You are on page 1of 13

Speckle Interferometry (Bispectrum)

Short exposure image contain high frequency information that can be


exploited to reconstruct distorted image

Images can be described as the convolution of the image of the true


object () with a blurring function ()

Note the blurring function changes with each image


Speckle Interferometry

The blurring transfer function can be expressed in the Fourier domain

We need to estimate the magnitude and phase of for each image


Basic Speckle Imaging Algorithm

Ref: Carmen J. Carrano, "Speckle imaging over horizontal paths," Proc. SPIE 4825, High-Resolution Wavefront Control: Methods, Devices, and Applications IV,
(6 November 2002);
Application to Images
The transfer function varies across the image
(anisoplanatism)
Solution:
1. Divide the image into overlapping regions
2. Estimate object for each region
(Formula used for Amplitude)
3. Recombine the regions using weighted
averages (Hanning filter)
GPU Specifications

Quadro RTX 5000

CUDA Parallel Processing Cores 3072

NVIDIA Tensor Cores 384

NVIDIA RT Cores 48

GPU Memory 16 GB GDDR6

FP32 Performance 11.2 TFLOPS

Max Power Consumption 265 W


Optimization of SI Algorithm using CUDA
Streams
- Figure shows overlapping of CPU to GPU data transfer(Gold) with execution
of algorithm functions on GPU.
- Streams take advantage of overlapping by transferring data in chunks
- A chunk of data is processed while a different chunk is being transferred to
GPU on a different stream
7 Image Quality Metrics
Image Quality Metrics are used to quantify the quality of images
Image quality assessment metrics fall largely into three categories
 FR – full reference
 RR – reduced reference
 NR – no reference

 An ideal reference image is typically not available for


atmospherically distorted videos.
8 No-reference Image Quality Metrics
There are a number of NR IQA metrics. Most of these are trained
using human rated data (opinion aware)
DIVINE, BLIINDS, CBIQ, LBIQ

I chose to start with two spatial domain algorithm


 BRISQUE - spatial domain
 NIQE (not OA) - spatial domain

These use Natural Scene Statistics (NSS) and the distance from a
multi-variate Gaussian distribution to return a metric
9 Full Reference IQA
When a reference is available, there are a number of Quality
metrics

 imse – image means squared error


 SSIM – structural similarity index
 PSNR – peak signal to noise ratio

 The median image will be used as the reference for the


background. It largely removes moving objects.
References

 C. J. Carrano, "Speckle imaging over horizontal paths " Proc. SPIE 4825, High-Resolution
Wavefront Control: Methods, Devices, and Applications IV, 2002.
 D. Korff, "Analysis of a method for obtaining near-diffraction-limited information in the
presence of atmospheric turbulence," J. Opt. Soc. Am, vol. 63, pp. 971-980, 1973.
 A. Lohmann, G. Weigelt, and B. Wirnitzer, "Speckle Masking in Astronomy: Triple
Correlation Theory and Applications," Applied Optics, vol. 22, No. 4, Dec. 15, 1983.
 Taylor W. Lawrence, J. Patrick Fitch, Dennis M. Goodman, Norbert A. Massie, Robert J.
Sherwood, and Erik M. Johansson "Extended-image reconstruction through horizontal path
turbulence using bispectral speckle interferometry," Optical Engineering 31(3), (1 March
1992)
11 References

 Mittal, A. K. Moorthy and A. C. Bovik, "No-Reference Image Quality


Assessment in the Spatial Domain," in IEEE Transactions on Image Processing,
vol. 21, no. 12, pp. 4695-4708, Dec. 2012, doi: 10.1109/TIP.2012.2214050.

 A. Mittal, R. Soundararajan and A. C. Bovik, "Making a “Completely Blind”


Image Quality Analyzer," in IEEE Signal Processing Letters, vol. 20, no. 3, pp.
209-212, March 2013, doi: 10.1109/LSP.2012.2227726.
Speckle Interferometry: MTF

MTF is found by trying several values in the eqn

The magnitude of the object is estimated as


Speckle Interferometry: Phase estimate

The phase of the object can be estimated by recursively using the equation

Where the bispectrum is defined as

The phase at the centre of the spectrum is set to 0 and recursively computed

Note: The bispectrum is 4 dimension but can be sped up using symmetry

You might also like