Tecnura: A Quantitative and Qualitative Performance Analysis of Compressive Spectral Imagers

Tecnura
http://revistas.udistrital.edu.co/ojs/index.php/Tecnura/issue/view/819
DOI: http://dx.doi.org/10.14483/udistrital.jour.tecnura.2017.2.a04
Investigación
A quantitative and qualitative performance analysis

of compressive spectral imagers
Una comparación cuantitativa y cualitativa de análisis de rendimiento
de las cámaras espectrales compresiva
Ferley Medina Rojas1, Henry Arguello Fuentes2, Cristina Gómez Santamaría3
Fecha de recepción: 22 de septiembre de 2016 Fecha de aceptación: 15 de febrero de 2017
Cómo citar: Ferley Medina Rojas; Henry Arguello Fuentes y Cristina Gómez Santamaría. (2017). A quantita-
tive and qualitative performance analysis of compressive spectral imagers. Revista Tecnura, 21(52), 53-67. doi:
10.14483/udistrital.jour.tecnura.2017.2.a04
Abstract and qualitative comparison of these systems in the

Context: Spectral images (SI) contain spatial-spectral same scenarios.
information about a scene arranging in a data cube, Results: This paper analyzes the qualitative and
which often comprises a significant amount of data. quantitative performance of these four compressi-
However, traditional (SI) systems acquire data igno- ve spectral imaging systems to evaluate them in the
ring the high correlation between the measurements same scenarios. For that, the architectures are mo-
and the samples are redundant. Compressive spec- deled as a system of linear equations; then, image
tral imaging systems compress spectral data in the reconstructions are accomplished through the same
acquisition step, so it allows reducing redundancy optimization approach, transmittance, coded aper-
and the data amount. Recently, several spectral ima- ture, and shot numbers.
ging systems have become available, providing new Conclusion: Results show that the performance of
functionality for users and opening up the field to the SSCSI system attains better quality reconstruc-
a wide array of new applications. For instance, the tion in terms of PSNR.
CASSI, SCSI, SSCS, and HYCA systems are four of the Keywords: sampling matrix, compressive sampling
most outstanding systems. architectures, spectral image.
Methods: Some review works have provided com-
prehensive surveys of the available technologies and Resumen
have shown how the new capabilities of spectral Contexto: Las imágenes espectrales (SI) contienen
imaging approaches can be utilized. However, se- información espacio-espectral acerca de una esce-
lecting a specific architecture requires a quantitative na disponible en un cubo de datos que usualmente
1 Systems engineer, agricultural engineer, specialist in regional development management, specialist in telecommunication networks, master
in telematics, candidate for doctor in engineering. Professor of the Universidad Cooperativa de Colombia. Neiva, Colombia.
Contact: ferley.medina@campusucc.edu.co
2 Electronic engineer, master in engineering, doctor in engineering. Professor of the Universidad Industrial de Santander. Bucaramanga, Co-
lombia. Contact: henarfu@uis.edu.co
3 Electronic engineer, master in engineering, doctor in engineering. Professor of the Universidad Pontificia Bolivariana. Medellin, Colombia.
Contact: cristina.gomez@upb.edu.co
Tecnura • p-ISSN: 0123-921X • e-ISSN: 2248-7638 • Vol. 21 Número 52 • Abril - Junio de 2017 • pp. 53-67
[ 53 ]
A quantitative and qualitative performance analysis of compressive spectral imagers
Ferley Medina Rojas; Henry Arguello Fuentes y Cristina Gómez Santamaría.
comprende una cantidad significativa de éstos. Los Sin embargo, para la selección de una arquitectura es-
sistemas tradicionales de (SI) adquieren datos redun- pecífica se requiere una comparación cuantitativa y
dantes ignorando la alta correlación entre las me- cualitativa de estos sistemas en los mismos escenarios.
diciones y las muestras redundantes. Los sistemas Resultados: En este trabajo se analiza el rendimiento
de compresión de imágenes espectrales comprimen cualitativo y cuantitativo de estos cuatro sistemas de
los datos espectrales en la etapa de adquisición, lo compresión de imágenes espectrales para evaluar-
que permite reducir la cantidad de datos y la re- los en los mismos escenarios. Para ello, cada arqui-
dundancia. Actualmente, existen varios sistemas de tectura se modela como un sistema de ecuaciones
imágenes espectrales disponibles que proporcionan lineales y el proceso de reconstrucción de las imá-
nuevas funciones para los usuarios y abren un am- genes se logra con el mismo enfoque de optimiza-
plio campo de nuevas aplicaciones. Por ejemplo, los ción transmitancia, código de apertura y número de
sistemas de CASSI, SCSI, SSCS, y HYCA son cuatro proyecciones.
de los más destacados. Conclusión: Se muestra que el sistema SSCSI alcan-
Método: La revisión de algunos trabajos provee am- za el mejor rendimiento en la reconstrucción con el
plios estudios de tecnologías disponibles y muestra valor más alto PSNR.
cómo se pueden utilizar las nuevas capacidades de Palabras clave: matriz de muestreo, arquitecturas de
los enfoques de formación de imágenes espectrales. muestreo compresivo, imágenes espectrales.
INTRODUCTION measure the similarity between two spectra or de-

tect a target by exploring the data contained in the
Remote-sensed spectral images (SI) contain hi- L-dimensional space by applying detection and
gh-resolution spectral information about mate- classification algorithms (Zare, Bolton, Chanussot,
rials or vegetation present over the earth (Griffin, & Gader, 2014).
May-Hsu, Burke, Orloff, & Upham, 2005), which Spectral remote sensing is a task that contains
corresponds to the surface radiance and reflectan- a large volume of data, regularly redundant. This
ce. SI images are used in areas such as environ- volume of information demands memory, trans-
ment (Mishne, Talmon, & Cohen, 2015), mining mission bandwidth, and information recovery. Re-
(Soydan, Koz, Azebnem Düzgün, & Aydin Alatan, cently, compressive sampling (CS) theory has been
2015), geology (Camacho-Velasco, Vargas-García, developed to make efficient signal sampling and
Rojas-Morales, Castillo-Castelblanco, & Argue- reduce redundancy by taking advantage of the co-
llo-Fuentes, 2015), and soil characterization (Oz- rrelation between measurements (Baraniuk, 2007).
turk, Esin, & Artan, 2015), (Backhaus, Bollenbeck, CS establishes that the Nyquist criteria can
& Seiffert, 2011). Specifically, in the agricultural be overcome (Morteza, 2015; Nyquist, 1928) by
area for monitoring, tracking and controlling the compressing and sampling simultaneously (Ha-
growth of the crop, and identifying weeds, pests, yashi, Nagahara, & Tanaka, 2013). This theory
diseases, and nutrient deficiencies (Arafat, Aboel- is based on the principles of sparseness and in-
ghar, & Ahmed, 2013; Lelong et al., 2010), (Ve- coherence (Arce, Brady, Carin, Arguello, & Ki-
lasco, García, & Fuentes, 2016). Each pixel in an ttle, 2014), (Espitia, Mejía, & Arguello, 2016). In
SI can be expressed as a vector ࢜ ൌ ሾ‫ݒ‬ଵ ǡ ‫ݒ‬ଶ ǡ ‫ ڮ‬ǡ ‫ݒ‬௅ ሿ this way, calculating a set of sparse coefficients
which contains one or more mixed spectra from in some basis (e.g. Wavelet), corresponding to a
different materials of a scene (Golbabaee, Arberet, vector �f � f �� , that
& Vandergheynst, 2013). Therefore, it is possible to represents the spectral image where M and N
[ 54 ]
represent the spatial dimension and L the spectral on compression blocks of pixels in the spectral
dimension), allows to recover the signal without bands according to the compression window size,
distortion (Baraniuk, 2007). The vector f can also which allows simultaneous sampling and com-
be expressed by f �Ψ ss where ‫ א ܛ‬Թே‫ڄ‬ே‫ڄ‬௅ con- pression of SI.
tains only ܵ ‫ ܮ ڄ ܰ ڄ ܰ ا‬non-zero elements and This paper analyzes qualitative and quantitati-
શ ‫ א‬Թே‫ڄ‬ே‫ڄ‬௅ൈே‫ڄ‬ே‫ڄ‬௅ is a sparseness operator. The ve performance of these four compressive spectral
measurements vector ܏ ‫ א‬Թ௤ is obtained as g=Hf, imaging systems in the same conditions. For that,
where ۶ ‫ א‬Թ௤ൈே‫ڄ‬ே‫ڄ‬௅ is a sensing matrix. By choo- the architecture are modeled as a system of linear
sing q projections of H and assuming sparseness equations, and then image reconstructions are ac-
of f in the ψ domain, the signal f can be recovered complished with the same optimization approach.
from model g ��H� ss (Donoho, 2006). This paper is organized as follows: section II shows
In recent years, CS was adopted as an SI sam- the formulation of the sampling model for each
pling protocol. The following architectures were system, section III presents the description of the
selected because they are widely used in CS stu- reconstruction algorithm; simulation results are
dies such as Coded Aperture Snapshot Spectral presented in section IV. Finally, conclusions are
Imaging system (CASSI) (Arguello & Arce, 2010). described in section V.
In this case, the main elements are coded apertu-
re, a dispersive element, and the sensor responsi- COMPRESSIVE SAMPLING
ble for capturing the energy of the encoded scene, ARCHITECTURES
the coded apertures are matricial arrays composed
by translucent optical elements that block or un- Compressive sampling architectures can be mode-
block the path of light through the system. Another led by the system as g = hf, where H is the sam-
well-know system is the Spatial Spectral encoded pling matrix that characterizes the system. The
Compressive hyper-Spectral Imaging system (SSC- parameters of the sampling matrices depend on
SI) (Lin, Liu, Wu, & Dai, 2014). In this system the the specific architecture (CASSI, SSCSI, SCCSI, or
spatial-spectral sampling encoded scheme modu- HYCA). In this section, the formulation of the sam-
lates the RGB bands achieving a spatially chan- pling model of each system is shown by characte-
ging spectral coding that provides a higher degree rizing sampling matrices, correlation between the
of randomness in the measured projections, and spectral and spatial dimensions of the data cube
a coding mask that encodes each spectral chan- components, coded aperture pattern, and disper-
nel independently. On the other hand, a recent sive elements.
architecture is the Snapshot Colored Compressi-
ve Spectral Imager system (SCCSI) (Correa, Argue- Coded aperture snapshots spectral imaging,
llo, & Arce, 2014). In this architecture, an array of CASSI
optical filters encodes the input source, and then
a dispersive element disperses it; both processes Figure 1 shows the Compressive Spectral Imaging
are integrated along the spectral range sensitivi- Acquisition system (CASSI) based on coded aper-
ty of the detector. The last architecture presented tures, represented by T(x,y), which modulates a
in this paper is the Hyper-spectral Coded Apertu- spatial-spectral scene f0 (x,y,λ). The resulting coded
re system (HYCA) (Martin, Bioucas-Dias, & Plaza, field f1 (x,y,λ) is dispersed by a dispersive element,
2015), which is a theoretical architecture based resulting in equation (1).
݂ଶ ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ ൌ ඵ ܶሺ‫ ݔ‬ᇱ ǡ ‫ ݕ‬ᇱ ሻ݂଴ ሺ‫ ݔ‬ᇱ ǡ ‫ ݕ‬ᇱ ǡ ߣሻ ൈ ݄ሺ‫ ݔ‬െ ‫ ݔ‬ᇱ െ ‫ݏ‬ଵ ߣǡ ‫ ݕ‬െ ‫ ݕ‬ᇱ ሻ݀‫ ݔ‬ᇱ ݀‫ ݕ‬ᇱ (1)
[ 55 ]
Ray opticals
Scene
Relay lens
Diffraction grating
Relay lens
Spectral plane
Coding element aperture

& sensor (coded aperture)
Figure 1. Optics CASSI architecture show scene capture until coding element aperture to create a block or unblock
each coded spectral channel.
Source: own work
Where T(x’,y’) represents the transfer function of accounts for the effects of the coded aperture and
the coded aperture, ݄ሺ‫ ݔ‬െ ‫ ݔ‬ᇱ െ ‫ݏ‬ଵ ߣǡ ‫ ݕ‬െ ‫ ݕ‬ᇱ ሻ is the the dispersive element. Figure 2 shows the CASSI
మ
impulse response of the optical system, and s1λ is sampling matrix۶஼஺ௌௌூ ‫ א‬Թ൫ேሺேା௅ିଵሻ൯௤ൈே ௅ , where
the dispersion induced by the dispersive element, q is the number of coded projections (shots) for
assumed as linear dispersion. Compressed measu- N = 5, L = 3 and q = 3. The number of measu-
rements are acquired when the field ݂ଶ ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ is in- rements is ‫ ݌‬ൌ ܰሺܰ ൅ ‫ ܮ‬െ ͳሻ‫ݍ‬. The diagonal pat-
tegrated into the detector (Arguello & Arce, 2010). terns that show in the horizontal direction are the
Equation (1) can be expressed as a system of coded aperture pattern, each time with an N unit
linear equation g = HCASSIf where HCASSI repre- downward shift, as many times as the quantity of
sents the CASSI sensing matrix. The matrix HCASSI spectral bands.
[ 56 ]
N2 L Spectral band1 Spectral band 2
2N
q=1
q=2
q=3
Figure 2. Sampling matrix HCASSI , N = 5, L = 3 and q = 3 and coded aperture pattern with downward shift for each
number of spectral bands
Source: own work
Ray opticals
Scene
Relay lens
Diffraction grating
Relay lens
Spectral plane
Mask da
dm
Figure 3. Optics SSCSI architecture show scene capture until coding element aperture to create each coded spectral
channel.
Source: own work.
[ 57 ]
Spatial-spectral encoded compressive HS Figure 3 shows the calibration parameter for

imager, SSCSI converting spectral coordinate to spatial coordinate;
and b is the distance relation between mask to sen-
The Spatial-Spectral encoded Compressive HS sor and between spectral planes to sensor. The equa-
imager (SSCSI) is illustrated in figure 3. This system tion (2) can be modeled as the matrix system SSCSI
మ మ
allows the reconstruction of a high-resolution spec- g = HSSCSIf where ۶ௌௌ஼ௌூ ‫ א‬Թே ௤ൈே ௅ . The number
tral image from a single sensor, due to the design represents the SSCSI sensing matrix. The matrix
of the optical camera being based on the sampling ‫ ݌‬ൌ ܰ ଶ ‫ ݍ‬accounts for the effects of the coding mask
of spatial-spectral encoded image, which is given that encodes each spectral channel independently.
by ‫݋‬ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ for three-dimensional hyper-spectral Figure 4 shows the SSCSI sampling matrix
మ మ
images (Lin et al., 2014). The image is formed by ۶ௌௌ஼ௌூ ‫ א‬Թே ௤ൈே ௅ The number of measurements
the projection of the hyper-spectral image through is ‫ ݌‬ൌ ܰ ଶ ‫ݍ‬. The diagonal patterns repeated in the
the spectral dimensions Ωλ; it can be expressed as horizontal direction correspond to each coded
equation (2). spectral channel pattern independently, as many
times as the quantity of spectral bands.
݅ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ න ‫݋‬ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ݀ߣ (2)
ஐಓ
Snapshot colored compressive spectral
The coded sensor image can be expressed as imager, SCCSI
equation (3).
Figure 5 Optics SCCSI architecture shows that the
݅ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ න ݂ሺ‫ ݔ‬൅ ܾሺܽߣ െ ‫ݔ‬ሻǡ ‫ݕ‬ሻ‫݋‬ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ݀ߣ (3) Snapshot Colored Compressive Spectral Imager
ஐಓ
(SCCSI) replaces the traditional focal plane array
q=1
q=2
q=3
Figure 4. Sampling matrix HSSCSI, N = 5, L = 3, q = 3 and each coded spectral channel pattern for each number of
spectral bands.
Source: own work.
[ 58 ]
Ray opticals
Scene
Relay lens
Diffraction grating
Relay lens
Spectral plane

Figure 5. Optics SCCSI architecture show scene capture until coding element aperture to create each coded
spectral channel by optical filters array.
Source: own work.
(FPA) by sensors with patterns of optical filters. Fur- reconstruction problem to be better conditioned. The
thermore, a dispersive element allows capturing small number of optical elements makes a compact
both spatial and spectral information from the spec- system that generates a considerable cost reduction
tral image in only one shot. This architecture has an and helps its implementation (Correa et al., 2014).
arrangement of optical filters in the detector, with In the SCCSI system, a dispersive element spec-
different spectral response, which allows obtaining trally disperses the multispectral image ݂଴ ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ
specific spectral components in each pixel. In this is spectrally dispersed by, generating the field
way, the coding scheme is enriched, allowing the ݂ଵ ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ given by equation (4).
݂ଵ ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ ൌ ඵ ݂଴ ሺ‫ ݔ‬ᇱ ǡ ‫ ݕ‬ᇱ ǡ ߣሻ݄൫‫ ݔ‬ᇱ െ ‫ݔ‬ǡ ‫ ݕ‬ᇱ െ ‫ ݕ‬െ ‫ݏ‬ଵ ሺߣሻ൯݀‫ ݔ‬ᇱ ݀‫ ݕ‬ᇱ (4)
Where ‫ݏ‬ଵ ሺߣሻ is the dispersion and h is the im- detector, to obtain a coded and dispersed version
pulse response of the system, f1 is encoded by the of the original signal, expressed as equation (5).
optical filters array ‫ܥ‬ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ located in the FPA
݂ଶ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ න ݂ଵ ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ‫ܥ‬ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ݀ߣ (5)

ஃ
[ 59 ]
Where the system output �� is obtained by ܎ ‫ א‬Թ୑‫୐ڄ୒ڄ‬, and the measurement vector ܏ ‫ א‬Թெ is
integrating product of f1. The equation (5) can be modeled as equation (5).
expressed as g = HSCCSIf where HSCCSI represents the
SCCSI sensing matrix. The matrix HSCCSI is determi- ܏ ൌ ‫ܪ‬ሺ܎ሻ ൅ ‫( ݓ‬6)
ned by the spectral response of the optical filter
matrix structure and represents the proper opera- Where ‫ܪ‬ǣ Թேൈேൈ௅ ՜ Թௐ is an operator that cal-
tion of dispersive elements. culates M inner products between the hyper-spec-
Figure 6 shows the sensing matrix of the SCC- tral vector and f, and w represents the noise level.
మ
SI system ۶ௌ஼஼ௌூ ‫ א‬Թேሺேା௅ିଵሻൈே ௅ , for N = 5, As H is a linear operator, then H(f) can be rewrit-
మ
L = 3 and q = 3. The number of measurements is ten as HHYCAf where ۶ு௒஼஺ ‫ א‬Թெൈ൫ே ௅൯ is the ma-
‫ ݌‬ൌ ܰሺܰ ൅ ‫ ܮ‬െ ͳሻ‫ ݍ‬The diagonal patterns that re- trix modeling the linear operator H.
peat in the horizontal direction correspond to spec- The matrix H is constructed as equation (7).
tral response of the optical filter matrix structure
pattern every time with a unit N downward shift, as ۶ǣ ሺ۶ଵ ǡ ‫ ڮ‬ǡ ۶ேൈே ሻ (7)
many times as the quantity of spectral bands.
Where ܾ݀݅ܽ݃ሺήሻ is the block diagonals, and
Hyper-spectral coded aperture, HYCA ۶௜ ‫ א‬Թ௤ൈ௅ is the matrix that measures each pixel
along the spectral dimension (Martin et al., 2015).
The methodology based on Hyper-spectral Co- In contrast to other architectures, HYCA measu-
ded Aperture (HYCA) relies on two characteristics rement process selects sub-compression matrices
of hyper-spectral images: first, the low dimen- using Gaussian or Bernoulli i.e. random projections
sional space where the hyper-spectral vector be- ‫ܣ‬௜ ‫ א‬ሺ‫ܪ‬ଵ ǡ ‫ ڮ‬ǡ ‫ܪ‬ேൈே ሻ as illustrated in figure 7, where
longs; second, the high correlation between the the data cube is split into a non-overlapping square
spectral and spatial dimensions of the data cube window ws x ws, and ws is the window size.
components. The vectored data cube is defined as
q=1 2N
q=2
q=3
Figure 6. Sampling matrix HSCCSI, N = 5, L = 3, q = 3 and coded aperture pattern with downward shift for each
number of spectral bands.
Source: own work.
[ 60 ]
Figure 8 shows the sensing matrix of the HYCA Coded aperture and transmittance
మ మ
system ۶ୌଢ଼େ୅ ‫ א‬Թே ௤ൈே ௅ , for N = 5, L = 3 and
q = 3. The number of measurements is ‫ ݌‬ൌ ܰ ଶ ‫ ݍ‬. The coded aperture and transmittance define the
The diagonal patterns that repeat correspond to information that the architecture has on each
equation (7). scene. Each one of the architectures has their
Figure 7. Strategy to take measures hyper-spectral data cube, (a) is the way as the data cube is Split into a non-
overlapping square windows ws x ws, (b) code aperture the matrix that measures each pixel along the spectral
dimension.
Source: own work
Figure 8. Sampling matrix HHYCA, N = 5, L = 3, q = 3 and block diagonals matrix that measures each pixel along the
spectral dimension.
Source: own work
[ 61 ]
self-definition of these elements. The coded aper- Summary of architectures

ture is an element that encodes the light before
it is integrated with the detector. The amount of Table 1 shows the characteristics of the different
light that passes is given by the transmittance, as architectures, such as the compression, matrix H
illustrated in figure 9 with different transmittan- and the way each element is coded. Furthermo-
ce percentages. Transmittance can be written as re, the architectures with the best compression rate
equation (8). are SSCSI and, HYCA. However, HYCA is not a real
� architecture, but a theoretical method.
1
�� (8)
�
�� Image reconstruction
Where is t element of the coded aperture and,
th
Q is the amount of total elements in the coded aper- The reconstruction is based on responding a linear
ture. CASSI has values xi ϵ {0,1} (Arguello & Arce, undetermined system of equations from estimating
2010); SSCSI, xi ϵ {0,1} (Lin et al., 2014); SCCSI, f with a minimization function, which is expressed
xi ϵ {0,…,1} (Correa et al., 2014), and HYCA, xi ϵ as equation (9).
{0,…,1} (Martin et al., 2015).
Figure 9. Transmittance percentage, (a) is 30%, (b) is 50%, (c) is 70%.
Source: own work.
Table 1. Features of the compressive sampling architectures
features CASSI SSCSI SCCSI HYCA

Compression ( M ( N  L  1)) q
MNLqMNL MN+L-1qMNL MNLqMNL
ratio MNL
Sensing మ௅
۶஼஺ௌௌூ ‫ א‬Թ൫ேሺேା௅ିଵሻ൯௤ൈே ۶ௌௌ஼ௌூ ‫ א‬Թே
మ ௤ൈே మ ௅
۶ௌ஼஼ௌூ ‫ א‬Թேሺேା௅ିଵሻሻ௤ൈே
మ௅
۶ୌଢ଼େ୅ ‫ א‬Թே
మ ௤ൈே మ ௅
matrix
Block or unblock elements Mask modulate the the arrangement of optical Approaches of physical
Coding
spectral response images in spatial and filters in the detector, with a spatial/spectral coding
element
spectral dimensions different spectral response
Source: own work.
[ 62 ]
example the GPSR algorithm (Figueiredo, Nowak,

� � ��‖� � ��‖ �� ‖�‖�) (9) & Wright, 2007); the TWIST algorithm, which uses
�
an implicit dispersed dictionary as total variation
Where the parameter τ is regularization cons- to solve the optimization problem (Bioucas-Dias
tant and H can be the mentioned sensing matri- & Figueiredo, 2007); the SPARSA algorithm, desig-
ces in table 1. The reconstruction requires using ned to solve optimization problems on a large sca-
algorithms to converge quickly such problems. For le, and it involves the sum of a term of soft error
Table 2. Mean reconstruction PSNR in [dB] compressive sampling architectures
CASSI SSCSI SCCSI HYCA

shots compression % PSNR [dB] compression % PSNR [dB] compression % PSNR [dB] compression % PSNR [dB]
1 13,18 30,69 12,50 32,91 13,18 29,18 12,50 31,49
2 26,36 32,69 25,00 35,06 26,37 30,18 25,00 34,42
3 39,55 33,81 37,50 36,65 39,55 30,53 37,50 37,65
4 52,73 34,96 50,00 37,57 52,73 30,61 50,00 37,59
Source: own work.
and a regularizing term correlated to the GPSR transmittance in the random matrices was 50%,
that is not limited to the regularization l1 (Wright, and shots up to q = 4.
Nowak, & Figueiredo, 2009).
RESULTS
METHODOLOGY
For simulations, sensing matrices depicted in table
The simulations for this paper were developed on 1 performed the compressive measurements. Ta-
two test datasets and several experiments with me- ble 2 details the PSNR as a function of the number
asurements obtained from datasets with different shots, the data compression, and the PSNR values
compression rates from 12,50% to 52,73%, in or- obtained in the reconstruction for CASSI, SSCSI,
der to analyze the reconstruction performance of SCCSI, and HYCA measurements. It also shows
each architecture. We used the Indian Pines and that SSCS from 1 to 2 shots has the highest value
Salinas datasets from AVIRIS sensor for the simu- reached (32,91 [dB] and 35,05 [dB]), while HYCA
lations (Indiana’s, 1992). Indian Pines dataset had from 3 to 4 shots has the highest value reached
145 x 145 pixels, 220 spectral bands ranging from (37,65 [dB] and 37,59 [dB]). Additionally, in all ar-
0.2 to 2.4 mm with a spatial resolution of 20 m. chitectures the higher the number of shots, the hi-
Salinas dataset had 512 x 217 pixels, 224 spec- gher the PSNR of reconstructed images.
tral bands, and spectral resolution of 3.7 m. (Grupo Figure 10 and figure 11 show the Indian Pines
de Inteligencia Computacional de la Universidad and Salinas data cube reconstructions for CASSI,
del País Vasco (UPV/EHU), 2014). We used 8 spec- SSCSI, SCCSI and HYCA architectures, with their
tral bands with 128x128 pixels from each data- respective PSNR. In addition, the best reconstruc-
set. The reconstruction algorithm used was GPSR tion was 35,06 [dB], performed with a 25% com-
(Wright et al., 2009) with up to 200 iterations; the pression of the SSCSI data measurements.
[ 63 ]
Figure 12 and figure 13 show the spectral sig- Pines spectral signature and ends up at 33% abo-
natures of the pixels in the spatial coordinates ve; HYCA spectral signature has intensity value at
(80, 60) and (110, 96) generated for CASSI, SSC- 20% below and finishes at 11% above; and SSCSI
SI, HYCA, and SCCSI architectures of the Salinas follows that behavior. In the case of spectral sig-
and Indian Pines images. CASSI spectral signatu- nature of the SALINAS, spectral signature of the
re has a starting point value at 60% above Indian CASSI starting point value was 140% above and
ORIGINAL CASSI SSCSI SCCSI HYCA
REFERENCE 32.69 [dB] 35.06 [dB] 30.18 [dB] 34.42 [dB]
Figure 10. INDIAN PINES image reconstruction with CASSI, SSCSI, SCCSI, HYCA architectures
Source: own work.
ORIGINAL CASSI SSCSI SCCSI HYCA
REFERENCE 28.81 [dB] 29.72 [dB] 27.61 [dB] 29.05[dB]
Figure 11. SALINAS image reconstruction with CASSI, SSCSI, SCCSI, and HYCA architectures.
Source: own work.
[ 64 ]
(a)
(b)
(a) (b)
Figure 12. The spectral signature of a pixel in the image Indian Pines in CASSI, SSCSI, SCCSI, and HYCA
architectures, (a) is 80,60 point, (b) is 110,96 point.
Source: own work.
(a)
(b)
(a) (b)
Figure 13. The spectral signature of a pixel the Salinas with CASSI, SSCSI, SCCSI, and HYCA
architectures, (a) is 80,60 point, (b) is 110,96 point.
Source: own work.
ends at 10% below; HYCA spectral signature has CONCLUSIONS

intensity value at 80% above and finishes at 10%
below. This can be generalized for both pixels in In this work, we presented a coded aperture pat-
both images. tern, data compression, peak signal-to-noise ratio
(PSNR) and dispersive elements analysis of CASSI,
[ 65 ]
SSCSI, SCCSI, and HYCA compressive sampling Baraniuk, R.G. (2007). Compressive Sensing. IEEE Sig-
systems. In order to compare qualitatively, the sam- nal Processing Magazine, 118–124.
pling matrices of each architectures were depicted. Bioucas-Dias, J.M & Figueiredo, M.A.T. (2007). A new
The main difference between the architectures is TwIST: Two-step iterative shrinkage/thresholding al-
the way to compress the image; CASSI spatially en- gorithms for image restoration. IEEE Transactions on
codes each spectral band equally, and the mea- Image Processing, 16(12), 2992–3004.
surements are the integration on the detector of a Camacho-Velasco, A.; Vargas-García, C.A.; Rojas-Mo-
coded and shifted version of the spectral bands. On rales, F.A.; Castillo-Castelblanco, S.F. & Argüe-
the contrary, SCCSI codes in a space-spectral way, llo-Fuentes, H. (2015). Aplicaciones y retos del
so coding each band is different. The SSCSI measu- sensado remoto hiperespectral en la geologia co-
rements are multiplexed without a shift; and HYCA lombiana Applications and challenges of hypers-
is a theoretical compression method that poses the pectral remote sensing in the colombian geology
spectral attenuation of pixels multiplexed without Aplicações e desafios do sensoriamento remoto hi-
a shift in the spectral bands. perespectral na geologia colombiana. Revista Fa-
The construction of the measurements affects cultad de Ingeniería (Fac. Ing.), 24(40), 17–29.
the quality of the reconstruction in each architec- Correa, C.V.; Argüello, H. y Arce, G.R. (2014). Com-
ture, and this is proved in the results of the simu- pressive spectral imaging with colored-patterned
lations, in which the same parameters are set up detectors. In Acoustics, Speech and Signal Proces-
while comparing the different architectures. SSCSI sing (ICASSP), 2014 IEEE International Conference
reaches the highest PSNR value (35.06 [dB]) with on (pp. 7789–7793).
the 25 % of data compression. Donoho, D.L. (2006). Compressed sensing. IEEE Tran-
sactions on Information Theory, 52(4), 1289–1306.
REFERENCES http://doi.org/Doi 10.1109/Tit.2006.871582
Espitia, M.O.; Mejía, M.Y. & Argüello, F.H. (2016). To-
Arafat, S.M.; Aboelghar, M.A. & Ahmed, E.F. (2013). magrafía computariza: proceso adquisición, tec-
Crop Discrimination Using Field Hyper Spec- nología y estado actual. Revista Tecnura, 20(47),
tral Remotely Sensed Data. Advances in Remo- 119-135.
te Sensing, 02(02), 63–70. http://doi.org/10.4236/ Figueiredo, M.A.T.; Nowak, R.D. & Wright, S.J. (2007).
ars.2013.22009 Gradient projection for sparse reconstruction:
Arce, G.R.; Brady, D.J.; Carin, L.; Arguello, H. & Kittle, Application to compressed sensing and other in-
D.S. (2014). Compressive coded aperture spectral verse problems. IEEE Journal on Selected Topics
imaging: An introduction. IEEE Signal Processing Ma- in Signal Processing, 1(4), 586–597. http://doi.
gazine. http://doi.org/10.1109/MSP.2013.2278763 org/10.1109/JSTSP.2007.910281
Arguello, H. & Arce, G. (2010). Code aperture design Foucart, S. & Rauhut, H. (2013). A Mathematical Intro-
for compressive spectral imaging. In 18th European duction to Compressive Sensing. New York: Springer.
Signal Processing Conference (pp. 1434–1438). Golbabaee, M.; Arberet, S. & Vandergheynst, P. (2013).
http://doi.org/10.1086/180230 Compressive Source Separation: Theory and Me-
Backhaus, A.; Bollenbeck, F. & Seiffert, U. (2011). Ro- thods for Hyperspectral Imaging. IEEE Transactions
bust classification of the nutrition state in crop on Image Processing, 22(12), 5096–5110. http://
plants by hyperspectral imaging and artificial neu- doi.org/10.1109/TIP.2013.2281405
ral networks. In 2011 3rd Workshop on Hypers- Griffin, M.K.; May-Hsu, S.; Burke, H.K.; Orloff, S.M. &
pectral Image and Signal Processing: Evolution in Upham, C.A. (2005). Examples of EO-1 Hyperion
Remote Sensing (WHISPERS) (pp. 1–4). IEEE. http:// data analysis. Lincoln Laboratory Journal, 15(2),
doi.org/10.1109/WHISPERS.2011.6080898 271–298.
[ 66 ]
Grupo de Inteligencia Computacional de la Universi- Morteza, H. (2015). Reducing ADC Sampling Rate
dad del País Vasco (UPV/EHU). (2014). Hyperspec- with Compressive Sensing. ArXiv Preprint ar-
tral Remote Sensing Scenes. Retrieved from http:// Xiv:1503.00311, 1–8.
www.ehu.eus/ccwintco/index.php?title=Hypers- Nyquist, H. (1928). Certain Topics in Telegraph Trans-
pectral{_}Remote{_}Sensing{_}Scenes mission Theory. Transactions of the American Insti-
Hayashi, K.; Nagahara, M. & Tanaka, T. (2013). A user’s tute of Electrical Engineers, 47, 617–644.
guide to compressed sensing for communications Ozturk, S.; Esin, Y.E. y Artan, Y. (2015). Object Detec-
systems. IEICE Transactions on Communications, tion in Rural Areas using Hyperspectral Imaging,
E96-B(3), 685–712. http://doi.org/10.1587/trans- 96432M. http://doi.org/10.1117/12.2195326
com.E96.B.685 Soydan, H.; Koz, A.; Azebnem Düzgün, H. & Aydin
Indiana’s, A.N.W. (1992). Indiana’s Indian Pines Da- Alatan, A. (2015). Hydrocarbon microseepage ma-
taset. Retrieved from https://engineering.purdue. pping using signature based target detection techni-
edu/~biehl/MultiSpec/hyperspectral.html ques. In SPIE Remote Sensing (pp. 96440I-96440I).
Lelong, C.C.D.; Roger, J.M.; Brégand, S.; Dubertret, F.; International Society for Optics and Photonics.
Lanore, M.; Sitorus, N.A.; … Caliman, J.P. (2010). http://doi.org/10.1117/12.2195105
Evaluation of oil-palm fungal disease infesta- Velasco, A.C.; García, C.A. & Fuentes, H.A. (2016). Un
tion with canopy hyperspectral reflectance data. estudio comparativo de algoritmos de detección de
Sensors, 10(1), 734–747. http://doi.org/10.3390/ objetivos en imágenes hiperespectrales aplicados
s100100734 a cultivos agrícolas en Colombia. Revista Tecnura,
Lin, X.; Liu, Y.; Wu, J. & Dai, Q. (2014). Spatial-spec- 20(49), 86-99.
tral encoded compressive hyperspectral imaging. Wright, S.J.; Nowak, R.D. & Figueiredo, M.A.T. (2009).
ACM Trans. Graph., 33(6), 233:1–233:11. http:// Sparse reconstruction by separable approxima-
doi.org/10.1145/2661229.2661262 tion. IEEE Transactions on Signal Processing, 57(7),
Martin, G.; Bioucas-Dias, J.M. & Plaza, A. (2015). 2479–2493.
HYCA: A New Technique for Hyperspectral Com- Zare, A.; Bolton, J.; Chanussot, J. & Gader, P. (2014). Fo-
pressive Sensing. IEEE Transactions on Geoscience reword to the Special Issue on Hyperspectral Image
and Remote Sensing, 53(5), 2819–2831. http://doi. and Signal Processing. Selected Topics in Applied
org/10.1109/TGRS.2014.2365534 Earth Observations and Remote Sensing, IEEE Jour-
Mishne, G.; Talmon, R. & Cohen, I. (2015). Graph-ba- nal of, 7(6), 1841–1843. http://doi.org/10.1109/
sed supervised automatic target detection. IEEE JSTARS.2014.2341811
Transactions on Geoscience and Remote Sen-
sing, 53(5), 2738–2754. http://doi.org/10.1109/
TGRS.2014.2364333
[ 67 ]

Tecnura: A Quantitative and Qualitative Performance Analysis of Compressive Spectral Imagers

Cargado por

Información del documento

Descripción original:

Título original

Derechos de autor

Formatos disponibles

Compartir este documento

Compartir o incrustar documentos

Opciones para compartir

¿Le pareció útil este documento?

¿Este contenido es inapropiado?

Copyright:

Formatos disponibles

Tecnura: A Quantitative and Qualitative Performance Analysis of Compressive Spectral Imagers

Cargado por

Copyright:

Formatos disponibles

Tecnura

A quantitative and qualitative performance analysis

Ferley Medina Rojas1, Henry Arguello Fuentes2, Cristina Gómez Santamaría3

Fecha de recepción: 22 de septiembre de 2016 Fecha de aceptación: 15 de febrero de 2017

Abstract and qualitative comparison of these systems in the

INTRODUCTION measure the similarity between two spectra or de-

Coding element aperture

Source: own work

N2 L Spectral band1 Spectral band 2

Source: own work

Source: own work.

Spatial-spectral encoded compressive HS Figure 3 shows the calibration parameter for

N2 L Spectral band1 Spectral band 2

Source: own work.

Coding element aperture

Source: own work.

݂ଶ ሺ‫ݔ‬ǡ ‫ݕ‬ሻ ൌ න ݂ଵ ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ‫ܥ‬ሺ‫ݔ‬ǡ ‫ݕ‬ǡ ߣሻ݀ߣ (5)

N2 L Spectral band1 Spectral band 2

Source: own work.

Source: own work

Source: own work

self-definition of these elements. The coded aper- Summary of architectures

Figure 9. Transmittance percentage, (a) is 30%, (b) is 50%, (c) is 70%.

Source: own work.

Table 1. Features of the compressive sampling architectures

features CASSI SSCSI SCCSI HYCA

Source: own work.

example the GPSR algorithm (Figueiredo, Nowak,

Table 2. Mean reconstruction PSNR in [dB] compressive sampling architectures

CASSI SSCSI SCCSI HYCA

Source: own work.

ORIGINAL CASSI SSCSI SCCSI HYCA

REFERENCE 32.69 [dB] 35.06 [dB] 30.18 [dB] 34.42 [dB]

Source: own work.

ORIGINAL CASSI SSCSI SCCSI HYCA

REFERENCE 28.81 [dB] 29.72 [dB] 27.61 [dB] 29.05[dB]

Source: own work.

ends at 10% below; HYCA spectral signature has CONCLUSIONS

También podría gustarte