Skip to main content

Table 1 Quantitative comparison under the CDI modality

From: Large-scale phase retrieval

Algorithm

SNR = 20dB

SNR = 25dB

SNR = 30dB

PSNR

SSIM

TIME

PSNR

SSIM

TIME

PSNR

SSIM

TIME

AP

18.46

0.50

819.67

21.75

0.58

854.37

22.29

0.65

863.14

WF

19.05

0.52

+ 27.15

20.84

0.62

+ 31.98

21.27

0.70

+ 32.41

RWF

18.52

0.50

+ 25.69

21.98

0.61

+ 27.53

22.41

0.71

+ 27.98

AF

16.55

0.42

+ 28.61

19.63

0.49

+ 29.74

19.83

0.54

+ 27.29

TAF

18.57

0.53

+ 26.04

21.81

0.59

+ 25.99

22.30

0.65

+ 26.49

RAF

18.52

0.53

+ 22.55

21.79

0.58

+ 21.80

22.27

0.65

+ 22.19

PLIFT

✘-memory limitation

✘-memory limitation

✘-memory limitation

PLAMP

✘-memory limitation

✘-memory limitation

✘-memory limitation

PMAX

16.64

0.42

+ 38.48

19.73

0.49

+ 39.04

19.97

0.54

+ 38.11

CD

✘-no convergence

✘-no convergence

✘-no convergence

KAC

✘-no convergence

✘-no convergence

✘-no convergence

prDeep

20.60

0.52

+ 49.01

21.83

0.58

+ 43.36

23.33

0.65

+ 35.46

LPR

23.30

0.79

+28.52

25.52

0.83

+ 29.97

28.11

0.86

+ 27.19

  1. CD and KAC fail with no convergence. PLIFT and PLAMP are out of computer memory. Most of the conventional algorithms produce little improvement than AP. LPR outperforms the other algorithms, with as much as 6dB (SNR = 30) and 0.29 (SNR = 20) improvement on PSNR and SSIM, respectively. We use the excess time beyond AP as the other algorithms’ running time, which shows that prDeep consumes the most running time. In comparison, LPR takes the same level of running time as the conventional methods