Professional Documents
Culture Documents
D1 - Us20170205337a1
D1 - Us20170205337a1
insigrg Re:cistrictics
process process
Patent Application Publication Jul. 20, 2017. Sheet 1 of 13 US 2017/0205337 A1
N
Patent Application Publication Jul. 20, 2017 Sheet 2 of 13 US 2017/0205337 A1
S.
w
saw
N cy
:: SES...St.E. ::::::::
-: is
-: &
s: :-
Patent Application Publication Jul. 20, 2017 Sheet 3 of 13 US 2017/0205337 A1
vem
S. S
ZZZZZZZZ,
Fig. 1C
Patent Application Publication Jul. 20, 2017 Sheet 4 of 13 US 2017/0205337 A1
Patent Application Publication Jul. 20, 2017 Sheet 5 of 13 US 2017/0205337 A1
3.
2 3. i 5 S. FO 3. S
Eac rurae
& 2
re
s
s
E
Q.15
3. 4. SO s 7 SO 3.
Band nurther
Fig. 2B
Patent Application Publication Jul. 20, 2017 Sheet 6 of 13 US 2017/0205337 A1
. S3
.2
3. SC 9
an inter
.8
S
.
.
8
0.06 2 3. O S. f 8 3
Band firs:
Fig. 2C
Patent Application Publication Jul. 20, 2017 Sheet 7 of 13 US 2017/0205337 A1
Fig. 3A
Patent Application Publication Jul. 20, 2017 Sheet 8 of 13 US 2017/0205337 A1
Fig. 3B
Patent Application Publication Jul. 20, 2017 Sheet 9 of 13 US 2017/0205337 A1
SS { Sa: s:
58 s 5. OO | 8 533 8. SS: s: s:
25 S38 3. 32 85.
75 3i SS s
58 52 St.
Fig. 3C
Patent Application Publication Jul. 20, 2017. Sheet 10 of 13 US 2017/0205337 A1
12
Patent Application Publication Jul. 20, 2017. Sheet 11 of 13 US 2017/0205337 A1
3.
Wavelength
8.
Wavetergh
Fig. 5A
Patent Application Publication Jul. 20, 2017. Sheet 12 of 13 US 2017/0205337 A1
Wavelength
Fig. 5B
Patent Application Publication Jul. 20, 2017. Sheet 13 of 13 US 2017/0205337 A1
gae*
ž*.… : &~* …. *
** …. «* &….
*~ . , º ….
S.
xx-xx-xx-xxssssssssssssssss-xx-xx-xx-xxsssks-stax-stax-s-s-s-s-s
SSS SSSSSS
Fig. 6
…«^ SS-S-S-S-S-S-XXXX-XXXXXX-SS
: 38
8.
US 2017/0205337 A1 Jul. 20, 2017
of data from less than 5% of its size, as opposed to 33% the plurality of SFA elements to filter one or more spectral
Subsampling rate of color imaging. bandwidths centered each at specific wavelengths corre
0022. Interferometric Spectral Filter Array sponding to that SFA element, whereby all of the plurality of
0023. Due to the very stringent requirements of spectral SFA elements taken together cover a determined spectral
imaging, a spectral filter array is much more difficult to range; and setting the specific wavelengths of each SFA
produce than the color arrays used in commercial color element of the plurality of SFA elements on the surface such
cameras. Color cameras generally employ pigment-based to obtain a uniform and aperiodic spatial distribution of all
filters whose spectral transmission bandwidth is too wide of the plurality of SFA elements across the surface. The
and irregular for the radiometric measurements of spectral sampling of the spectral property of the image of the object
imaging. The number of different filters that can be produced further comprises providing an image sensor configured to
in a pigment-based color filter array is also limited by the record at each pixel the light filtered by one of the plurality
number of pigments used. The cost of a color filter array thus of SFA elements or a subset of the plurality of SFA elements
greatly increases with the increase in the number of different thereby producing one intensity value of light filtered by the
filters. one of the plurality of elements or the subset of the plurality
0024 Interferometric filters, such as Fabry-Perot filters, of SFA elements per pixel; forming the image of the object
are based on a different principle than pigment filters and on the SFA through a lens or group of lenses; and recording
enjoy a number of properties useful for imaging spectros for all of the pixels of the image sensor the spectrally filtered
copy: intensity values thereby obtaining a 2-dimensional array of
0025 finely tunable central wavelength: the central the intensity values corresponding to the image of the object.
wavelength is given by the distance between the reflec The method for obtaining spectral imaging data further
tive surfaces used to make light interfere with itself. comprises the step of reconstructing a full 3 dimensional
Varying this distance changes the central wavelength of spectral data cube of the imaged object from the sampled
the filter without significantly altering the shape of its 2-dimensional array.
spectral transmission curve; 0032. In a preferred embodiment the step of reconstruct
0026 symmetry around central wavelength: the filter ing comprises a mathematical modelling and computational
transmission is locally symmetric with respect to the numerical optimization.
central wavelength, due to the natural interferometric 0033. In a further preferred embodiment, the mathemati
process; cal modelling and computational numerical optimization
(0027 finely tunable bandwidth: the filter bandwidth is comprises a machine learning method which enables an
given by the reflectivity of the surfaces which reflect absence of transmission measurement by the SFA.
the light;
0028 the production cost of a large number of different 0034. In a further preferred embodiment, the mathemati
wavelength interferometric filters is much lower than cal modelling and computational numerical optimization
for pigment-based filters. comprises at least a convex optimization method based on
transmission measurement of the SFA.
0029 While their properties make them an obvious
choice for imaging spectroscopy, until recently, interfero 0035. In a further preferred embodiment, the step of
metric filters were not feasible to create at the size of a single reconstructing comprises providing measured transmissions
pixel. Fabry-Perot filters were thus used to produce spectral of individual SFA elements of the SFA; creating a system
images by band-sequential scanning, as large global filters, design matrix from measured transmissions of individual
which cover the entire sensor surface 8. However, recent SFA elements and image sensor radiometric calibration; and
developments in miniaturization and nanotechnology have inferring non sampled spectral information using deconvo
led to the development of interferometric spectral filter lution or non-linear sparse reconstruction methods based on
arrays with the individual filters being no larger than a pixel the system design matrix.
4. 0036. In a further preferred embodiment, the step of
0030. One of the main disadvantages of interferometric setting the specific wavelengths of each SFA element of the
filters is their second and higher order transmissions. These plurality of SFA elements on the surface such to obtain the
higher order transmissions of the filters let light pass not uniform and aperiodic spatial distribution of the specific
only at the desired central wavelength, but also at twice that wavelengths of the SFA is deterministic, thereby comprising
wavelength, three times that wavelength and so on. Inter an aperiodic tiling Such as Wang or Penrose.
ferometric filters can also be configured as multiband filters, 0037. In a further preferred embodiment, the step of
due to the higher order transmissions which include multiple setting the specific wavelengths of each SFA element of the
specific wavelengths for which they transmit light. In most plurality of SFA elements on the surface such to obtain the
cases, the higher order transmissions are undesired effects uniform and aperiodic spatial distribution of the specific
and need to be nullified. wavelengths of the SFA further comprises filling the SFA
Surface with a repeating pattern, then continuously and
SUMMARY OF THE INVENTION randomly inter-changing the SFA elements until the mea
0031. In a first aspect, the invention provides a method sured entropy of the SFA central wavelengths is sufficiently
for obtaining spectral imaging data. The method comprises high.
at least the steps of receiving a sample set of data generated 0038. In a further preferred embodiment, the step of
by sampling a spectral property of an image of an object in setting the specific wavelengths of each SFA element of the
a spatial basis, wherein the sampling of the spectral property plurality of SFA elements on the surface such to obtain the
of the image of the object comprises providing a Spectral uniform and aperiodic spatial distribution of the specific
Filter Array (SFA) by arranging a plurality of SFA elements wavelengths of the SFA further comprises a random sam
together to form a Surface; configuring each SFA element of pling of the SFA elements from a uniform distribution.
US 2017/0205337 A1 Jul. 20, 2017
0039. In a further preferred embodiment, the step of form aperiodic Spectral Filter Array (SFA) by arranging a
setting the specific wavelengths of each SFA element of the plurality of SFA elements together to form a surface, con
plurality of SFA elements on the surface such to obtain the figuring each SFA element of the plurality of SFA elements
multiband distribution of the specific wavelengths of the to filter multiple spectral bands centered at specific wave
SFA further comprises a sampling of multiband SFA ele lengths corresponding to that SFA element, whereby all of
ments such that neighboring SFA elements are spectrally the plurality of SFA elements taken together cover a deter
orthogonal. mined spectral range, and setting the central wavelength of
0040. In a second aspect, the invention provides a method each SFA element of the plurality of SFA elements on the
for Spectral Filter Array (SFA) element second-degree trans Surface Such that neighboring SFA elements are spectrally
mission cancellation through SFA design and Subsequent orthogonal; from a 2-dimensional array of pixel intensity
response Subtraction. The method comprises at least design values, spatially interpolating each channel corresponding to
ing a uniformly and aperiodically distributed Spectral Filter an SFA element configuration independently at pixel loca
Array (SFA) by arranging a plurality of SFA elements tions where other SFA element configurations are present;
together to form a Surface; configuring each SFA element of and averaging the independently interpolated channels, to
the plurality of SFA elements to filter a spectral bandwidth create a high resolution monochromatic image.
centered at a central wavelength corresponding to that SFA 0043. In a fifth aspect, the invention provides a method
element, whereby all of the plurality of elements taken for generating a spectral 3D-model of an imaged object and
together cover a determined spectral range; and setting the increasing the spectral reconstruction quality of an imaged
central wavelength of each SFA element of the plurality of scene, comprising taking multiple images of the scene, each
SFA elements on the surface such to obtain a distribution of of the images comprising at least an array of pixel intensi
all the central wavelengths to be uniform and aperiodic over ties; using monochromatic previews to optically align the
the surface. The method further comprises identifying the images through image registration, thus generating the
SFA elements with second degree transmission in the sen 3D-model of the scene; grouping values from different ones
sitivity range of an image sensor, changing the central of the arrays of pixel intensities corresponding to different
wavelength of SFA elements neighboring the SFA elements images of the multiple images, if the pixels are aligned to a
with second degree transmission, as to match the central same point in the imaged scene; based on grouped values,
wavelength of the second degree transmission; re-arranging re-modeling a system design matrix corresponding to mul
the SFA element central wavelengths which were not tiple arrays of pixel intensities, taken from different loca
changed or do not have a second degree transmission, as to tions; and reconstructing the spectral texture of the 3D
maintain the uniformity of the SFA central wavelength model based on the cross-image system design matrix.
distribution; and in a 2-dimensional array of pixel intensity 0044. In a sixth aspect, the invention provides a computer
values, Subtracting the response corresponding to the program stored on a memory device, the computer program
changed neighboring elements from the response corre when executed by a computing device reading the memory
sponding to the SFA elements which have a second degree device enabling to reconstruct a 3D-hyperspectral image
transmission. from a 2D-spatial-spectral dataset, by implementing the
0041. In a third aspect, the invention provides a method steps of the method for obtaining spectral imaging data as
described herein above.
for obtaining a real-time monochromatic preview of an 0045. A corresponding hyperspectral image reconstruc
imaged object, the imaged object having been obtained tion module comprises a single- or multi-processor comput
using the method of obtaining spectral imaging data
described herein above, the method for obtaining the real ing device, or a computer network programmed to recon
time monochromatic preview comprising designing a uni struct a 3D hyperspectral image from a 2D dataset. The
form aperiodic Spectral Filter Array (SFA) by arranging a underlying algorithmic framework of the hyperspectral
plurality of SFA elements together to form a surface, con image reconstruction may include deconvolution, non-linear
figuring each SFA element of the plurality of SFA elements sparse reconstruction, regressive machine learning methods,
to filter a spectral bandwidth centered at a central wave projection of the acquired spectral samples onto a set of a
length corresponding to that SFA element, whereby all of the priori known spectral signatures, as well as any combination
plurality of SFA elements taken together cover a determined of the herein three methods. Algorithmic methods are also
spectral range, and setting the central wavelength of each provided for variable spatio-spectral resolution reconstruc
SFA element of the plurality of SFA elements on the surface tion of the imaged scene, based on a fixed configuration
such to obtain a distribution of all the central wavelengths to imager.
be uniform and aperiodic over the Surface; generating a BRIEF DESCRIPTION OF THE FIGURES
subset of SFA elements by a 2-dimensional periodic selec
tion; replacing the SFA elements belonging to the above 0046. The invention will be better understood in view of
subset with SFA elements of identical transmission; and the description of preferred embodiments and in reference to
from a 2-dimensional array of pixel intensity values, using the drawings, wherein:
a Subset of values, corresponding to the Subset of periodic 0047 FIG. 1A is a schematic overview of the core
SFA elements with identical transmission, to create a lower imaging design and process;
resolution monochromatic image. 0048 FIG. 1B is a variation of the 2D spectral sampling
0042. In a fourth aspect, the invention provides a method imager based on a spectral filter array adapter fitted between
for obtaining a real-time monochromatic preview of an the main imaging lens and the imaging sensor,
imaged object, the imaged object having been obtained 0049 FIG. 1C is a variation of the 2D spectral sampling
using the method of obtaining spectral imaging data imager where the reconstructor is integrated into the imager;
described herein above, the method for obtaining the real 0050 FIG. 2A shows a color preview of an aerial image
time monochromatic preview comprising designing a uni scene, the corresponding sensor readout and the color pre
US 2017/0205337 A1 Jul. 20, 2017
view of the reconstructed 100 band spectral data cube, as data cube as it is reconstructed from the relatively small 2D
computed by a simulation. Four points of interest are marked set of optical samples that does not exceed the number of
on the reconstructed image: pixels of the imaging sensor, while the full data cube
0051 FIG. 2B-C shows a spectral comparison between contains as many elements as the number of pixels on the
the original image scene sampled at 100 bands (marked sensor multiplied by the number of registered spectral
“Original') and the reconstructed spectral data cube (marked bands. The size ratio between the 2D set of samples and the
“Reconstructed’), for each of the four points marked in FIG. reconstructed hyperspectral data cube can vary depending
2: on the configuration, but will typically be between 1/10 and
0052 FIG. 3A shows an example of a uniformly distrib 1A2OO.
uted and aperiodic 12x12 element spectral filter array pat 0063 Camera
tern given by central wavelength in nanometers. 61 possible 0064. A core embodiment of the invention comprises an
filters were considered with central wavelengths between imaging spectroscopy camera system depicted in FIG. 1A,
400 and 100 nm, at a step of 10 nm: employed for producing a 2D set of spectral samples 8 and
0053 FIG. 3B is a variation of FIG. 3A where elements a computational method which is utilized to reconstruct a
having a non-negligible second degree transmission are complete 3D spectral data cube from the aforementioned 2D
always placed in the direct vicinity of elements filtering at set of optical samples. More specifically, the camera com
twice their central wavelength. These filter pairs are high prises an optical lens 5, or a group of lenses, used for
lighted in gray; producing an image of the object of interest 1 on the image
0054 FIG. 3C is a variation of FIG. 3A where the plane. The camera further comprises a spectral filter array
aperiodic pattern is interleaved with a periodic pattern of (SFA) 6 wherein the various spectral components of the
non-filtering elements or elements having a common spec objects image are transmitted to the image sensor, and the
tral transmission. The elements of the periodic pattern are image sensor array 7 arranged to detect the light transmitted
highlighted in gray; by the individual SFA elements. The camera also contains an
0055 FIG. 4 depicts an image filtering configuration analog-to-digital signal converter and a storage medium 9.
where the spectral filter array is preceded by an anti-aliasing on which the resultant 2D datasets are stored.
filter and followed by a color filter array, along the optical 0065. The corresponding spectral data cube reconstruc
path of the light to the imaging sensor, tion module 3, or reconstructor, comprises a single- or
0056 FIG. 5A-B contain example plots of the continuous multi-processor computing device, or a computer network
and discretized spectral transmission of a Fabry-Perot spec programmed to reconstruct a 3D spectral image from the 2D
tral array element with a central wavelength of 700 nm and set of optical samples.
a full width at half maximum of 100 nmi; and 0066. The underlying algorithmic framework of the spec
0057 FIG. 6 shows example plots of transmissions of tral image reconstruction may include deconvolution, non
two Fabry-Perot spectral array elements designed for mul linear sparse reconstruction, projection of the acquired spec
tiband filtering. Variations in both multiband wavelength tral samples onto a set of a priori known spectral signatures,
number and bandwidth are illustrated. A wideband filter as well as any combination of the herein described three
designed with 6 specific wavelengths (above) and a narrow methods.
band filter designed with 5 multiband wavelengths (below) 0067. The imaging process is as follows:
are plotted. 0068. The imaged scene is focused by the lens onto the
DETAILED DESCRIPTION OF THE
plane of the SFA.
INVENTION 0069. Each element of the SFA lets only certain fre
quencies of light pass through it, depending on its
0058. In this section, we outline the advantages of the transmission curve.
presented solution with respect to prior art and Subsequently 0070 Light transmitted by the SFA falls onto the
focus on the technical details pertaining to three major imaging sensor which generates, for each pixel, an
aspects of the invention, namely: electrical charge in a manner proportional to the
0059) 1... the camera design and functionality: amount of light received by the pixel.
0060 2. the spectral filter array design; and 0071. The electrical charges in the imaging sensor are
0061 3. the reconstructor embodiments and preferred discretized and digitized by an analog-to-digital con
implementation. Verter, producing the sensor readout.
0062. The present invention improves upon prior art in 0072 The sensor readouts are then stored onto the
multiple ways. With respect to spatial or spectral scanning storage medium.
spectral imaging systems, the present invention overcomes 0073. The reconstructor will receive the sensor read
the need to scan, sampling the imaged scene simultaneously outs from the storage medium and will reconstruct the
while providing similar resolution. Due to the extremely full 3D spectral image.
short optical path of the present invention, a much smaller 0074. In some embodiments, the camera of this invention
and lighter spectral image acquisition system can be built is based on an interchangeable lens camera as depicted in
than those requiring spectrally dispersive elements. The FIG. 1B. The spectral filter array 6 and a refocusing lens 14
reduced complexity of the optical and mechanical parts also are built into an adapter 13 which fits in between the lens 5
significantly lowers the production cost when compared to and the imaging sensor 7 of the interchangeable lens camera.
spectral imaging systems with dispersive elements. With The image from the lens is focused on the spectral filter
respect to state-of-the-art Snapshot spectral acquisition sys array, then refocused onto the imaging sensor by the refo
tems, the present invention generates significantly higher cusing lens. In these embodiments, any interchangeable lens
resolution spectral data cubes. The present invention greatly color camera becomes the core of this invention by using
reduces the amount of storage required to acquire a spectral Such an adapter 13.
US 2017/0205337 A1 Jul. 20, 2017
0075 Various embodiments of the main imaging lens 5 filter, a pigment-based color filter, or a non-filtering element
include a fixed focal optical lens, or lens groups, while transmitting all wavelengths of light. The pixels under these
others feature a variable focal and variable focus lens. periodic elements will be used to create preview images 15
0076 Some embodiments see the storage medium 9 of the imaged scene. These images can be used as real-time
integrated in the camera while others have no storage previews of the imaged scene or directly in image registra
medium, the images being sent to a computer or a computer tion and image based 3D reconstruction.
network directly after acquisition and analog-to-digital con 0088. Some embodiments of the invention include a
version. camera where an anti-aliasing filter 12 is placed before the
0077 Spectral Filter Array spectral filter array, as depicted in FIG. 4, for homogenizing
0078 Most embodiments of the present invention include the light reaching the different filters. The anti-aliasing filter
a uniformly distributed aperiodic spectral filter array 6, may be comprised by: birefringent filters extending over
which we refer to as SFA, composed of individual spectral more than 2x2 of the sensor elements, a defocused objective
filters, referred to as elements. An example of such an SFA lens with a matched aperture stop, and/or a degraded imag
embodiment is shown in FIG. 3A where each element is ing lens.
represented by its central wavelength. In other implemen I0089 Spectral Cube Reconstruction
tations, the SFA may also contain non-filtering elements, 0090. In various embodiments of the invention, the
while preserving an aperiodic distribution of interferometric reconstructor 3 is either a single computer, a programmable
filters across the majority of its surface, as shown in FIG.3C. single or multi-processor 10 device or a computer network
In the preferred implementation, the SFA is composed of programmed to reconstruct the hyperspectral image from an
interferometric filters such as Fabry-Perot filters. acquired two-dimensional dataset 8 through non-linear
007.9 The uniform aperiodic distribution may be sparse reconstruction. The reconstructor can also be a virtual
obtained in several ways: machine, accessible through the internet and ran on cloud
0080 deterministically through aperiodic tiling such as based computational servers.
Wang or Penrose tiling. 0091 Various embodiments feature the reconstructor
I0081 by filling the SFA surface with a repeating pat integrated in the camera and producing the hyperspectral
tern, then continuously and randomly inter-changing image immediately after having acquired the two-dimen
the SFA elements until the measured entropy of the sional set of optical samples. FIG. 1C depicts such an
SFA central wavelengths distribution is sufficiently integrated camera, composed of the imaging optics 5 which
high. focus the light onto the SFA 6 that is deposited over the
I0082 by random sampling the SFA elements from a imaging sensor 7. Data from the imaging sensor 8 is then
uniform distribution. saved on the storage medium 9 and converted to spectral
0083. Some embodiments of the invention include SFA data by the reconstructor 3, composed of multiple process
designs where the second and higher degree transmissions of ing units 10.
the interferometric spectral filters are nullified by means of 0092. The mathematical model describing the sampling
several methods: and discretization of the imaged scene by the hyperspectral
I0084. A method for second degree cancellation con imager of this invention is described by the following linear
sists of placing the SFA 6 over an existing color filter system:
array 11 (Such as a Bayer pattern filter array), matching Ax=y, (1)
the individual SFA elements transmission so that only
the first order response will be transmitted by the color where A is the system design matrix containing the spectral
filter array, as shown in FIG. 4. transmission of the individual elements of the SFA, y is the
0085. Another method for second order transmission sensor readout, and X is the hyperspectral data cube corre
cancellation is always placing a SFA element with its sponding to the spatio-spectral properties of the imaged
transmission equal to that of the second degree of one scene. Vectors x and y are thus serialized versions of the 3D
of its adjacent elements, if those filters have 2nd degree hyperspectral and 2D sensor datasets.
transmission in the sensitivity range of the sensor. The (0093. While the SFA elements sample the continuous
design depicted in FIG. 3B, allows for subtracting the incoming light spectrum, the design matrix A contains
second degree transmission based on adjacent filters, discrete versions of the SFA transmissions, whose spectral
prior to reconstruction. resolution will be equal to the number of spectral bands of
0.086. Some embodiments of the invention use interfero the reconstructed data cube. This property allows for the
metric filters with multiband wavelength transmissions, reconstruction of a various number of spectral bands, this
rather than central wavelengths, as the elements of the SFA. number having only a lower bound, dictated by the mini
Examples of these multiband filter transmissions are shown mum number of bands required to properly represent the
in FIG. 6. In these embodiments, the spatial uniform ape SFA element transmission curves. In FIGS. 5A and 5B, the
riodic distribution of the elements is chosen to maximize the effects of the aforementioned discretization and lower bound
spectral orthogonality of the multiband filters. By spectral are depicted: while 30 and 10 bands produce an adequate
orthogonality between filters we mean that they respond representation of the SFA element transmission, 3 bands are
differently to the same wavelengths. Maximizing spectral insufficient.
orthogonality translates to choosing filters which do not 0094 Under the hypothesis of local spectral homogeneity
correlate spectrally. of the imaged scene, the number of spectral samples in y can
0087. In various embodiments, an element at periodic be artificially increased by pixel grouping, resulting in a
positions in the SFA has the same transmission across the non-diagonal system design matrix A. This hypothesis can
entire spectral filter array as shown in FIGS. 4 and 1A. This be further enforced by the use of an anti-aliasing filter placed
periodic element can be either an interferometric spectral in front of the SFA. Each spatial pixel of the image sensor
US 2017/0205337 A1 Jul. 20, 2017
readout constitutes a single spectral sample of the complete age-thresholding algorithm (FISTA) introduced in 10.
spectral response of the corresponding point in the imaged using the total variation (TV) norm as the penalty term. The
scene. The reconstruction process may include the grouping reconstructed data cube is thus found by minimizing:
of multiple spectral samples per pixel, which are derived argmin(Ax-y2+ x(t) (4)
from the spatial pixels in the close vicinity of the processed
pixel. where ||x||. can either be the 3D TV norm of the entire
0095 One method for pixel grouping may be a sliding 2D reconstructed hyperspectral data cube, or the sum of the 2D
window, which groups all spectral samples from pixels TV norms of the individual spectral bands of the recon
falling inside its area into the central pixel. If the SFA structed hyperspectral data cube. Using the sum of 2D TV
contains a periodic interleaved pattern, another method for norms rather than the 3D TV norm has the advantage of
pixel grouping may be guided by the contrast gradients independently processing spectral bands and allowing for
observed in the dense image directly obtained from the parallelization of the most computationally intensive steps
pixels behind periodic elements. Using this dense image of FISTA.
allows for grouping of pixels which do not span across edges 0101 If the measurements y were produced by a system
or high contrast areas, thus reducing the possible artefacts configuration having an anti-aliasing filter in the optical
produced by pixel grouping. path, the quality and speed of the reconstruction produced by
0096. The variable resolution discrete filter representa FISTA can be improved by independently blurring the
tion and the pixel grouping methods described above work spectral bands of the estimated reconstruction at each itera
in unison to allow for the adjustment of the reconstructed tion of the algorithm, in an amount proportional to the
hyperspectral data cube's spatial and spectral resolutions. blurring induced by the anti-aliasing filter. This process can
Depending on the desired hyperspectral data cube resolution be parallelized.
desired, only A and y need to be recomputed accordingly, 0102. A simulation of the imaging and reconstruction
while the hyperspectral imager configuration remains fixed. process of the invention has been used to validate the
0097. Reconstruction of the hyperspectral cube can be principles of the present invention, the results of which are
achieved by applying regressive machine learning methods presented in FIGS. 2A and 2B. The target compression ratio
12 Such as neural networks or random forests to the sensor was 1/100, meaning the hyperspectral data cube was recon
readout or pixel-grouped sensor readout. These methods structed from 100 times fewer measurements than its total
take in an input signal or feature (y) and output an estimation number of elements. ASFA model based on Wang tiling and
of the original signal (X). Through training with a large Fabry-Perot filter transmissions was employed, containing
number of examples, models such as neural networks can 100 filters of central wavelengths between 400 and 1000 nm,
learn to un-mix and correct spectral data from the sensor and 100 nm full width at half maximum (FWHM). The SFA
readout. element's transmissions were discretized in 100 spectral
0098. Another method for reconstructing the hyperspec bands. The SFA also contained a periodic pattern of elements
tral data cube X from the measurements y can be modelled having the central wavelength at 700 nm, as shown in FIGS.
as a convex optimization problem: 3C and 5. Gradient-based pixel grouping was applied,
argmin(Ax-y2+c(x)) (2) grouping together 4 pixels from a vicinity of 5x5 pixels.
where the function c(x) is an always positive penalty term, FISTA was configured to minimize the sum of 2D TV norms
of independent bands in parallel and included the blurring
constraining the search space for the optimal X based on a step at each iteration.
priori knowledge of the properties of X. The function c(x) (0103. The obtained simulated sensor readout as well as a
will have low values for instances of X corresponding to a color preview of the hyperspectral data cube are shown in
priori knowledge, while |Ax-y will have low values for FIG. 2A. Four points of interest have been marked on the
instances of X fitting the measured data y. The convex reconstructed image, presenting various spectral properties
optimisation (2) can thus be interpreted as finding the as well as being located in areas of various spatial detail
spectral data cube X which simultaneously best fits the across the imaged scene. The spectral properties of the
measured data y and the a priori knowledge about its original spatio-spectral image scene and the reconstructed
structure c(X). hyperspectral data cube are compared in these four points,
0099 Variations of the aforementioned reconstruction the results shown in FIGS. 2B and 2C. The spectral reso
model include changing the representation of X to a math lution is notable, where the reconstructed spectrum exhibits
ematical basis in which X becomes sparse. For instance, detail at much finer resolution than the FWHM of the SFA
basis such as Fourier, direct cosine transform, wavelet or elements used to sample the spectrum. If a spectral camera
gradient are known to be sparse for natural images 9. In with 100 nm wide filters at half-maximum had directly
these cases, (2) may become: sampled Such a spectrum, sharp spectral features such as
argmin(ASx-y+WSx) (3) peaks would have been smoothed out by these filters through
where S is the matrix representation of the transformation the sampling process. However, in our reconstructions we
from the chosen sparse mathematical basis and c(x) is clearly see sharp peaks and Valleys spanning less than 10
replaced with the L norm which enforces sparsity on the bands (or 60 nm), well below the FWHM of the SFA
representation of X, while A controls the strength of penal elements used for the spectral sampling.
ization. For hyperspectral data cubes, other transformations 0104 Commercial Applications
can be envisioned, such as projecting the measured data onto 0105 Imaging spectroscopy technology has numerous
a known set of spectral vectors, obtained through principal and proven applications in research and commercial
component analysis, for instance. domains, including agriculture, natural resource manage
0100. The preferred implementation of the reconstruction ment, mineralogy, medicine and manufacturing among oth
algorithm is based on a version of the fast iterative shrink ers. Specifically, the disclosed invention facilitates the
US 2017/0205337 A1 Jul. 20, 2017
providing an image sensor configured to record at each 10. A method for Spectral Filter Array (SFA) element
pixel the light filtered by one of the plurality of SFA second-degree transmission cancellation through SFA
elements or a subset of the plurality of SFA elements design and Subsequent response Subtraction comprising at
thereby producing one intensity value of light filtered least:
by the one of the plurality of elements or the subset designing a uniformly and aperiodically distributed Spec
of the plurality of SFA elements per pixel, tral Filter Array (SFA) by
forming the image of the object on the SFA through a arranging a plurality of SFA elements together to form
lens or group of lenses, a Surface,
recording for all of the pixels of the image sensor the configuring each SFA element of the plurality of SFA
spectrally filtered intensity values thereby obtaining elements to filter a spectral bandwidth centered at a
a 2-dimensional array of the intensity values corre central wavelength corresponding to that SFA ele
sponding to the image of the object, ment, whereby all of the plurality of elements taken
reconstructing a full 3 dimensional spectral data cube of together cover a determined spectral range, and
the imaged object from the sampled 2-dimensional setting the central wavelength of each SFA element of
array. the plurality of SFA elements on the surface such to
2. The method of claim 1, wherein the step of reconstruct obtain a distribution of all the central wavelengths to
ing comprises a mathematical modelling and computational be uniform and aperiodic over the surface;
numerical optimization. identifying the SFA elements with second degree trans
3. The method of claim 2, wherein the mathematical mission in the sensitivity range of an image sensor,
modelling and computational numerical optimization com changing the central wavelength of SFA elements neigh
prises a machine learning method which enables an absence boring the SFA elements with second degree transmis
of transmission measurement by the SFA. sion, as to match the central wavelength of the second
4. The method of claim 2, wherein the mathematical degree transmission,
modelling and computational numerical optimization com re-arranging the SFA element central wavelengths which
prises at least a convex optimization method based on were not changed or do not have a second degree
transmission measurement of the SFA. transmission, as to maintain the uniformity of the SFA
5. The method of claim 1, wherein the step of reconstruct central wavelength distribution, and
ing comprises providing measured transmissions of indi in a 2-dimensional array of pixel intensity values, Sub
vidual SFA elements of the SFA, creating a system design tracting the response corresponding to the changed
matrix from measured transmissions of individual SFA neighboring elements from the response corresponding
elements and image sensor radiometric calibration, and to the SFA elements which have a second degree
transmission.
inferring non sampled spectral information using deconvo
lution or non-linear sparse reconstruction methods based on 11. A method for obtaining a real-time monochromatic
the system design matrix. preview of an imaged object, the imaged object having been
6. The method of claim 1, wherein the step of setting the obtained using the method of obtaining spectral imaging
specific wavelengths of each SFA element of the plurality of data of claim 1, the method for obtaining the real-time
SFA elements on the surface such to obtain the uniform and monochromatic preview comprising the steps of
aperiodic spatial distribution of the specific wavelengths of designing a uniform aperiodic Spectral Filter Array (SFA)
the SFA is deterministic, thereby comprising an aperiodic by arranging a plurality of SFA elements together to
tiling Such as Wang or Penrose. form a surface, configuring each SFA element of the
7. The method of claim 1, wherein the step of setting the plurality of SFA elements to filter a spectral bandwidth
specific wavelengths of each SFA element of the plurality of centered at a central wavelength corresponding to that
SFA elements on the surface such to obtain the uniform and SFA element, whereby all of the plurality of SFA
aperiodic spatial distribution of the specific wavelengths of elements taken together cover a determined spectral
the SFA further comprises range, and setting the central wavelength of each SFA
filling the SFA surface with a repeating pattern, then element of the plurality of SFA elements on the surface
such to obtain a distribution of all the central wave
continuously and randomly inter-changing the SFA ele lengths to be uniform and aperiodic over the Surface,
ments until the measured entropy of the SFA central generating a Subset of SFA elements by a 2-dimensional
wavelengths is Sufficiently high. periodic selection,
8. The method of claim 1, wherein the step of setting the replacing the SFA elements belonging to the above subset
specific wavelengths of each SFA element of the plurality of with SFA elements of identical transmission, and
SFA elements on the surface such to obtain the uniform and from a 2-dimensional array of pixel intensity values,
aperiodic spatial distribution of the specific wavelengths of using a Subset of values, corresponding to the Subset of
the SFA further comprises periodic SFA elements with identical transmission, to
a random sampling of the SFA elements from a uniform create a lower resolution monochromatic image.
distribution. 12. A method for obtaining a real-time monochromatic
9. The method of claim 1, wherein the step of setting the preview of an imaged object, the imaged object having been
specific wavelengths of each SFA element of the plurality of obtained using the method of obtaining spectral imaging
SFA elements on the surface such to obtain the multiband data of claim 1, the method for obtaining the real-time
distribution of the specific wavelengths of the SFA further monochromatic preview comprising the steps of
comprises designing a uniform aperiodic Spectral Filter Array (SFA)
a sampling of multiband SFA elements such that neigh by arranging a plurality of SFA elements together to
boring SFA elements are spectrally orthogonal. form a surface, configuring each SFA element of the
US 2017/0205337 A1 Jul. 20, 2017
plurality of SFA elements to filter multiple spectral using monochromatic previews to optically align the
bands centered at specific wavelengths corresponding images through image registration, thus generating the
to that SFA element, whereby all of the plurality of SFA 3D-model of the scene,
elements taken together cover a determined spectral grouping values from different ones of the arrays of pixel
range, and setting the central wavelength of each SFA intensities corresponding to different images of the
element of the plurality of SFA elements on the surface multiple images, if the pixels are aligned to a same
Such that neighboring SFA elements are spectrally point in the imaged scene,
orthogonal, based on grouped values, re-modeling a system design
from a 2-dimensional array of pixel intensity values, matrix corresponding to multiple arrays of pixel inten
spatially interpolating each channel corresponding to sities, taken from different locations, and
an SFA element configuration independently at pixel reconstructing the spectral texture of the 3D model based
locations where other SFA element configurations are on the cross-image system design matrix.
present, and
averaging the independently interpolated channels, to 14. A non-transitory computer readable medium having a
create a high resolution monochromatic image. computer-executable program stored thereon, the computer
13. A method for generating a spectral 3D-model of an executable program, when executed by a computing device,
imaged object and increasing the spectral reconstruction configured to perform a method according to claim 1 to
quality of an imaged scene, comprising: reconstruct a 3D-hyperspectral image from a 2D-spatial
taking multiple images of the scene, each of the images spectral dataset.
comprising at least an array of pixel intensities,