WO1992005502A1 - Systeme d'analyse spectrale pour extraire des spectres residuels de donnees spectrales - Google Patents

Systeme d'analyse spectrale pour extraire des spectres residuels de donnees spectrales Download PDF

Info

Publication number
WO1992005502A1
WO1992005502A1 PCT/US1991/006952 US9106952W WO9205502A1 WO 1992005502 A1 WO1992005502 A1 WO 1992005502A1 US 9106952 W US9106952 W US 9106952W WO 9205502 A1 WO9205502 A1 WO 9205502A1
Authority
WO
WIPO (PCT)
Prior art keywords
spectrum
data
res
data elements
multispectral
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/US1991/006952
Other languages
English (en)
Inventor
Robert L. Huguenin
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
APPLIED ANALYSIS Inc
Original Assignee
APPLIED ANALYSIS Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by APPLIED ANALYSIS Inc filed Critical APPLIED ANALYSIS Inc
Publication of WO1992005502A1 publication Critical patent/WO1992005502A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01J—MEASUREMENT OF INTENSITY, VELOCITY, SPECTRAL CONTENT, POLARISATION, PHASE OR PULSE CHARACTERISTICS OF INFRARED, VISIBLE OR ULTRAVIOLET LIGHT; COLORIMETRY; RADIATION PYROMETRY
    • G01J3/00—Spectrometry; Spectrophotometry; Monochromators; Measuring colours
    • G01J3/28—Investigating the spectrum
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/40—Analysis of texture
    • G06T7/41—Analysis of texture based on statistical description of texture
    • G06T7/44—Analysis of texture based on statistical description of texture using image operators, e.g. filters, edge density metrics or local histograms
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
    • G01N2021/178—Methods for obtaining spatial resolution of the property being measured
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G06T2207/10036—Multispectral image; Hyperspectral image

Definitions

  • the invention relates to extracting information from low resolution multispectral data.
  • Multispectral analysis has been used for over a decade in the area of laboratory absorption spectroscopy.
  • the reflected light which is visible to humans (“color") is the reflected result, with absorbed light removed.
  • the wavelength positions where the material of interest (referred to hereinafter as the MOI) absorbs light are physical characteristics of the MOI itself, and are not affected by, light level, atmospheric conditions, angle of illumination, angle of observation or environment, unless the environment causes a chemical or physical change in the material properties of the MOI. Materials also emit energy in specific characteristic wavebands when properly stimulated. This is the basis of flame spectroscopy and radio-astronomy.
  • the field of image processing in general has begun to use multispectral analysis.
  • multispectral analysis technigues for analyzing image data to detect the presence of a particular MOI.
  • Such techniques typically involve a search of multispectral image data for a characteristic spectral signature of the MOI.
  • the multi-spectral image data is generally
  • each pixel includes spectral data for more than one wavelength within the energy band of the characteristic spectral signature of the MOI.
  • the multispectral pixel data is tested against the spectral signature of the MOI to identify which pixels contain the characteristic spectral signature, thereby suggesting the presence of the MOI.
  • the ability to successfully detect the presence of the MOI depends upon a variety of factors. These include, among others, whether there is enough of the substance present to produce a sufficiently strong feature; whether there are other materials present with features that can mask (i.e., strongly overlap) the features of the MOI and prevent its detection; and whether the nature and quality of the image data are adequate to resolve the features of the MOI. For
  • the invention is a system for processing multispectral data including a plurality of data elements.
  • the system includes means for identifying for a given data element of the
  • multispectral data an associated subset of the plurality of data elements; and means for extracting some
  • the background spectrum is derived from the associated subset of data elements.
  • the extraction means includes means for determining a correct amount of the background spectrum, where the correct amount is that amount which achieves a best fit between the residual spectrum and the spectrum of a material of interest.
  • the extraction means subtracts the correct amount of the background spectrum from the spectrum for a given element to generate the residual spectrum.
  • the invention is a system for processing multispectral data in preparation for determining whether any of the plurality of data elements of the multispectral data contain a material of interest characterized by a spectrum.
  • the correct amount is that amount which achieves a best fit between the residual spectrum and the spectrum of the material of interest.
  • the multispectral data is image data and each data element corresponds to a different pixel of said image.
  • the background spectrum is derived from a subset of the plurality of data elements, where the subset is associated with the given element.
  • the subset includes data elements that are spatially local to said given element. More
  • the subset is made of all data elements that are contiguous to said given element.
  • the multispectral data is temporal data and each data element corresponds to a different time.
  • the extraction means uses a least squares technique to determine the best fit and it employs the following equation to model the given data element:
  • S CTR [ ] k*S B [ ] + (1-k)*RES[ ], where S CTR [ ] is the multispectral data for the given data element, S B [ ] is the background spectrum, k is a number between 0 and 1 representing the proportion of S CTR accounted for by the background spectrum and RES[ ] is the residual spectrum for the given element.
  • extraction means includes means for computing RES[ ] for a selected value of k; means for computing the sum of the squares of the difference of RES[ ] and the spectrum of the material of interest; and means for changing the selected value of k so as to find the k at which the computed sum of the squares is a minimum.
  • the extraction means includes means for computing RES[ ] for a selected value of k; means for computing a set of ratios from RES[ ] and another set of ratios for the spectrum of the material of interest; means for computing the sum of the squares of the difference of the ratios for RES[ ] and the ratios for the spectrum of the material of interest; and means for changing the selected value of k so as to find the k at which the computed sum of the squares is a minimum.
  • the system also includes means for doping the given data element with the spectrum of the material of interest to generate the spectrum associated with the given element.
  • the invention is a system for processing multispectral data in
  • the system includes means for doping a selected one of the plurality of data elements with the spectrum of the material of interest to produce a doped spectrum; and means for removing a background spectrum from the doped spectrum to generate a residual spectrum, the background spectrum being derived from a subset of data elements of the multispectral data.
  • the multispectral data is image data and each data element corresponds to a different pixel of said image.
  • the background spectrum is derived from a subset of the plurality of data elements, where the subset associated with the given element. The subset includes data elements that are spatially local to said given element. More
  • the subset is made of all data elements that are contiguous to said given element.
  • the multispectral data is temporal data and each data element corresponds to a different time.
  • the background removal means uses a least squares technique to determine the best fit and it employs the following equation to model the given data element:
  • S CTR [ ] K*S B [ ] + (1-k)*RES[ ], where S CTR [ ] is the multispectral data for the given data element, S B [ ] is the background spectrum, k is a number between 0 and 1 representing the proportion of S CTR accounted for by the background spectrum and RES[ ] is the residual spectrum for the given element.
  • RES[ ] for a selected value of k; means for computing the sum of the squares of the difference of RES[ ] and the spectrum of the material of interest; and means for changing the selected value of k so as to find the k at which the computed sum of the squares is a minimum.
  • the background removal includes means for computing RES[ ] for a selected value of k; means for computing a set of ratios from RES[ ] and another set of ratios for the spectrum of the material of interest; means for computing the sum of the squares of the difference of the ratios for RES[ ] and the ratios for the spectrum of the material of interest; and means for changing the selected value of k so as to find the k at which the computed sum of the squares is a minimum.
  • An advantage of the invention is its ability to remove background information so as to unmask any MOI which might be present so that it may more easily be detected. More specifically, the invention has the advantage of being able to determine what background information should be removed and to determine how much of that background should be removed.
  • Another advantage of the invention is its ability to identify the presence and/or location of an MOI in low resolution data with a high degree of reliability.
  • the invention for example, is capable of identifying the presence of the MOI even though it occupies only about 25% of the total area represented by the pixel, i.e., it represents only 25% of the total signal for that pixel.
  • the doping technique provides a way of evaluating how the presence of background alters the spectrum of the MOI, thereby further enhancing the sensitivity of the system in some applications. Using the doping technique, one can discover pixels containing MOI that would have gone undetected without using such a technique.
  • Fig. 1 is a block diagram of an image analysis system that analyzes image data generated by a sensor
  • Fig. 2 shows pixel locations of a representative portion of an image
  • Fig. 3 is the flow diagram describing the operation of the spectral processing module shown in Fig. 1;
  • Figs. 4a and 4b show a flow diagram of the disparate signature detection algorithm that is
  • Fig. 5 is a flow diagram of an algorithm for refining the residual spectrum for the center pixel.
  • Fig. 6 is a flow diagram of the "sanity" check which compares the shape of the residual spectrum to that of the MOI;
  • Fig. 7 is a representative set of residual spectra for doped pixels.
  • Fig. 8 is a representative residual spectrum for use in describing the ratio method of characterizing spectrum shape.
  • a multispectral sensor 4 converts an image of a scene 6 into a multi-plane, nxm pixel array of digital image data 8 that is stored in a memory 10 so that it may be processed at a later time by an image analyzing system 12 to determine whether a material of interest (i.e., MOI) is present in the scene.
  • Each plane of stored image data 8 represents a different point of the spectrum for the image (e.g., a different wavelength of light).
  • the total number of image planes may range from three to many times that amount, the upper limit being primarily established by the limitations of sensor 4.
  • Image analyzing system 12 locates pixels within image data 8 that contain an MOI, which is some substance or material believed to be present in scene 6 and for which the spectral characteristics are known or can be derived. In general, to accomplish this, image analyzing system 12 extracts a sub-pixel constituent from each pixel and then tests the sub-pixel constituent for the presence of the MOI. In the described embodiment, image analyzing system 12 performs the following two functions for each pixel within the image. First, it determines what makes that pixel different from its neighbor pixels, i.e., it determines a residual spectrum. Second, it determines whether the difference has the same spectral characteristics as that of the MOI.
  • a preprocessing module 14 converts stored image data 8 into preprocessed image data 16 that is passed to an spectral processing (SP) module 18. For each pixel of
  • SP module 18 generates a
  • Analyzer 22 may employ any one of a number of known spectrum analysis techniques, such as, for example, any of a variety of classification techniques, principal components analysis, Fourier analysis modeling, or peak onset detection.
  • a doping module 24 selectively dopes one or more pixel locations within image data 8 with varying proportions of MOI.
  • the output of doping module 24 is a set of doped image pixels 26, which, when processed by SP module 18 as actual image pixels, yield a measure of how the presence of other materials (i.e., background
  • processing the doped pixels as actual center pixels yields an allowed range of values for each point of the spectrum of the MOI.
  • the allowed range of values indicate how much the corresponding point of the spectrum for the MOI will vary in the presence of background material.
  • each pixel of image data 8 is represented by at least three points of the energy spectrum.
  • the energy spectrum may be an electrooptical spectrum, a radar cross-section spectrum or any other energy spectrum useful for characterizing scene 6 from which image data 8 is derived. It is desirable that the points of the energy spectrum (e.g., the wavelengths) be selected from the same energy band as defined
  • the same energy band refers to any range of the energy spectrum characterized by a homogeneous set of physical phenomena (e.g., emission or absorption). For example, from ultraviolet to near infra-red (i.e., from about 0.4 microns to about 2.4 microns) the dominating phenomenon is absorption and thus defines an energy band. From about 8 to 12 microns the dominant phenomenon is emission and thus defines another energy band. In contrast, in the intermediate range from about 3 to 5 microns, a mixed phenomenon of absorption and emission occur and thus points selected from this region of energy space do not qualify as being from the same energy band.
  • a homogeneous set of physical phenomena e.g., emission or absorption
  • each center pixel also has a number of neighbor pixels, defined as the pixels
  • a center pixel can have either three, five or eight neighbor pixels, depending upon where in the image the center pixel is located. If the center pixel is at a corner of the image, as is the case for pixel A, then it has three neighbors, namely, pixels B, D, and E. If it is along an edge, as is the case for pixel B, then it has five neighbors, namely, pixels A, C, D, E, and F. If it is not at a boundary of the image, as is the case for pixel E, then it has eight neighbors, namely, pixels A, B, C, D, F, G, H, and I.
  • the size of the MOI is generally smaller than the resolution of a single pixel.
  • preprocessing module 14 is optional.
  • One of its objectives is to account for the ways in which the energy spectrum from scene 6 is modified before it was reduced to image data 8.
  • sensor 4 may include filters having non- uniform transmission characteristics over the energy band of interest. Indeed, the detector used in sensor 4 may itself have non-uniform sensitivity over this band and thus alter the shape of the image spectrum.
  • the intervening media might add spectral information in a predictable and known way, it might have a known non-uniform transmission characteristic over the energy band of interest or it may reflect energy from the sun to add spectral energy to the scene as viewed by sensor 4.
  • CTR' [ ] is the transformed value
  • ⁇ is a generic operator representing one or more of the mathematical operations of addition, subtraction, multiplication and division
  • MODSPEC[ ] is an array of transform values. If, for example, a filter in sensor 4 is being taken into account, then M0DSPEC[ ] is specified by the transmission characteristics of the filter and ⁇ is the multiplication operator. Since image data 6 is preprocessed one pixel at a time, preprocessing module 14 may perform different preprocessing for each pixel, if appropriate.
  • SP module 18 implements the algorithm shown in Fig. 3. For each neighbor of the center pixel that is being processed, SP module 18 first preprocesses the neighbor to generate a preprocessed neighbor value (step 102). This preprocessing step uses the same
  • SP module 18 optionally scales the preprocessed neighbor value to the center pixel (step 104).
  • a preselected one of the points (elements) of the center pixel array is used as a reference element for this scaling step.
  • the element which tends to have the largest magnitude i.e., the brightest element
  • SP module 18 computes a ratio of the reference element of the center pixel to the reference element of the preprocessed neighbor and then multiplies each element of the
  • preprocessed neighbor pixel by the computed ratio to produce a scaled neighbor array having a reference element of the same magnitude as the reference element of the center pixel being processed.
  • the scaling step serves to account for scalar intensity differences between pixels that may be due to intensity variations across scene 6. For example, if image analysis system 12 is being used to locate
  • SP module 18 determines how much of the spectral shape of the center pixel is accounted for by the scaled neighbor (step 106). For this purpose, it runs a disparate signal detection algorithm 200 (see Figs. 4a and 4b) which employs the following model for the center pixel:
  • S CTR [ ] k*S N [ ] + (1-k)*RES[ ]. (1)
  • S CTR [ ] is the value of the center pixel
  • k is a number between zero and one
  • S N [ ] is the scaled neighbor value
  • RES[ ] is the residual value after the
  • the model assumes that the center pixel comprises a fraction k of the spectrum of the neighbor pixel and a residual
  • SP module 18 determines the value of k which achieves a best fit between the residual spectrum (i.e., RES[ ]) and the spectrum of the MOI. If a best fit for a valid k is found, SP module 18 remembers the values for RES[ ], a measure of the fit and k. If a best fit for a valid k is not found, SP module 18 rejects that center pixel
  • SP module 18 determines whether there any more neighbor pixels for the center pixel (step 108). If there are, SP module 18 selects the next neighbor (step 110) and returns to step 102 and computes a residual spectrum (RES n [ ], where n identifies the neighbor pixel) using the next neighbor pixel. After all neighbors have been processed for that center pixel, SP module 18 selects the RES n [ ] which yielded the best fit (step 112). If the process did not find any neighbors which yielded a best fit for a valid k, it rejects that center pixel as a candidate for the presence of MOI.
  • RES n [ ] residual spectrum
  • SP module 18 optionally performs a "sanity" check on the resulting RES[ ] to make sure that it has a shape similar to that of the spectrum for the MOI (step 114) . If the shape is not similar, the center pixel is rejected. If the shape passes the "sanity" check", SP module 18 performs a second screening test to determine whether the points of RES[ ] fall within the allowed range of values established by the doped pixels (step 116). Again, if any of the points fail this allowed- range-of-values test, that center pixel is rejected. If all points pass this test, that center pixel is passed to analyzer 22 to determine whether RES[ ] represents MOI.
  • DSD algorithm 200 uses an iterative technique to arrive at the best fit.
  • the value of k is incremented one step at a time by an amount K STEP between a lower value, K MIN , and an upper value, K MAX .
  • the fit between the estimated residual spectrum and the spectrum of the MOI is determined based upon a sum of the squares of the differences of the two spectra. The fit is then compared to the fit computed for the previous step to arrive at a minimum least squares value, if one exists.
  • SP module 18 initializes three variables, namely K LOWER , K UPPER , and K FINAL (step 202).
  • K LOWER is set equal to K MIN
  • K UPPER is set equal to K MAX
  • K FINAL is set equal to K LOWER .
  • K MIN and K MAX define the range of permissible values for k and are set externally by the user. For example, they may equal 0 and 1, respectively.
  • SP module 18 sets the initial value of K STEp equal to 0.1*[K MAX - K MIN ] .
  • SP module 18 determines whether K STEP is greater than a threshold value, which in the described embodiment is equal to 10 -5 (step 206). At this point in the algorithm, K STEP is greater than the threshold, so SP module 18 sets a variable BESTFIT equal to a large number (e.g., 10 30 ) and sets k, the weighting factor representing the
  • BESTFIT is a variable that takes on the best fit value computed by algorithm 200. As will become more apparent, BESTFIT is initially set to a high number to assure that it is larger than any conceivable value that might be generated for the first computed fit value, and thus BESTFIT will take on the first computed fit value after the first pass through algorithm 200.
  • SP module 18 compares k to K UPPER (step 210). If k is not greater than K UPPER , SP module 18 computes the residual spectrum, i.e., RES[ ], for the center pixel in accordance with the following equation (step 212): (2 )
  • RES[ ] is scaled to the spectrum for the MOI, i.e., MOI[ ], using the same principles as were previously employed to scale the preprocessed neighbor to the center pixel (step 214).
  • the output of step 214 is a scaled residual array designated as RES'[ ].
  • SP module 18 determines whether RES'[ ] passes a minimal test for a valid spectrum. More
  • SP module 18 makes sure that every element of the computed RES'[ ] is greater than zero (step 216). If any element of RES'[ ] is less than or equal to zero, the residual spectrum has no meaningfully significant physical interpretation and thus, that particular center pixel, neighbor pixel combination is rejected.
  • SP module 18 computes the fit between RES'[ ] and the spectrum of the MOI, i.e., MOI[ ], and sets a variable FIT equal to the result (step 218).
  • SP module 18 uses the sum of the squares of the differences to generate a measure of the fit. The equation is:
  • FIT ⁇ i (RES'[i] - MOI[i]) 2 .
  • SP module 18 compares it to BESTFIT (step 220). If FIT is less than BESTFIT (as it most certainly will be for the first time through the algorithm), BESTFIT is set equal to FIT (step 222). SP module 18 also sets K FINAL equal to the current value for k (step 224) and then increments the value of k for the next pass through the inner loop (step 226).
  • SP module 18 branches back to step 210 to repeat the above described sequence of steps for the next value of k.
  • step 220 for some incremented value of k, the computed FIT value will not be less than BEST FIT.
  • SP module 18 determines whether the computed FIT value equals BESTFIT (step 228). If the two values are equal, SP module 18 generates a new value for K FINAL that is the average of the current value of k and the current value of K FINAL (step.230). Then, SP module 18 sets K LOWER equal to K FINAL minus the current value for K STEP' thereby establishing a new lower bound for the next pass through the outer loop using a smaller step size.
  • step 2208 if FIT does not equal BESTFIT, SP module 18 branches directly to step 232 to establish a new value for K LOWER .
  • K LOWER is compared to K MIN to make sure that it is not outside the range of permissible values for k (step 234). If K LOWER is less than K MIN (i.e., it is outside the permissible range), SP module 18 set K LOWER equal to K MIN thereby moving it back into the permissible range (step 236).
  • K UPPER is established. SP module 18 sets K UPPER equal to K FINAL plus the current value for K STEP (step 238). Then, K UPPER is compared to K MAX to make sure that it is not outside the range of permissible values for k (step 240). If K UPPER is greater than K MAX (i.e., it is outside the permissible range), SP module 18 set K UPPER equal to K MAX thereby moving it back into the permissible range (step 242).
  • SP module 18 After the value for K UPPER has been established, SP module 18 reduces the step size to 10% of its current value (step 244) and then branches back to step 206 to repeat the outer loop for the new step size. Note that for the next series of iterations of the inner loop (i.e, incrementing k by a fixed step size), Sp module 18 has redefined K LOWER and K UPPER so that the search range brackets the k value at which the current BEST FIT was found.
  • SP module 18 will detect in step 206 that K STEP is no longer greater than the threshold. In that case, SP module 18 determines whether an actual minimum has been found. First, SP module 18 determines whether K FINAL equals K MIN (i.e., zero) (step 246). Note that it is possible for the RES'[ ] curve to be monotonically increasing over the range of permissible values for k, in which case, it does not have a minimum (i.e., a best fit) within that range. In that event, K FINAL will remain pinned at K MIN . If that is the case, SP module 18 rejects that particular center pixel, neighbor pixel combination.
  • SP module 18 determines whether K FINAL equals K MAX (i.e., one) (step 248). Note that it is also possible for the RES'[ ] curve to be monotonically decreasing over the range of permissible values for k, in which case, it again does not have a minimum within that range. In that event, K FINAL w ill become pinned at K MAX . If that is the case, SP module 18 rejects that particular center pixel, neighbor pixel combination.
  • SP module 18 computes RES[ ] for k equal to K FINAL (step 250). Then, it scales RES[ ] to MOI[ ] (as previously described for step 214) to generate RES'[ ] and outputs RES'[ ], K FINAL , and BESTFIT (step 252).
  • DSD algorithm 200 shown in Figs. 4a and 4b is run for each pixel in the
  • the result is a list of (FIT) N values and their associated residual spectra, RES N [ ]. If there are eight neighbors, this list may have anywhere from zero entries, if no best fit was returned for any center pixel, neighbor pixel combination, to eight entries, if all combinations returned a best fit. The best of the returned BEST FIT's is selected to characterize the center pixel.
  • a residual spectrum can be derived for the center pixel based upon contributions from more than one neighbor.
  • a refine algorithm 300 which
  • Fig. 5 the valid results of the analysis of each center pixel, neighbor pixel combination (i.e., the fit value and the residual spectrum for each neighbor that returns a valid k) are stored in a list (step 302). After all center pixel, neighbor pixel combinations have been examined, the resulting list of FIT values and associated RES[ ] is sorted in ascending order according to FIT value, the best fit first (step 304). Then, a check is made to determine whether the list contains any entries (step 306). If it does not, algorithm 300 is terminated.
  • the neighbor yielding the best fit value i.e., the first entry in the list
  • its computed FIT value, k, and RES[ ] are used as the current fit value, k, and residual spectrum, respectively, for the center pixel (step 308).
  • algorithm 300 determines whether there is a next best entry in the sorted list of fit values (step 310). If no next best entry exists, algorithm 300 terminates and the resulting residual spectrum for the center pixel is used for the subsequent determination of whether MOI is present. If a nest best entry exists, algorithm 300 computes a new RES[ ], k, and FIT value for the center pixel and that next best entry neighbor (step 312). It does this by invoking DSD algorithm 200 shown in Figs. 4a and 4b. Next, algorithm 300 determines whether DSD
  • algorithm 200 has returned a valid k value (step 314). If no valid k value was found, algorithm 300 rejects that, particular neighbor pixel and branches back to step 310 to identify the next best fit in the sorted list. If a valid k was found, algorithm 300 compares the computed FIT value associated with that neighbor pixel to the current FIT value for the center pixel (step 316). If the computed FIT value is not better than the current FIT value, algorithm 300 rejects that particular neighbor pixel and branches back to step 310 to examine the rest of the sorted list of FIT values.
  • algorithm 300 updates the current FIT value to equal the computed FIT value, sets the residual spectrum of the center pixel equal to the computed RES[ ] for that
  • step 318 After updating the values for the center pixel, algorithm 300 branches back to step 310 and continues this process down the remainder of the sorted list. At the conclusion of algorithm 300, i.e., after it has found and processed the last entry in the sorted list, the resulting residual spectrum is then passed to analyzer 22 (see Fig. 1) to determine whether MOI is present.
  • the “sanity” check is performed on the resulting residual spectrum for the center pixel to determine whether further testing against the spectrum of the MOI is warranted.
  • the “sanity” check compares the shape of the residual spectrum with the shape of the spectrum of the MOI spectrum.
  • SP module 18 does the shape comparison on a pair-by-pair basis, that is, by comparing two points at a time of the
  • a "sanity" check algorithm 400 for carrying out this comparison is shown in Fig. 6.
  • SP module 18 sets p, an integer index identifying the points of the spectrum, equal to one (step 402).
  • Index p can take on values from 0 to P max -1, where P max is the total number of points or image planes of the multispectral image data.
  • SP module 18 computes MOI[p] - MOI[p-1] and sets a variable A equal to the result (step 404).
  • A is divided by MOI[p] and the resulting ratio is compared to 0.10 (step 406). If the computed ratio is greater than a preselected
  • SP module 18 computes RES[p] - RES[p-1] and sets a variable B equal the result (step 408). Then, SP module 18 compares the sign of A to the sign of B (step 410). If the signs are different, that center pixel is rejected as a candidate for the presence of MOI. If the signs are the same, SP module 18
  • step 412 determines whether there are any more points of the spectrum to be analyzed (i.e., does p equal P max -1?) (step 412). If there are more points, SP module 18 moves to the next point (step 414) and returns to step 404 to perform the slope comparison for that point.
  • step 406 if the computed ratio is not greater than 0.01, a slope comparison is not done for those two points. Rather, SP module 18 moves onto the next point if a next point exists.
  • SP module 18 When SP module 18 reaches the last point of the spectrum and assuming all points have passed the slope comparison test, it detects that p equals P max -1 in step 410 and reports that the residual spectrum for that center pixel passes the "sanity" check.
  • the step performed by doping module 24 is
  • doping module 24 adds in varying amounts of MOI, i.e., doping module 24 dopes background image with MOI.
  • the number of pixels selected for the doping operation is variable. For small images, it may be preferable to perform this step on all pixels of the image. On the other hand, for large images, it may be preferable to select only a
  • Doping module 24 generates a doped pixel value S D [ ] according to the following equation:
  • d is the doping level and has a value between zero and one, inclusive
  • S B [ ] is the value of the pixel from the image
  • MOI[ ] is the spectrum of the MOI.
  • This may be computed for only one value of m selected from the range of permissible values (e.g., 0.15) or it may be computed for a set of values of d.
  • choosing more than one value of m provides more information about how the background alters the spectrum of the MOI and may be useful in selecting the decision threshold so as to reduce the likelihood of false detections. For example, by selecting a range of values of d, one may learn that clustering occurs above a certain value of d. Thus, to reduce the possibility of false detections the detection criteria can be selected over the more limited cluster ranges.
  • the set of doped pixels within the image are processed as center pixels, using the same procedures as outlined above. That is, the system attempts to identify a valid k for the doped pixel in combination with each of its neighbor pixels. Note that doped pixels are only used as center pixels and not used as neighbor pixels. The neighbor pixels are derived from the image as they were for the actual center pixels. This follows from the purpose of doped pixels, namely, to determine how
  • step 114 of Fig. 3 For each center pixel that passes the shape test (step 114 of Fig. 3), the second screening test (step 114 of Fig. 3).
  • SP module 18 determines whether each point of the residual spectrum falls within the allowed range of values
  • analyzer 22 uses the computed FIT value to decide whether the resulting residual spectrum for center pixels which pass all previous tests represents MOI. For example, any FIT value which is less than a predetermined level returns an affirmative indication as to the presence of the MOI.
  • the shape ratios are defined as the ratios of the intensities in the different image planes. This can be more easily understood with the aid of Fig. 8, which shows a representative spectrum of five points of a residual spectrum, labelled I through V.
  • One set of ratios describing the shape of this spectrum might be all or some subset of the following ratios: (V/I), (V/II), (V/III), (V/IV), (IV/I), (IV/II), (IV/III), (III/I), (III/II), (II/I).
  • the ratios rather than comparing the actual residual spectrum to the MOI spectrum, the ratios
  • the neighborhood from which the "neighbor" pixels are selected need no be limited to pixels that are contiguous to the center pixel nor need they even be limited to an area that is local to the center pixel. It may be desirable under some
  • the image data could be a sequence of blood analyses taken over an extended period of time and the objective is to find the presence of a substance (i.e, an MOI) in the blood samples.
  • a substance i.e, an MOI
  • the above-described methods can be used to extract background information from the samples to isolate a residual spectrum that can then be compared to the spectrum of the MOI.

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Theoretical Computer Science (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)

Abstract

Système (12) servant à traiter des données multispectrales composées d'une pluralité d'éléments de données. Ledit système comprend un dispositif (14) permettant d'identifier un sous-ensemble associé de la pluralité des éléments de données pour un élément de données déterminé des données multispectrales; un dispositif (18) pour extraire un certain pourcentage d'un spectre d'arrière-plan à partir d'un spectre associé avec l'élément de données déterminé pour générer un spectre résiduel, le spectre d'arrière-plan étant dérivé du sous-ensemble d'éléments de données associé.
PCT/US1991/006952 1990-09-25 1991-09-24 Systeme d'analyse spectrale pour extraire des spectres residuels de donnees spectrales Ceased WO1992005502A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US58800090A 1990-09-25 1990-09-25
US588,000 1990-09-25

Publications (1)

Publication Number Publication Date
WO1992005502A1 true WO1992005502A1 (fr) 1992-04-02

Family

ID=24352044

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US1991/006952 Ceased WO1992005502A1 (fr) 1990-09-25 1991-09-24 Systeme d'analyse spectrale pour extraire des spectres residuels de donnees spectrales

Country Status (1)

Country Link
WO (1) WO1992005502A1 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0604124A3 (fr) * 1992-12-17 1995-01-18 Trw Inc Technique d'extraction de signature multispectrale.

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4300833A (en) * 1979-10-26 1981-11-17 The United States Of America As Represented By The Secretary Of Agriculture Method for background corrected simultaneous multielement atomic absorption analysis
US4660151A (en) * 1983-09-19 1987-04-21 Beckman Instruments, Inc. Multicomponent quantitative analytical method and apparatus

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4300833A (en) * 1979-10-26 1981-11-17 The United States Of America As Represented By The Secretary Of Agriculture Method for background corrected simultaneous multielement atomic absorption analysis
US4660151A (en) * 1983-09-19 1987-04-21 Beckman Instruments, Inc. Multicomponent quantitative analytical method and apparatus

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ANALYTICAL CHEMISTRY, Vol. 48, No. 3, March 1976, (J.D. GANJEI et al.), "Multielement Atomic Spectrometry with a computerized vidicon detector", pages 505-510. *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0604124A3 (fr) * 1992-12-17 1995-01-18 Trw Inc Technique d'extraction de signature multispectrale.
US5479255A (en) * 1992-12-17 1995-12-26 Trw Inc. Multispectral signature extraction technique

Similar Documents

Publication Publication Date Title
Briottet et al. Military applications of hyperspectral imagery
Gu et al. A selective KPCA algorithm based on high-order statistics for anomaly detection in hyperspectral imagery
US6208752B1 (en) System for eliminating or reducing exemplar effects in multispectral or hyperspectral sensors
Uezato et al. A novel endmember bundle extraction and clustering approach for capturing spectral variability within endmember classes
CN109858477A (zh) 用深度森林在复杂环境中识别目标物的拉曼光谱分析方法
WO1999045497A1 (fr) Systeme d'elimination ou de reduction d'effets de bruit typique (serene)
Lopez et al. A new preprocessing technique for fast hyperspectral endmember extraction
RU2695963C1 (ru) Способ распознавания образов биохимических наземных объектов с тонкими спектральными различиями на основе гиперспектральных данных
Zhang et al. Target identification and detection using hyperpectral signature transformation
Vanzella et al. Multicolor Observations of the Hubble Deep Field South
CN118212481A (zh) 一种基于深度学习的高光谱图像目标智能检测方法
WO1992005502A1 (fr) Systeme d'analyse spectrale pour extraire des spectres residuels de donnees spectrales
CN109682821A (zh) 一种基于多尺度高斯函数的柑橘表面缺陷检测方法
McCann et al. Novel histogram based unsupervised classification technique to determine natural classes from biophysically relevant fit parameters to hyperspectral data
Kasen et al. Target detection in hyperspectral images based on multicomponent statistical models for representation of background clutter
Drory Yet another object detection application (YODA)-Object detection and photometry for multi-band imaging data
Brunzendorf et al. A QSO survey via optical variability and zero proper motion in the M 92 field-IV. More QSOs due to improved photometry
Ientilucci Hyperspectral sub-pixel target detection using hybrid algorithms and Physics Based Modeling
Riley et al. Anomaly detection in noisy hyperspectral imagery
CN118443599B (zh) 一种面向新鲜鸡肉的品质光谱检测方法、设备及系统
Tsang et al. Evaluation of endmember selection techniques and performance results from ORASIS hyperspectral analysis
de Souza et al. SAGUI: SED-based Segmentation of Multi-band Galaxy Images--Application to JADES in GOODS-South
Younis Investigation of Hyperspectral Data Unmixing and Its Impact on the Task of Anomaly Detection
Friesen et al. Contextual anomaly detection cueing methods for hyperspectral target recognition
Hoff et al. Enhanced classification performance from multiband infrared imagery

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A1

Designated state(s): CA JP SU

AL Designated countries for regional patents

Kind code of ref document: A1

Designated state(s): AT BE CH DE DK ES FR GB GR IT LU NL SE

122 Ep: pct application non-entry in european phase
NENP Non-entry into the national phase

Ref country code: CA