Source-linked AI summary
Compressive adaptive computational ghost imaging
Marc Aßmann, Manfred Bayer
TL;DR
High-resolution compressive imaging was limited by the long measurement durations required. The paper develops adaptive CCGI that measures directly in a sparse basis, reporting acquisition rates 15–20 times below the Shannon limit and better signal-to-noise ratio than RMCS.
Problem
High-resolution imaging with the discussed compressive techniques had, to the authors' knowledge, not been carried out because the required measurement durations were long.
Source: Experimental results · p. 9Method
Adaptive CCGI records images directly in a sparse basis and uses thresholds to take an image in a fixed amount of time.
Source: DISCUSSION · p. 10, Experimental results · p. 10Results
Images were recorded with a single-pixel camera at an acquisition rate 15–20 times below the Shannon limit, and CCGI-based methods tend to achieve better signal-to-noise ratio than RMCS.
Source: DISCUSSION · pp. 10–11Takeaways & Limitations
The paper presents recording directly in a sparse basis as a way to reach the Shannon limit.
Source: DISCUSSION · p. 10Unlike RMCS, CCGI requires a predefined sparse basis; for very sparse images, this can impose a penalty of up to 50%.
Source: DISCUSSION · p. 11
Abstract
from arXiv · showhide
Compressive sensing is considered a huge breakthrough in signal acquisition. It allows recording an image consisting of $N^2$ pixels using much fewer than $N^2$ measurements if it can be transformed to a basis where most pixels take on negligibly small values. Standard compressive sensing techniques suffer from the computational overhead needed to reconstruct an image with typical computation times between hours and days and are thus not optimal for applications in physics and spectroscopy. We demonstrate an adaptive compressive sampling technique that performs measurements directly in a sparse basis. It needs much fewer than $N^2$ measurements without any computational overhead, so the result is available instantly.
The adaptive compressive CGI algorithm
The adaptive CCGI scheme measures selectively in a sparse wavelet basis, using coefficient thresholds to refine detail across scales. Coarse-scale measurements are retained where finer-scale edge detail is not indicated.
The adaptive compressive CGI algorithm
- Adaptive CCGI uses measurements in a sparse basis and adaptive sampling to avoid computational overhead after measurements are complete.The approach replaces random speckle patterns with patterns forming the sparse basis.Source: The adaptive compressive CGI algorithm · p. 3
- The 2D Haar transform sums and differences adjacent pixels across rows and columns, producing four quadrants with a coarse image in the upper left and edge information in the others.Further transforms of the upper-left quadrant yield coarser scales; the upper-left pixel ultimately contains the image's mean intensity.Source: The adaptive compressive CGI algorithm · p. 3
- At each scale, the method checks wavelet coefficients against a threshold and measures the four children at the next finer scale when a coefficient exceeds it.It repeats this refinement with a new threshold until the finest image at scale j=1 is constructed.Source: The adaptive compressive CGI algorithm · p. 5
- No wavelet coefficient requires more than four real-space measurements because each coefficient spans exactly four pixels at its scale.A parent coefficient covers the same image region as its four children at the next finer scale.Source: The adaptive compressive CGI algorithm · pp. 4–5
- The finer image combines newly measured transmission values in sharp-edge regions with previously measured coarse-scale values in regions without edges.The image is transformed again at each finer scale to identify where further measurements are needed.Source: The adaptive compressive CGI algorithm · p. 5
Experimental results
The metal-plate image was faithfully reconstructed with approximately 5–7% of full-resolution measurements, while a more complex resolution chart prompted the adaptive algorithm to take more measurements. Image quality improved little beyond α >7%, and periodic chart structures exposed an algorithm weakness.
Experimental results
- The adaptive procedure measures children wavelet coefficients at finer scales when parent coefficients exceed a threshold, then applies the inverse wavelet transform.In the illustrated example, measurements were roughly 40% of the image pixel count.Source: Experimental results · p. 6
- At α around 2%, the holes could already be distinguished and counted; image quality looked reasonable around 4% and good around 7%.Only small improvements were seen for larger α.Source: Experimental results · pp. 7, 9
- The mean squared error roughly decayed exponentially and saturated near 0.055, verifying little quality improvement for α >7%.The remaining error mainly came from hole edges falling between pixel borders, producing intermediate transmission unlike the reference drawing.Source: Experimental results · p. 7
- With the metal-plate threshold retained, the resolution chart prompted almost three times as many measurements, reaching α=36.9%, and its image remained good.Lines about 125 µm in size could still be resolved, though the image showed artifacts.Source: Experimental results · p. 7
- Strictly periodic parallel lines may appear to have no edge at a coarser scale, a weakness behind the chart artifacts.The text notes that more advanced algorithms considering more than the parent wavelet value may overcome this problem.Source: Experimental results · p. 7
- Thresholds let users control image quality and avoid computationally intensive recovery algorithms; the SLM can operate at up to 60 Hz.The text describes the prospect of recording reasonable-resolution, reasonable-quality images within a few minutes with readout components on a similar timescale.Source: Experimental results · pp. 7, 9
- Approximately 5–7% of full-resolution measurements sufficed for a faithful image of the metal plate.The acquisition rate α is the total measurements performed relative to the number of pixels on the finest scale.Source: Experimental results · pp. 5, 8
DISCUSSION
The discussion weighs adaptive CCGI’s fast, reconstruction-free sparse-basis imaging against RMCS, highlighting measurement, signal-to-noise, and application trade-offs.
DISCUSSION
- Images were recorded with a single-pixel camera at acquisition rates 15–20 times below the Shannon limit through direct sparse-basis sampling.Source: DISCUSSION · p. 10
- CCGI avoids computational image reconstruction while drastically reducing measurement time for high-resolution images captured with single-pixel detectors.Source: DISCUSSION · p. 11
- RMCS typically achieves near-ideal reconstruction with fewer measurements and finds a sparse basis automatically, while CCGI tends to deliver better signal-to-noise at comparable measurement counts.Source: DISCUSSION · p. 11
- For very sparse images, CCGI’s measurement penalty can reach 50%, and required counts are hard to predict because sparsity in the wavelet basis varies.Source: DISCUSSION · p. 11
- Noise comparable to the chosen threshold can become severe, making CCGI unsuitable for images with strong noise or weak signals.Source: DISCUSSION · p. 11
- Across examined image sizes, measurement time exceeded wavelet-transform and sampling-queue computation time enough that acquisition delay was not noticeable.Source: DISCUSSION · p. 11
- Immediate results make adaptive, spatially resolved pump-probe spectroscopy and other delicate high-resolution spectroscopic measurements possible.Source: DISCUSSION · p. 12
METHODS
The experiment measured object transmission at the SLM focal plane using a photodiode and multimeter, with a hole test plate and pulsed Ti:Sapph laser.
METHODS
- The objects were placed at the SLM focal plane (f=36 cm).Source: METHODS · p. 12
- A photodiode measured transmission through each object, and a Keithley 2000 multimeter recorded its output.Transmission through the object was focused onto the photodiode.Source: METHODS · p. 12
- The test plate had twelve holes: six with 2 mm diameters and six with 1 mm diameters, averaging around 1.5 mm apart.Source: METHODS · p. 12
- Transmission measurements used a pulsed Ti:Sapph laser with approximately 2 ps pulses at 800 nm.Source: METHODS · p. 12