Point Target Detection Research Papers

Compressive Sensing Hyperspectral Imaging by Spectral Multiplexing with Liquid Crystal

2022, Journal of Imaging

Hyperspectral (HS) imaging involves the sensing of a scene’s spectral properties, which are often redundant in nature. The redundancy of the information motivates our quest to implement Compressive Sensing (CS) theory for HS imaging. This... more

descriptionView Paper arrow_downwardDownload

Compressive Sensing Hyperspectral Imaging by Spectral Multiplexing with Liquid Crystal

by Vladimir Farber

2022, Journal of Imaging

Hyperspectral (HS) imaging involves the sensing of a scene’s spectral properties, which are often redundant in nature. The redundancy of the information motivates our quest to implement Compressive Sensing (CS) theory for HS imaging. This... more

descriptionView Paper arrow_downwardDownload

Compressive Sensing Hyperspectral Imaging by Spectral Multiplexing with Liquid Crystal

by Yitzhak Y August

2022, Journal of Imaging

Hyperspectral (HS) imaging involves the sensing of a scene’s spectral properties, which are often redundant in nature. The redundancy of the information motivates our quest to implement Compressive Sensing (CS) theory for HS imaging. This... more

Figure 1. Spectral multiplexing. The figure represents three different examples of spectral multiplexing. Each sub-figure illustrates multiplexing of a few spectral bands onto a FPA. Within the framework of CS, the CS-MUSI camera can reconstruct HS images with hundreds of spectral bands from only spectrally multiplexed shots, numbering an order of magnitude less than would be required using conventional systems. Furthermore, the CS-MUSI camera benefits from high optical throughput, small volume, light weight, and reduced acquisition time. In this paper, we review the evolution of the CS-MUSI camera and its different applications.

Figure 7. Staring mode reconstruction result of three LED arrays. (a) RGB color image of three LED arrays that were used as objects to be imaged with CS-MUSI. (b-e) Representative single exposure images for LC cell voltage of 0 V, 5.8373 V, 7.6301 V and 8.6552 V, respectively. (f) RGB representation of the reconstructed HS image (700 x 700 pixels x 391 bands). (g-i) Reconstructed images at 460 nm, 520 nm and 650 nm, respectively. (k—m) Spectrum reconstruction for three points in the HS datacube and comparison to the measured spectra of the three respective LEDs with a commercial grating-based spectrometer. images (HS and ultra-spectral images) attained by the CS-MUSI camera in staring mode. Figure 7 presents the results of an experiment where the emission spectra of three arrays of red, green and blue light sources (Thorlabs LIU001, LIU002 and LIU003 LED arrays) were imaged using the CS-MUSI camera. Figure 7a shows the image of the light sources captured by a standard RGB color camera. The imaging experiment was performed by capturing 32 spectrally multiplexed images containing 1024 x 1280 pixels. Figure 7b-e shows four captured images that represent four single grayscale frames from the spectrally-compressed measurements. The images show the total optical intensity that has passed through the LC phase retarder and was collected by the sensor array at a given shot with a given LC voltage. From the captured data, a window of 700 x 700 pixels was used in the reconstruction process. Using the TWIST solver [33] and orthogonal Daubechies-5 wavelet as the sparsifying operator, a HS datacube with 391 spectral bands (410-800 nm) was reconstructed, yielding a compression ratio of about 12:1. Figure 7f presents a pseudo-color image obtained by projecting the reconstructed HS datacube onto the RGB space. Figure 7g—i displays three images from the reconstructed datacube at different wavelengths (460 nm, 520 nm and 650 nm). Figure 7k—-m demonstrate spectrum reconstruction for three points in the HS datacube and a comparison to the measured spectra of the three respective LEDs with a commercial grating-based spectrometer. The reconstruction PSNR is 32.4 dB, 34.8 dB and 27.9 dB for the blue, green and red LED points, respectively.

Figure 2. LC cell phase retarder. The LC phase retarder is made of a Nematic LC layer (blue arrow) sandwiched between two glass plates and two linear polarizers (green layers). The glass plates are coated with Indium Tin Oxide (ITO, pink layers) and a polymer alignment layer (purple layers). ‘igure 2. LC cell phase retarder. The LC phase retarder is made of a Nematic LC layer (blue arrow)

Figure 4. (a) CS-MUSI acquisition process. (b) CS-MUSI optical scheme diagram. The HS object F(x,y,A) is modulated according to fic (A, V;), yielding the multiplexed measurement G;(x, 1). As the sensor samples discrete values, and for CS analysis and reconstruction purposes, it is more convenient to present the sensing process from Equation (2) in a matrix-vector format. Let us denote the spectral signal with N spectral bands by f €¢ RN*! and the multiplexed measured spectral signal with M entries by g € R“™*!. By using these vectors, the measurement process can be described by:

Figure 8. Staring mode reconstruction result of six different markers. (a) RGB representation of the reconstructed HS image (800 x 900 pixels x 1171 bands). (b) Four reconstructed images at four different wavelengths (470 nm, 530 nm, 580 nm, and 630 nm). Figure 8. Staring mode reconstruction result of six different markers. (a) RGB representation of the

Figure 3. Measured spectral responses (intensity transmission vs. wavelength in nm) of the fabricated LC phase retarder. Each graph represents the spectral modulation with a different voltage applied on the LC cell (15 different voltages).

Figure 6. CS-MUSI spectral response map for voltages from 0 V to 10 V (left map) and a zoom in on the area where the voltages are from 1.3 V to 3.5 V (right map). The theoretical expression in Equation (1) cannot be used directly in order to obtain the system-sensing matrix, because the dependence of the birefringence on the voltage and the materia dispersion are unknown. Consequently, a calibration process in which the spectral responses of the camera were precisely measured, was performed once. Using a point light source as the object and by replacing the sensor of the camera with a commercial high-precision grating spectrometer, the spectra responses of the camera were measured. The calibration process was performed by using a halogen light source with a pre-measured spectrum and by applying voltages from 0 V to 10 V on the LC cel with steps of 2 mV. Figure 6, presents the CS-MUSI system’s spectral response map that was measured in the calibration process.

Figure 10. CS-MUSI camera along-track scanning. Each shot of the CS-MUSI camera, Gj, captures a shifted scene with a different LC spectral transmission (which depends on the voltage vj). The CS-MUSI camera can also be applied in a mode where the camera, the scene, or objects in the scene, are not stationary [44]. Such scenarios include microscope applications with moving cells or scanning platforms, and airborne and remote sensing systems. By capturing a sequence of spectrally multiplexed registration. scanning [44] shots and tracking the object, it is possible to reconstruct HS data by an appropriate t is required that the object appears in M shots. For example, in the case of along-track (Figure 10) the CS-MUSI camera needs 2M measurements in order to capture a scene of the size of the camera FOV. A second requirement is image registration, since the FOV of each shot is slight y different. As a result, before solving Equation (4) it is necessary to register all the measured images along a common spatial grid. This can be done with one of the many available algorithms [45-47].

Figure 9. Staring mode reconstruction results with the dictionary of (a) outdoor and (b) indoor HS images taken with CS-MUSI camera. The figures show RGB representation of the reconstructed HS datacube.

Figure 11. Scanning mode (Figure 10) reconstruction result. (a) RGB color image of three LED arrays. (b-e) representative single exposure images (frame #30, #90, #150 and #300, respectively) and (f-i) the RGB representation of the reconstructed HS image up to the appropriate column.

Figure 13. (a) 4D Spectro-Volumetric imaging. (b) Grayscale representation of HS images at three different depths (225 cm, 254 cm and 270 cm). 6. Target Detection One key usage of spectral imagery is subpixel target detection, when an a priori known spectral signature is sought in each pixel of the spectral datacube. Previous researches dealt with target detection in the reconstructed domain [59], but in the case of the CS-MUSI camera, target detection can be applied in the compressed domain [60,61] since the CS-MUSI camera performs compression only in the spectral domain, without any spatial multiplexing. This yields a significant reduction of processing time and memory storage compared to non-compressing systems, of around an order of magnitude. In order to test the subpixel target detection performance, we used the match filter (MF) algorithm [62], which can be derived by maximizing the SNR, or even by simply considering two 1 wal zz a ee wih Be ee

Figure 5. (a) Realization of the CS-MUSI camera. (b) In-house manufactured LC cell.

Figure 14. (a) Comparison of ROC curves for target detection in conventional (dotted lines) and CS-MUSI (solid lines) HS datacubes. (b) RGB representation of the four HS datacubes in the comparison [42,64,65].

Figure 12. Compressive HS synthetic aperture InIm acquisition setup. After the acquisition of the data and reconstruction of the spectral information from its compressed version, acquired with the moving CS-MUSI camera, the 3D image for each spectral channel (in terms of focal-stack) can be reconstructed numerically in different ways [49,56-58]. One of the most popular methods is based on back-projection, also known as shift-and-add. In the case of synthetic aperture InIm, the refocusing process can be performed as follows [49]:

descriptionView Paper arrow_downwardDownload

Compressive Sensing Hyperspectral Imaging by Spectral Multiplexing with Liquid Crystal

by Adrian Stern

2022, Journal of Imaging

Hyperspectral (HS) imaging involves the sensing of a scene’s spectral properties, which are often redundant in nature. The redundancy of the information motivates our quest to implement Compressive Sensing (CS) theory for HS imaging. This... more

descriptionView Paper arrow_downwardDownload