 877

Image : Monolake

Size: 1040 X 1040

Image

MATLAB

Conventional

Winograds

Image

MATLAB

Conventional

Winograds

Band1

1.9374

0.767

0.209

Band1

2.2344

1.0039

0.2369

Band2

1.7344

0.672

0.200

Band2

2.1719

0.9089

0.2180

Band3

1.6719

0.612

0.141

Band3

2.1250

0.8909

0.2180

Band4

1.7031

0.617

0.200

Band4

2.0938

0.8738

0.1548

Band5

1.5

0.609

0.140

Band5

2.1250

0.9139

0.2182

Band6

1.7813

0.673

0.202

Band6

2.2656

1.0045

0.2370

Band7

1.6875

0.616

0.144

Band7

2.1094

0.8739

0.1850

Total

12.0156

4.566

1.236

Total

15.1251

6.4698

1.4679

14

12.0156

12
10
8
6

4.566

4
1.236

2
0
MATLAB

Conventional

Winograds

16
14
12
10
8
6
4
2
0

15.1251

6.4698
1.4679

MATLAB

Conventional

Winograds

Experimental Results
Image : Owensvalley

Size: 1047 X 917

Image

MATLAB

Conventional

Winograds

Band1

1.9219

0.9069

0.2630

Band2

1.8906

0.8589

0.2627

Band3

1.7813

0.8162

0.2330

Band4

1.9375

0.9071

0.2632

Band5

1.8438

0.8269

0.2423

Band6

1.7656

0.8140

0.2326

Band7

1.8281

0.8264

0.2410

Total

12.9688

5.9564

1.7378

14

12.9688

12
10
8

5.9564

6
4

1.7378

2
0
MATLAB

Conventional

Winograds

Image : Santacruz

Size: 1000 X 1000

Image

MATLAB

Conventional

Winograds

Band1

2.0625

0.8609

0.2520

Band2

1.9844

0.8119

0.2360

Band3

1.9063

0.7980

0.2190

Band4

1.8594

0.7960

0.2176

Band5

1.8125

0.7826

0.2150

Band6

0.8449

0.2518

Band7

1.8906

0.7970

0.2187

Total

13.5157

5.6913

1.6101

16
14
12
10
8
6
4
2
0

13.5157

5.6913
1.6101

MATLAB

Conventional

Winograds

Experimental Results
Image : Saltlake

Size: 1035 X 920

Image

MATLAB

Conventional

Winograds

Band1

1.9219

0.8429

0.3150

Band2

1.7188

0.7140

0.2350

Band3

1.8225

0.7149

0.2660

Band4

1.8438

0.7450

0.2740

Band5

1.7813

0.7148

0.2360

Band6

1.8594

0.7690

0.3130

Band7

1.8281

0.7244

0.2661

Total

12.7758

5.225

1.9051

14

12.7758

12
10
8
5.225

6
4

1.9051

2
0
MATLAB

Conventional

Winograds

Experimental Results
Image : Bolivia

Size: 1015 X 877

Image : Monolake

Size: 1040 X 1040

Image

MATLAB

Conventional

Winograds

Image

MATLAB

Conventional

Winograds

Band1

2.5469

1.3412

0.2240

Band1

3.0781

1.5740

0.2419

Band2

2.4063

1.2179

0.185

Band2

3.1406

1.5743

0.249

Band3

2.4688

1.2999

0.1940

Band3

3.0469

1.5480

0.2176

Band4

2.5469

1.3419

0.2268

Band4

3.0469

1.4670

0.2180

Band5

2.3906

1.2029

0.1850

Band5

3.0313

1.4630

0.2177

Band6

2.4844

1.3409

0.2011

Band6

3.0781

1.5700

0.2460

Band7

2.5469

1.3429

0.2270

Band7

3.1094

1.5741

0.2482

Total

17.3908

9.0876

1.4429

Total

21.5313

10.7704

1.6384

20

17.3908

25

21.5313

20

15

15

9.0876

10

10.7704

10

1.4429

1.6384

MATLAB

Conventional

Winograds

MATLAB

Conventional

Winograds

Experimental Results
Image : Owensvalley

Size: 1047 X 917

Image : Santacruz

Size: 1000 X 1000

Image

MATLAB

Conventional

Winograds

Image

MATLAB

Conventional

Winograds

Band1

2.8438

1.2510

0.2440

Band1

2.9844

1.4220

0.2630

Band2

2.8594

1.3290

0.2440

Band2

2.7500

1.1504

0.2110

Band3

2.8438

1.3267

0.2300

Band3

2.8750

1.4040

0.2510

Band4

2.8906

1.5025

0.2473

Band4

2.8281

1.1770

0.2010

Band5

2.7969

1.3220

0.2200

Band5

2.8438

1.1870

0.2190

Band6

2.8906

1.5030

0.2470

Band6

2.7500

1.1642

0.2109

Band7

2.8281

1.2487

0.2240

Band7

2.8281

1.1769

0.2200

Total

19.9532

9.4829

1.6563

Total

19.8594

8.6815

1.5759

25

25

19.9532
20

20

15

15

9.4829

10

19.8594

8.6815

10

1.6563

1.5759

MATLAB

Conventional

Winograds

MATLAB

Conventional

Winograds

Experimental Results
Image : Saltlake

Size: 1035 X 920

Image

MATLAB

Conventional

Winograds

Band1

2.6875

1.2720

0.2020

Band2

2.7344

1.3059

0.2190

Band3

2.7969

1.4250

0.2220

Band4

2.6563

1.2670

0.2020

Band5

2.6719

1.3900

0.2020

Band6

2.7031

1.2840

0.2182

Band7

2.5625

1.2360

0.1860

Total

18.8126

9.1799

1.4512

20

18.8126

15
9.1799

10
5

1.4512

0
MATLAB

Conventional

Winograds

Winograds Speed Up
DCT
Speed up of around 8 when compared with MATLAB
Speed up of around 3.5 when compared with conventional
method
IDCT
Speed up of around 12 when compared with MATLAB
Speed up of around 6 when compared with conventional method

Image Classification: The process of categorizing all pixels in an

image into land cover classes.
Classification

Supervised Classification

Parametric

Maximum Likelihood Classifier

Mahalanobis distance classifier
Euclidean distance classifier

Unsupervised Classification

Non Parametric

Nearest Neighborhood Classifier

K-Nearest Neighborhood Classifier

Training Sites

1.
2.
3.
4.

Barren Land
Shrub
Water Body
Forest

Maximum Likelihood Classifier

1. Take the instance X to be classified
2. Calculate the probability of X belonging to each of a predefined
set of k classes by using

pk ( X )

(2 )

Where

n/2

1/ 2

exp[1 / 2.( X M k )

1
k

( X M k )].

Pk (x)= likelihood of x belonging to class K

n = number of bands
x= A d-dimensional pixel vector
MK = mean vector of class k
| k |= determinant of covariance matrix k

1
k

= Inverse of covariance matrix k

3. Then it assigns the instance X to the class for which the

probability is highest

Mahalanobis Distance Classifier

1. Take the instance X to be classified
2. Find Mahalanobis distance between X and the mean vector
of each class by using
dk (X) = (X-Mk)T -1(X-Mk)
X = a d-dimensional pixel vector

dk (x) = mahalanobis distance of X from kth class mean

MK = mean vector of class k

= inverse of the pooled covariance matrix

3. Then it assigns the instance X to the class having minimum

mahalanobis distance

Euclidean Distance Classifier

1. Take the instance X to be classified
2. Find Euclidean distance between X and the mean vector
of each class by using
dk (X) = (X-Mk)T (X-Mk)
X = A d-dimensional pixel vector
dk (x) = Euclidean distance of X from kth class mean
MK = mean vector of class k
3. Then it assigns the instance X to the class having minimum
Euclidean distance

Nearest Neighborhood Classifier

1. Take the instance x to be classified

2. Find the nearest neighbor of x in the training data
3. Determine the class k of the nearest neighbor
4. Return the class k as the classification of x

K-Nearest Neighborhood Classifier

1. Take the instance x to be classified

2. Find the K nearest neighbors of x in the training data
3. Determine the class k of the majority of the instances among

the k nearest neighbor

4. Return the class k as the classification of x

Accuracy Assessment

Original Image

Reconstructed
Image

Classification
Result

Classification
Results

Classification Accuracy
Assessment

Accuracy Assessment

Overall Accuracy

Confusion Matrix
Kappa Coefficient

Confusion Matrix
Multi-level
Confusion Matrix

Predicted Class
Class 1

Class 2

-----

Class N

Marginal Sum of
Actual Values
N

Class 1

f11

f12

-----

f1N

f
j 1

1j

Class 2

f22

-----

f2N

-----

2j

------

-----

------

------

j 1

------

Actual
Class

f21

Class N

fN1

fN2

-----

fNN

f
j 1

Marginal Sum of
Predictions

f
i 1

i1

f
i 1

i2

-----

i 1

f iN

Nj

T f ij
i 1 j 1

Total

Overall Accuracy =

no of

correctly clossified pixels (diognal

Total no of test pixels

f11 f 22 ...... f NN
N

f
i 1 j 1

ij

elements )

Kappa Coefficient

P0 Pe
K
1 Pe
P0 is observed agreement
N

Total number of correctly classified pixels

Total number of test pixels
N

f X
i1

i 1

j 1

f
Pe is the Expected agreement =

i 1 j 1

ij

1j

f
i 1

i2

X 2j
j 1

f
i 1 j 1

.......... ...

ij

f
i 1

iN

X
j 1

f
i 1 j 1

ij

f
i 1 j 1

ij

Interpretation of Kappa

Agreement
Kappa

Poor Slight Fair Moderate

0.0 .20 .40
.60

Substantial
.80

Almost perfect
1.0

Nj

Image : Monolake

Classification : Maximumlikilihood
Original Image

Classified as group
Correct Classification
(%)

Number of
Samples used

1. barren land

100

500

2. Vegetation

91

3. Forest
4. Rock

Spectral Class

500

500

455

42

99.8

500

499

99.8

500

499

Misclassification= 2.35%

Overall accuracy= 97.65%

Kappa coefficient=0.9686

Standard JPEG Compression image

Classified as group

Spectral Class

Correct Classification
(%)

Number of
Samples used

1. barren land

99.8

500

499

2. Vegetation

89

500

445

53

3. Forest

99

500

495

4. Rock

100

500

500

Misclassification= 3.05%

Overall accuracy= 96.95%

Kappa coefficient=0.9593

Classification Accuracy (Maximumlikilihood)

Image : Monolake

Classification Accuracy (Org): 97.8 Classification accuracy (JPEG): 97.6

Compression

% of Saving
Bits

Classification
Accuracy

Compression with 2048 Clusters

41.114

96.2

Compression with 4096 Clusters

39.03

96.7

Compression with 8192 Clusters

16.167

97.00

Mean Filter

18.239

96.9

Outlier Mean Filter (C=1.645)

15.463

97.1

Outlier Mean Filter (C=1.28)

11.951

97.15

Median Filter

13.125

97.10

Outlier Median Filter (C=1.645)

10.366

97.5

Outlier Median Filter (C=1.28)

7.034

97.5

Bicubic

34.43

96.9

ste
rs

ed
ian

cu
bi

(1
.28
)
Bi

ed
ian

(1
.64
5)

M
ed
ian
M

ea
n

ea
n

(1.
64
5)
M
ea
n(
1.
28
)

40
96
Clu
ste
rs
81
92
Clu
ste
rs

20
48
Clu

120
100
80
60
40
20
0

Classification Accuracy (Mahalanobis)

Image : Monolake

Classification Accuracy (Org): 95.3

Classification accuracy (JPEG): 94.9

Compression

% of Saving
Bits

Classification
Accuracy

Compression with 2048 Clusters

41.114

92.05

Compression with 4096 Clusters

39.03

93.35

Compression with 8192 Clusters

16.167

93.65

Mean Filter

18.239

93.40

Outlier Mean Filter (C=1.645)

15.463

93.8

Outlier Mean Filter (C=1.28)

11.951

94.05

Median Filter

13.125

93.85

Outlier Median Filter (C=1.645)

10.366

94.30

Outlier Median Filter (C=1.28)

7.034

94.35

Bicubic

34.43

93.85

ed
ia
n

cu
bi

(1
.2
8)
Bi

(1
.6
45
)

M
n
ed
ia

ed
ia
M

ea
n

ea
n

te
rs

te
rs

te
rs

Clu
s

Clu
s

Clu
s

(1
.6
45
)
M
ea
n
(1
.2
8)

81
92

40
96

20
48

100
80
60
40
20
0

Classification Accuracy (Minimum Distance)

Image : Monolake

Classification Accuracy (Org): 80.0 Classification accuracy (JPEG): 79.1

Compression

% of Saving
Bits

Classification
Accuracy

Compression with 2048 Clusters

41.114

77.7

Compression with 4096 Clusters

39.03

78.1

Compression with 8192 Clusters

16.167

78.3

Mean Filter

18.239

78.15

Outlier Mean Filter (C=1.645)

15.463

78.4

Outlier Mean Filter (C=1.28)

11.951

78.5

Median Filter

13.125

78.4

Outlier Median Filter (C=1.645)

10.366

78.5

Outlier Median Filter (C=1.28)

7.034

78.9

Bicubic

34.43

78.1

ed
ian

cu
bi

(1
.28
)
Bi

ed
ian

(1
.64
5)

M
ed
ian
M

ea
n

ea
n

(1.
64
5)
M
ea
n(
1.
28
)

20
48
Clu
ste
rs
40
96
Clu
ste
rs
81
92
Clu
ste
rs

90
80
70
60
50
40
30
20
10
0

Classification Accuracy (Nearest Neighborhood)

Image : Monolake

Classification Accuracy (Org): 98.2

Classification accuracy (JPEG): 97.7

Compression

% of Saving
Bits

Classification
Accuracy

Compression with 2048 Clusters

41.114

95.35

Compression with 4096 Clusters

39.03

95.5

Compression with 8192 Clusters

16.167

96.25

Mean Filter

18.239

96.10

Outlier Mean Filter (C=1.645)

15.463

96.3

Outlier Mean Filter (C=1.28)

11.951

96.9

Median Filter

13.125

96.7

Outlier Median Filter (C=1.645)

10.366

96.95

Outlier Median Filter (C=1.28)

7.034

97.25

Bicubic

34.43

95.75

ste
rs

ed
ia n

cu
bi

(1
.28
)
Bi

ed
ian

(1
.64
5)

M
ed
ian
M

ea
n

ea
n

(1.
64
5)
M
ea
n(
1.
28
)

ste
rs
81
92
Clu
ste
rs

40
96
Clu

20
48
Clu

120
100
80
60
40
20
0

Classification Accuracy (K-Nearest Neighborhood)

Image : Monolake

Classification Accuracy (Org): 97.8 Classification accuracy (JPEG): 96.75

Compression

% of Saving
Bits

Classification
Accuracy

Compression with 2048 Clusters

41.114

95.75

Compression with 4096 Clusters

39.03

95.8

Compression with 8192 Clusters

16.167

96.5

Mean Filter

18.239

96.3

Outlier Mean Filter (C=1.645)

15.463

96.5

Outlier Mean Filter (C=1.28)

11.951

96.6

Median Filter

13.125

96.5

Outlier Median Filter (C=1.645)

10.366

96.7

Outlier Median Filter (C=1.28)

7.034

96.7

Bicubic

34.43

95.8

ed
ian

cu
bi

(1
.28
)
Bi

ed
ian

(1
.64
5)

M
ed
ian
M

ea
n

ea
n

(1.
64
5)
M
ea
n(
1.
28
)

ste
rs
40
96
Clu
ste
rs
81
92
Clu
ste
rs

20
48
Clu

120
100
80
60
40
20
0

CONCLUSIONS

The observations corroborate that Classification performance

of original images, conventional compressed images, and

proposed compressed images is almost the same. This supports

our hypothesis of sending compressed data from on-board satellite
system is no way hampering classification accuracy, so, mission
planners as well communicate compressed data to ground stations.
This saves bandwidth, space, power requirements and increases

satellite life.

List of Publications
Ch. Ramesh, Dr. N.B. Venkateswarlu and Dr. J.V.R. Murthy, Fast DCT Algorithm using
Winograds Method, International Journal of Electronics and Communication Engineering &
Technology, vol.3, issue.1, pp.98-110, January-June 2012, Online ISSN: 0976-6472, Print
ISSN: 0976-6464, Impact factor: 0.85.
Ch. Ramesh, Dr. N.B. Venkateswarlu and Dr. J.V.R. Murthy, A Novel K-Means Based JPEG
Algorithm for Still Image Compression, International Journal of Computer Engineering &
Technology, vol.3, issue.1, pp.339-354, January-June 2012, Online ISSN: 0976-6375, Print
ISSN 0976-6367,Impact factor: 1.0425.
Ch. Ramesh, Dr. N.B. Venkateswarlu and Dr. J.V.R. Murthy, Filter Augmented JPEG
Algorithms: A Critical Performance Study for Improving Bandwidth, International Journal of
Computer Applications, vol.60, issue.17, December 2012, Online ISSN: 0975-8887, Impact
factor: 0.821.
Ch. Ramesh, Dr. N.B. Venkateswarlu and Dr. J.V.R. Murthy, A New Classification
Performance aware Multi sensor, Multi resolution Satellite image compression Technique,
Global Journal of Computer Science and Technology, vol.13, issue.7, version.1.0, August
2013, Online ISSN: 0975-4172, Print ISSN: 0975-4350.
Ch. Ramesh, Dr. N.B. Venkateswarlu and Dr. J.V.R. Murthy, A Critical Performance
Evaluation of Classification Methods with Modified JPEG Decompressed Multiband Images,
Global Journal of Researchers in Engineering, vol.13, issue.16, version.1.0, December 2013.
Online ISSN: 2249-4596, Print ISSN: 0975-5861

Thank You