Skip to main content

Table 4 Comparison of the PSNR (in dB) for inpainting of images with various corruption levels based on analytic dictionaries, DCT, MCA, and dictionaries learned on all available corrupted image patches, BPFA, BPFAomp, wKSVD, and ITKrMM (modified inpainting is marked with a *)

From: Fast dictionary learning from incomplete data

 

Algorithm

Bar.

Cam.

Hou.

Man.

Pepp.

Pir.

30% corruption

Noisy Im.

11.17

10.81

10.11

10.82

11.18

11.70

 

DCT64

37.49

32.66

41.89

30.60

39.12

35.40

 

DCT144

37.08

32.41

41.49

30.86

38.90

35.42

 

MCA

35.89

32.45

39.62

28.38

35.59

33.35

 

BPFA

34.76

32.08

39.76

29.58

37.92

34.38

 

BPFAomp

35.36

32.23

41.09

30.81

38.66

35.42

 

wKSVD

35.87

32.62

41.42

30.41

38.64

35.09

 

ITKrMM1

36.12

32.80

41.97

30.85

39.20

35.60

 

ITKrMM3

37.16

3 3 . 0 4

4 2 . 3 0

3 0 . 9 2

3 9 . 8 0

3 6 . 0 8

50% corruption

Noisy Im.

8.95

8.59

7.88

8.60

8.96

9.47

 

DCT64

32.72

28.56

36.65

26.99

34.01

31.10

 

DCT144

32.46

28.46

36.40

27.25

33.93

31.17

 

MCA

32.50

28.99

36.54

25.34

32.35

29.86

 

BPFA

32.97

28.89

37.71

27.25

35.29

31.89

 

BPFAomp

32.98

28.87

37.88

27.29

35.41

32.18

 

wKSVD

33.23

2 9 . 5 5

3 8 . 2 1

27.79

3 5 . 4 1

32.12

 

ITKrMM1

33.28

29.44

37.75

27.96

35.31

32.14

 

ITKrMM3

3 3 . 8 2

29.48

38.04

2 7 . 9 7

35.30

3 2 . 2 6

70% corruption

Noisy Im.

7.48

7.13

6.42

7.13

7.50

8.01

 

DCT64

28.21

24.86

31.49

24.29

29.05

27.21

 

DCT144

28.09

24.81

31.37

24.44

28.81

27.32

 

MCA

28.74

25.71

33.42

23.29

28.56

26.55

 

BPFA

29.40

25.74

33.56

24.93

3 1 . 4 3

28.77

 

BPFAomp

29.22

25.61

33.05

25.10

31.12

28.63

 

BPFAomp*

29.23

25.60

33.04

25.11

31.18

28.74

 

wKSVD

29.70

25.89

33.96

25.09

31.17

28.76

 

wKSVD*

29.74

26.02

3 4 . 0 9

25.09

31.32

2 8 . 8 4

 

ITKrMM1

29.48

25.84

33.26

25.11

29.64

28.53

 

ITKrMM1*

2 9 . 9 3

2 6 . 3 4

33.65

2 5 . 1 2

31.26

28.83

  1. The results are averaged over 5 random masks and initialisations. The best result for each setting is marked in bold