Solmaz Hajmohammadi, Saeid Nooshabadi, and Jeremy P. Bos, "Massive parallel processing of image reconstruction from bispectrum through turbulence," Appl. Opt. 54, 9370-9378 (2015)

This paper presents a massively parallel method for the phase reconstruction of an object from its bispectrum phase. Our aim is to recover an enhanced version of a turbulence-corrupted image by developing an efficient and fast parallel image-restoration algorithm. The proposed massively parallel bispectrum algorithm relies on multiple block parallelization. Further, in each block, we employ wavefront processing through strength reduction to parallelize an iterative algorithm. Results are presented and compared with the existing iterative bispectrum method. We report a speed-up factor of 85.94 with respect to sequential implementation of the same algorithm for an image size of $1024\times 1024$.

Byungjae Hwang, Taeseong Woo, and Jung-Hoon Park Opt. Lett. 44(24) 5985-5988 (2019)

References

You do not have subscription access to this journal. Citation lists with outbound citation links are available to subscribers only. You may subscribe either as an OSA member, or as an authorized user of your institution.

You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an OSA member, or as an authorized user of your institution.

You do not have subscription access to this journal. Figure files are available to subscribers only. You may subscribe either as an OSA member, or as an authorized user of your institution.

You do not have subscription access to this journal. Article tables are available to subscribers only. You may subscribe either as an OSA member, or as an authorized user of your institution.

You do not have subscription access to this journal. Equations are available to subscribers only. You may subscribe either as an OSA member, or as an authorized user of your institution.

You do not have subscription access to this journal. Article level metrics are available to subscribers only. You may subscribe either as an OSA member, or as an authorized user of your institution.

Execution Times (in Seconds) and Approximate Number of Complex Operations for the Recursive Reconstruction of the 50 Frames of $\mathsf{256}\times \mathsf{256}$ Image Size^{a}

The experimentation platform is MATLAB, R2013b, 64-bit, on a six-core Intel Xeon CPU, X5650 2.67 GHz.
Table does not include the number of integer operations to compute 229,376,000 the indices into the 2D structure required for the bispectrum calculation.

Table 2.

Number of Complex Operations for Various Images Sizes for Complete Decomposition

INB-DP/INB-SP, interior blocks parallelized in double/single precision; INSDB-DP/INSDB-SP, interior and side blocks parallelized in double/single precision.

Table 4.

Computation Times (in Seconds) for Block Size of $\mathsf{16}\times \mathsf{16}$ and Image Size of $\mathsf{256}\times \mathsf{256}$

INB-DP/INB-SP, interior blocks parallelized in double/single precision; INSDB-DP/INSDB-SP, interior and side blocks parallelized in double/single precision.

Table 5.

Execution Times (in Seconds) for Different Image Sizes

Parallel Algorithm

Image Size

Recursive Algorithm

Mem-Tx

Compute

Total

Speed-up

$256\times 256$

69.00

0.00

2.61

2.61

26.43

$512\times 512$

319.98

0.00

4.68

4.68

68.37

$768\times 768$

695.00

4.82

6.86

11.68

59.50

$1024\times 1024$

1360.50

5.90

9.93

15.83

85.94

Table 6.

Execution Times (in Seconds) for Different Numbers of Subplanes for the Image Size of $\mathsf{256}\times \mathsf{256}$

Number of subplanes

5

6

7

8

9

10

11

12

13

14

15

16

Timing

2.61

2.87

3.14

3.4

3.83

4.26

4.75

5.25

5.83

6.46

7.17

7.81

Tables (6)

Table 1.

Execution Times (in Seconds) and Approximate Number of Complex Operations for the Recursive Reconstruction of the 50 Frames of $\mathsf{256}\times \mathsf{256}$ Image Size^{a}

The experimentation platform is MATLAB, R2013b, 64-bit, on a six-core Intel Xeon CPU, X5650 2.67 GHz.
Table does not include the number of integer operations to compute 229,376,000 the indices into the 2D structure required for the bispectrum calculation.

Table 2.

Number of Complex Operations for Various Images Sizes for Complete Decomposition

INB-DP/INB-SP, interior blocks parallelized in double/single precision; INSDB-DP/INSDB-SP, interior and side blocks parallelized in double/single precision.

Table 4.

Computation Times (in Seconds) for Block Size of $\mathsf{16}\times \mathsf{16}$ and Image Size of $\mathsf{256}\times \mathsf{256}$

INB-DP/INB-SP, interior blocks parallelized in double/single precision; INSDB-DP/INSDB-SP, interior and side blocks parallelized in double/single precision.

Table 5.

Execution Times (in Seconds) for Different Image Sizes

Parallel Algorithm

Image Size

Recursive Algorithm

Mem-Tx

Compute

Total

Speed-up

$256\times 256$

69.00

0.00

2.61

2.61

26.43

$512\times 512$

319.98

0.00

4.68

4.68

68.37

$768\times 768$

695.00

4.82

6.86

11.68

59.50

$1024\times 1024$

1360.50

5.90

9.93

15.83

85.94

Table 6.

Execution Times (in Seconds) for Different Numbers of Subplanes for the Image Size of $\mathsf{256}\times \mathsf{256}$