CN108009550B - Hyperspectral image characteristic detection method and device based on spectral curve fitting - Google Patents
Hyperspectral image characteristic detection method and device based on spectral curve fitting Download PDFInfo
- Publication number
- CN108009550B CN108009550B CN201711097228.7A CN201711097228A CN108009550B CN 108009550 B CN108009550 B CN 108009550B CN 201711097228 A CN201711097228 A CN 201711097228A CN 108009550 B CN108009550 B CN 108009550B
- Authority
- CN
- China
- Prior art keywords
- phi
- point
- lambda
- fitting
- hyperspectral image
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
Abstract
The invention is suitable for the technical field of characteristic point detection, and provides a hyperspectral image characteristic detection method and device based on spectral curve fitting, wherein the hyperspectral image characteristic detection method and device comprise the following steps: fitting the hyperspectral image f (x, y, lambda) in the spectral direction to obtain a new hyperspectral image phi (x, y, lambda'); constructing a new point p in the hyperspectral image phi (x, y, lambda')0And point p in its neighborhood1A weighted correlation function of; constructing a characteristic point response function according to the weighted correlation function; calculating a certain point p in the hyperspectral image phi (x, y, lambda') according to the characteristic point response function0The response value of the characteristic point of the point and the response values of the characteristic points of all the points in the neighborhood of the point; if a certain point p in the hyperspectral image phi (x, y, lambda')/is0If the response value of the feature point is greater than the response values of the feature points of all the points in the neighborhood, the point p is determined0The characteristic points of the hyperspectral image phi (x, y, lambda') are obtained; the method provided by the invention can be used for carrying out data compression and noise reduction on the whole hyperspectral image by a fitting method in the spectral direction, improves the operating efficiency and realizes the purpose of carrying out feature point detection on the hyperspectral image on the basis.
Description
Technical Field
The invention belongs to the technical field of characteristic point detection, and particularly relates to a hyperspectral image characteristic detection method and device based on spectral curve fitting.
Background
Aiming at a common image, such as a two-dimensional image like a gray image or a color image, algorithms for extracting feature points are spot algorithms SIFT, SURF, corner algorithms Harris, FAST, BRISK and the like, and the research on feature point detection algorithms of two-dimensional images is very mature.
Compared with the traditional gray-scale image, the hyperspectral image not only contains spatial information, but also contains information of the other aspect, namely spectral response information of an object, and the hyperspectral image organically combines the spectral information reflecting the radiation attribute of a substance with two-dimensional image information reflecting the spatial geometrical relationship of the object, so that the hyperspectral image can provide more information than the gray-scale image and the color image. However, for a three-dimensional data structure such as a hyperspectral image, a local feature detection method of a two-dimensional image is not suitable for the hyperspectral image; for example, a feature point detection algorithm of a two-dimensional image, such as a common Harris corner detection operator, can only act on a gray image or a color image, and cannot directly act on hyperspectral image data. On one hand, the amount of the hyperspectral image data is generally large, and especially the wave bands in the direction of the hyperspectral image spectrum are generally many, so that the processing efficiency is low; on the other hand, spectral curves are susceptible to noise interference, making the curves spiky.
Therefore, a method for detecting the feature points of the hyperspectral image is needed, and the method can solve the problems of low data processing efficiency and noise interference of the hyperspectral image.
Disclosure of Invention
The invention provides a hyperspectral image feature detection method and device based on spectral curve fitting, and aims to provide a method for performing data compression and noise reduction on the whole hyperspectral image by using a fitting method in a spectral direction and performing feature point detection on the hyperspectral image on the basis.
The invention provides a hyperspectral image feature detection method based on spectral curve fitting, which comprises the following steps:
step S1, fitting the hyperspectral image f (x, y, lambda) in the n order in the spectral direction to obtain a new hyperspectral image phi (x, y, lambda');
wherein x and y represent space domain coordinates, and lambda represents spectrum domain coordinates, wherein lambda is more than or equal to 1 and less than or equal to L, and lambda' is more than or equal to 1 and less than or equal to n + 1; λ represents a band index value before fitting, λ' represents a band index value after fitting, L represents the maximum band number before fitting, and n +1 represents the maximum band number after fitting;
step (ii) ofS2, constructing a new hyperspectral image phi (x, y, lambda') about a certain point p0And point p in its neighborhood1A weighted correlation function of;
step S3, constructing a characteristic point response function according to the weighted correlation function;
step S4, calculating a certain point p in the hyperspectral image phi (x, y, lambda') according to the characteristic point response function0The response value of the characteristic point of the point and the response values of the characteristic points of all the points in the neighborhood of the point;
step S5, if a certain point p in the hyperspectral image phi (x, y, lambda')0If the response value of the feature point is greater than the response values of the feature points of all the points in the neighborhood, the point p is determined0Namely the characteristic points of the hyperspectral image phi (x, y, lambda').
Further, the step S1 specifically includes:
step S11, constructing a spectral curve expression for all the band components with spatial coordinates (x, y) contained in the hyperspectral image f (x, y, λ) as follows:
g(x,y)＝[f(x,y,1),f(x,y,2),...,f(x,y,λ),...,f(x,y,L)]
wherein, λ is more than or equal to 1 and less than or equal to L, λ represents a band index value before fitting, L represents the maximum band number before fitting, and g (x, y) is a spectral curve expression of any spatial point before fitting;
step S12, for a specific point (x) in space in the spectral direction according to the above formula0,y0) G (x) of0,y0) Performing polynomial fitting of each term f (x)0,y0λ) the fitting result is:
wherein x is0,y0Is a constant, phi (x)0,y0,1),φ(x0,y0,2),φ(x0,y0,λ′),φ(x0,y0,n),φ(x0,y0N +1) are coefficients of the respective fitted terms, λ' is a coefficient index value which is an integer and1≤λ′≤n+1，
is to f (x)0,y0λ) fitting estimates;
step S13, extracting coefficients of the polynomial in the fitting result to reconstruct a new spectral curve:
G(x0,y0)＝[φ(x0,y0,1),φ(x0,y0,2),...,φ(x0,y0,λ′),...,φ(x0,y0,n+1)]
wherein λ' is a new band index value; g (x)0,y0) For the fitted specific point (x) in space0,y0) The spectral curve expression of (a);
step S14, performing curve fitting on a spectrum curve g (x, y) formed by the pixels (x, y) at any position of the whole hyperspectral image by combining the new spectrum curve to obtain:
G(x,y)＝[φ(x,y,1),φ(x,y,2),...,φ(x,y,λ′),...,φ(x,y,n+1)]
wherein phi (x, y, lambda ') is a new hyperspectral image expression, lambda ' is more than or equal to 1 and less than or equal to n +1, and lambda ' is a new waveband index value; n +1 represents the maximum number of bands after fitting, and G (x, y) is the spectral curve expression of the arbitrary point in space after fitting.
Further, the weighted correlation function is:
wherein, the point p0Is a pixel in the hyperspectral image phi (x, y, lambda ') with coordinates (x, y, lambda ') and phi (x, y, lambda ') as a point p0The DN value of the corresponding hyperspectral image; point p1Coordinates (x +. DELTA.x, y +. DELTA.y, lambda '+. DELTA.lambda) and phi (x +. DELTA.x, y +. DELTA.y, lambda' +. DELTA.lambda) as a point p1The corresponding DN value;
the window function ω (x, y, λ') employs a gaussian weighting function, as follows:
wherein, sigma is a scale factor of a Gaussian function;
wherein the content of the first and second substances,
for the convolution operation symbol, l is the length of the window function moving in the x direction, m is the length of the window function moving in the y direction, r is the length of the window function moving in the λ direction, l is 1, m is 1, r is 1, i.e., the window size is 3 × 3.
Further, in the weighted correlation function
Is shown as
Namely:
and, instead,
then the process of the first step is carried out,
wherein the content of the first and second substances,
in the formula, phix,φy,φλ′Respectively, the gradient of the image phi (x, y, lambda ') in three directions x, y, lambda', namely:
in the above formula, ω represents a Gaussian weighting function ω (x, y, λ'),
for the convolution symbols A, B, C, D, E, F correspond to each element of the matrix M, phi, respectivelyx 2,φy 2,φλ′ 2Respectively represents the gradient phi of the multispectral image in three directions of x, y and lambdax,φy,φλ′Square of (d), phixφyIs indicative of phixPhi and phiyProduct of (a), phiyφλ′Is indicative of phiyPhi and phiλ′Product of (a), phixφλ′Is indicative of phixPhi and phiλ′A, B, C, D, E, F correspond to the elements of the matrix M, respectively.
Further, the characteristic point response function is:
R＝det(M)-k(trace(M))3＝(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3；
wherein k is 0.001, and k is an empirical constant; det (M) represents the determinant of matrix M, trace (M) represents the traces of matrix M, and the expression is as follows:
det(M)＝λ1λ2λ3＝ABC+2DEF-BE2-AF2-CD2
trace(M)＝λ1+λ2+λ3＝A+B+C
wherein λ is1、λ2、λ3Respectively, the eigenvalues of the matrix M.
The invention also provides a hyperspectral image feature detection device based on spectral curve fitting, which comprises:
the fitting module is used for performing n-order fitting on the hyperspectral image f (x, y, lambda) in the spectral direction to obtain a new hyperspectral image phi (x, y, lambda');
wherein x and y represent space domain coordinates, and lambda represents spectrum domain coordinates, wherein lambda is more than or equal to 1 and less than or equal to L, and lambda' is more than or equal to 1 and less than or equal to n + 1; λ represents a band index value before fitting, λ' represents a band index value after fitting, L represents the maximum band number before fitting, and n +1 represents the maximum band number after fitting;
a weighted correlation function construction module for constructing a new hyperspectral image phi (x, y, lambda') about a certain point p0And point p in its neighborhood1A weighted correlation function of;
the characteristic point response function constructing module is used for constructing a characteristic point response function according to the weighted correlation function;
a characteristic point response value calculation module for calculating a certain point p in the hyperspectral image phi (x, y, lambda') according to the characteristic point response function0The response value of the characteristic point of the point and the response values of the characteristic points of all the points in the neighborhood of the point;
a characteristic point judging module for judging a certain point p in the hyperspectral image phi (x, y, lambda')0When the response value of the characteristic point is larger than the response values of the characteristic points of all the points in the neighborhood, the point p is judged0Namely the characteristic points of the hyperspectral image phi (x, y, lambda').
Further, the fitting module includes:
a first construction submodule, configured to construct a spectral curve expression for all band components with spatial coordinates (x, y) contained in the hyperspectral image f (x, y, λ) as follows:
g(x,y)＝[f(x,y,1),f(x,y,2),...,f(x,y,λ),...,f(x,y,L)]
wherein, λ is more than or equal to 1 and less than or equal to L, λ represents a band index value before fitting, L represents the maximum band number before fitting, and g (x, y) is a spectral curve expression of any spatial point before fitting;
a first fitting submodule for fitting a specific point (x) in space in the spectral direction according to the above formula0,y0) G (x) of0,y0) Performing polynomial fitting of each term f (x)0,y0λ) the fitting result is:
wherein x is0,y0Is a constant, phi (x)0,y0,1),φ(x0,y0,2),φ(x0,y0,λ′),φ(x0,y0,n),φ(x0,y0N +1) are each corresponding term coefficient of the fitting, λ 'is a coefficient index value which is an integer and λ' is not less than 1 and not more than n +1,
is to f (x)0,y0λ) fitting estimates;
a second construction submodule, configured to extract coefficients of the polynomial in the fitting result to reconstruct a new spectral curve:
G(x0,y0)＝[φ(x0,y0,1),φ(x0,y0,2),...,φ(x0,y0,λ′),...,φ(x0,y0,n+1)]
wherein λ' is a new band index value; g (x)0,y0) For the fitted specific point (x) in space0,y0) The spectral curve expression of (a);
and the second fitting submodule is used for performing curve fitting on a spectral curve g (x, y) formed by pixels (x, y) at any position of the whole hyperspectral image by combining the new spectral curve to obtain:
G(x,y)＝[φ(x,y,1),φ(x,y,2),...,φ(x,y,λ′),...,φ(x,y,n+1)]
wherein phi (x, y, lambda ') is a new hyperspectral image expression, lambda ' is more than or equal to 1 and less than or equal to n +1, and lambda ' is a new waveband index value; n +1 represents the maximum number of bands after fitting, and G (x, y) is the spectral curve expression of the arbitrary point in space after fitting.
Further, the weighted correlation function is:
wherein, the point p0Is a pixel in the hyperspectral image phi (x, y, lambda ') with coordinates (x, y, lambda ') and phi (x, y, lambda ') as a point p0The DN value of the corresponding hyperspectral image; point p1Coordinates (x +. DELTA.x, y +. DELTA.y, lambda '+. DELTA.lambda) and phi (x +. DELTA.x, y +. DELTA.y, lambda' +. DELTA.lambda) as a point p1The corresponding DN value;
the window function ω (x, y, λ') employs a gaussian weighting function, as follows:
wherein, sigma is a scale factor of a Gaussian function;
wherein the content of the first and second substances,
for the convolution operation symbol, l is the length of the window function moving in the x direction, m is the length of the window function moving in the y direction, r is the length of the window function moving in the λ direction, l is 1, m is 1, r is 1, i.e., the window size is 3 × 3.
Further, in the weighted correlation function
Is shown as
Namely:
and, instead,
then the process of the first step is carried out,
wherein the content of the first and second substances,
in the formula, phix,φy,φλ′Respectively, the gradient of the image phi (x, y, lambda ') in three directions x, y, lambda', namely:
in the above formula, ω represents a Gaussian weighting function ω (x, y, λ'),
for the convolution symbols A, B, C, D, E, F correspond to each element of the matrix M, phi, respectivelyx 2,φy 2,φλ′ 2Respectively represents the gradient phi of the multispectral image in three directions of x, y and lambdax,φy,φλ′Square of (d), phixφyIs indicative of phixPhi and phiyProduct of (a), phiyφλ′Is indicative of phiyPhi and phiλ′Product of (a), phixφλ′Is indicative of phixPhi and phiλ′A, B, C, D, E, F are respectively pairedThe individual elements of matrix M.
Further, the characteristic point response function is:
R＝det(M)-k(trace(M))3＝(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3；
wherein k is 0.001, and k is an empirical constant; det (M) represents the determinant of matrix M, trace (M) represents the traces of matrix M, and the expression is as follows:
det(M)＝λ1λ2λ3＝ABC+2DEF-BE2-AF2-CD2
trace(M)＝λ1+λ2+λ3＝A+B+C
wherein λ is1、λ2、λ3Respectively, the eigenvalues of the matrix M.
Compared with the prior art, the invention has the beneficial effects that: according to the hyperspectral image feature detection method and device based on spectral curve fitting, n-order fitting is carried out on a hyperspectral image f (x, y, lambda) in the spectral direction to obtain a new hyperspectral image phi (x, y, lambda'); constructing a point p in relation to the hyperspectral image phi (x, y, lambda')0And point p in its neighborhood1A weighted correlation function of; constructing a characteristic point response function according to the weighted correlation function; calculating a certain point p in the hyperspectral image phi (x, y, lambda') according to the characteristic point response function0The response value of the characteristic point of the point and the response values of the characteristic points of all the points in the neighborhood of the point; if a certain point p in the hyperspectral image phi (x, y, lambda')/is0If the response value of the feature point is greater than the response values of the feature points of all the points in the neighborhood, the point p is determined0The characteristic points of the hyperspectral image phi (x, y, lambda') are obtained; compared with the prior art, the method can perform data compression and noise reduction on the whole hyperspectral image by a fitting method in the spectral direction, reduce the data volume to be processed, improve the operating efficiency, retain the spectral curve characteristics of substances and achieve the purpose of noise reduction; further, feature point detection can be carried out on the hyperspectral image on the basis, and key information of the hyperspectral image is acquired, so that the hyperspectral image can be better detectedAnd the image is analyzed, so that the hyperspectral image pattern recognition effect is improved.
Drawings
FIG. 1 is a schematic flow chart of a hyperspectral image feature detection method based on spectral curve fitting according to an embodiment of the invention;
FIG. 2 is a schematic diagram showing a comparison between a raw spectral curve and spectral curves obtained by performing first-order, second-order and third-order simulations, respectively, according to an embodiment of the present invention;
fig. 3 is a schematic block diagram of a hyperspectral image feature detection apparatus based on spectral curve fitting according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The technical problems that the characteristic point detection cannot be carried out on the hyperspectral image and the low data processing efficiency and the noise interference of the hyperspectral image cannot be solved exist in the prior art.
In order to solve the technical problems, the invention provides a hyperspectral image feature detection method which is a brand new three-dimensional detection operator, performs data compression and noise reduction on the whole image by a method of fitting in the spectral direction, and performs hyperspectral image feature point detection on the basis of the method.
Referring to fig. 1, a hyperspectral image feature detection method based on spectral curve fitting provided by an embodiment of the invention includes:
step S1, fitting the hyperspectral image f (x, y, lambda) in the n order in the spectral direction to obtain a new hyperspectral image phi (x, y, lambda');
specifically, let F be a hyperspectral image, and the image size be M × N × L, then it can be expressed as:
F＝f(x,y,λ) (1)
in the formula, f (x, y, lambda) represents a function of the hyperspectral image, namely, represents a DN (Digital Number, remote sensing image pixel brightness) value at a point (x, y, lambda), (x, y, lambda) represents a 3-dimensional coordinate, x, y, lambda are integers, x and y represent space domain coordinates, x is larger than or equal to 1 and smaller than or equal to M, y is larger than or equal to 1 and smaller than or equal to N, lambda represents a spectral domain coordinate, lambda is larger than or equal to 1 and smaller than or equal to L, and L represents the maximum Number of wave bands in the spectral direction of the hyperspectral image.
Specifically, the step S1 includes:
step S11, constructing a spectral curve expression for all the band components with spatial coordinates (x, y) contained in the hyperspectral image f (x, y, λ) as follows:
g(x,y)＝[f(x,y,1),f(x,y,2),...,f(x,y,λ),...,f(x,y,L)] (2)
wherein, λ is more than or equal to 1 and less than or equal to L, λ represents a band index value before fitting, L represents the maximum band number before fitting, and g (x, y) is a spectral curve expression of a space arbitrary point before fitting.
Step S12, fitting it in the spectral direction to the nth order, at a specific point in space (x)0,y0) P (x) according to formula (2)0,y0) Performing polynomial fitting of each term f (x)0,y0λ) the fitting result is:
wherein x is0,y0Is a constant, phi (x)0,y0,1),φ(x0,y0,2),φ(x0,y0,λ′),φ(x0,y0,n),φ(x0,y0N +1) are each corresponding term coefficient of the fitting, λ 'is a coefficient index value which is an integer and λ' is not less than 1 and not more than n +1,
is to f (x)0,y0λ) fitting estimates;
step S13, extracting coefficients of the polynomial in the fitting result to reconstruct a new spectral curve:
G(x0,y0)＝[φ(x0,y0,1),φ(x0,y0,2),...,φ(x0,y0,λ′),...,φ(x0,y0,n+1)] (4)
wherein λ' is a new band index value, G (x)0,y0) For the fitted specific point (x) in space0,y0) The spectral curve of (1).
Step S14, performing curve fitting on a spectrum curve g (x, y) formed by the pixels (x, y) at any position of the whole hyperspectral image by combining the new spectrum curve to obtain:
G(x,y)＝[φ(x,y,1),φ(x,y,2),...,φ(x,y,λ′),...,φ(x,y,n+1)] (5)
wherein n +1 represents the maximum number of wave bands after fitting, 1 is not less than lambda '. is not less than n +1, lambda' represents the new wave band index value after fitting, and G (x, y) is the spectral curve expression of any point in space after fitting. Wherein, from the formula (5) and comparing the formulas (1) and (2), phi (x, y, lambda') is a new hyperspectral image expression. Through fitting, the maximum number of wave bands is reduced from original L wave bands to n +1 wave bands; fig. 2 shows a schematic diagram of a comparison between an original spectral curve and spectral curves obtained when first-order, second-order, and third-order fitting are performed respectively, and it can be seen from the diagram that the original hyperspectral images are relatively similar when the third-order fitting is performed, so that n in the embodiment of the present invention is 3, and the new maximum number of bands is n +1 — 4, which can be seen that the amount of data to be processed is greatly reduced by means of fitting.
Step S2, constructing a new hyperspectral image phi (x, y, lambda') about a certain point p0And point p in its neighborhood1A weighted correlation function of;
the invention relates to a method for detecting local characteristic points of a compressed hyperspectral image, which is improved based on a Harris two-dimensional image detection method.
Set point p0Is a pixel in the new hyperspectral image phi (x, y, lambda ') with coordinates (x, y, lambda') and a point p1Is p0A point on the neighborhood of (x +. DELTA.x, y +. DELTA.y, lambda' +. DELTA.lambda),then p is0And p1The correlation function is defined as follows:
c(△x,△y,△λ)＝[φ(x,y,λ′)-φ(x+△x,y+△y,λ′+△λ)]2 (6)
wherein phi (x, y, lambda') is a point p0The DN value of the corresponding new hyperspectral image, φ (x +. DELTA.x, y +. DELTA.y, λ' +. DELTA.λ) is the point p1The corresponding DN value.
The following describes the weighted correlation function for hyperspectral images;
the invention judges the correlation between the hyperspectral pixel and the neighborhood by adopting the convolution of the window function and the hyperspectral image, therefore, on the basis of the formula (6), the weighted correlation function is defined as follows;
in particular, point p0And p1The weighted correlation function is defined as follows:
where φ (x, y, λ') is a point p0The corresponding DN value of the hyperspectral image, φ (x +. DELTA.x, y +. DELTA.y, λ' +. DELTA.λ) is the point p1The corresponding DN value;
the window function ω (x, y, λ') employs a gaussian weighting function, as follows:
where σ is a scale factor of a gaussian function.
In the formula (7), the reaction mixture is,
for convolution operation, l is the length of the window moving along the x direction, m is the length of the window moving along the y direction, and r is the length of the window moving along the λ direction, in the embodiment of the present invention, the values of l, m, and r all take 1, that is, the window size is 3 × 3.
Step S3, constructing a characteristic point response function according to the weighted correlation function;
specifically, the implementation process is as follows:
will be given in formula (7)
Is shown as
Namely:
in the formula (9), phi (u +. DELTA.x, v +. DELTA.y, p +. DELTA.lambda) is obtained by translating the image phi (u, v, p) (DELTA.x, DELTA.y, DELTA.lambda); performing Taylor series expansion on phi (u +. DELTA.x, v +. DELTA.y, p +. DELTA.lambda), and taking a first order approximation as follows:
in the formula, phix,φy,φλ′Is the gradient of each point of the image phi (x, y, lambda ') in three directions x, y, lambda', i.e.:
thus, equation (9) can be approximated as:
wherein the content of the first and second substances,
in the formula, phix 2,φy 2,φλ′ 2Respectively represents the gradient phi of the hyperspectral image in the three directions of x, y and lambdax,φy,φλ′Square of (d), phixφyIs indicative of phixPhi and phiyProduct of (a), phiyφλ′Is indicative of phiyPhi and phiλ′Product of (a), phixφλ′Is indicative of phixPhi and phiλ′ω is a Gaussian weighting function ω (x, y, λ') in the formula (8),
for the convolution symbols A, B, C, D, E, F correspond to the elements of matrix M respectively.
The characteristic point distinguishing method does not need to calculate a specific characteristic value, but calculates a characteristic point response value R to judge the characteristic point; specifically, the characteristic point response function formula is:
R＝det(M)-k(trace(M))3＝(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3 (16)
in the formula, k is an empirical constant, and the value is 0.001; det (M) is the determinant of matrix M, trace (M) is the trace of matrix M, and the expression is as follows:
det(M)＝λ1λ2λ3＝ABC+2DEF-BE2-AF2-CD2 (17)
trace(M)＝λ1+λ2+λ3＝A+B+C (18)
wherein λ is1、λ2、λ3Respectively are the eigenvalues of the matrix M; that is, in practice, although the eigenvalues of the matrix M are not specifically found, the eigenvalues are already included in det (M) and trace (M).
Step S4, calculating a certain point p in the hyperspectral image phi (x, y, lambda') according to the characteristic point response function0Is characterized by the soundResponse values of all the characteristic points of all the points in the neighborhood of the response values;
step S5, if a certain point p in the hyperspectral image phi (x, y, lambda')0If the response value of the feature point is greater than the response values of the feature points of all the points in the neighborhood, the point p is determined0The characteristic points of the hyperspectral image phi (x, y, lambda') are obtained;
specifically, the step S5 specifically includes: comparing a certain point p in the hyperspectral image phi (x, y, lambda')0(x, y, λ') and its 3 x 3 neighborhood, if point p0(x, y, λ ') in its 3 × 3 × 3 neighborhood, with R (x, y, λ') at a maximum, point p0(x, y, lambda') is the characteristic point of the hyperspectral image.
According to the hyperspectral image characteristic detection method based on spectral curve fitting, provided by the invention, the whole hyperspectral image can be subjected to data compression and noise reduction by a fitting method in the spectral direction, the data volume to be processed is reduced, the operation efficiency is improved, the spectral curve characteristic of a substance is reserved, and the purpose of noise reduction is achieved; furthermore, feature point detection can be carried out on the hyperspectral image on the basis, key information of the hyperspectral image is obtained, the hyperspectral image can be better analyzed, and the hyperspectral image pattern recognition effect is improved. The local feature point detection method can be applied to the aspects of hyperspectral image classification and identification, hyperspectral image target detection, material sorting and the like; the method has good effect in the classification experiment of the hyperspectral image, and makes a contribution to the scientific field of local feature detection of the hyperspectral image.
Referring to fig. 3, a hyperspectral image feature detection apparatus based on spectral curve fitting according to an embodiment of the present invention includes:
the fitting module 1 is used for performing n-order fitting on the hyperspectral image f (x, y, lambda) in the spectral direction to obtain a new hyperspectral image phi (x, y, lambda');
wherein x and y represent space domain coordinates, and lambda represents spectrum domain coordinates, wherein lambda is more than or equal to 1 and less than or equal to L, and lambda' is more than or equal to 1 and less than or equal to n + 1; λ represents a band index value before fitting, λ' represents a band index value after fitting, L represents the maximum band number before fitting, and n +1 represents the maximum band number after fitting;
a weighted correlation function construction module 2 for constructing a new hyperspectral image phi (x, y, lambda') about a certain point p0And point p in its neighborhood1A weighted correlation function of;
a characteristic point response function constructing module 3, configured to construct a characteristic point response function according to the weighted correlation function;
a characteristic point response value calculating module 4, configured to calculate a certain point p in the hyperspectral image phi (x, y, λ') according to the characteristic point response function0The response value of the characteristic point of the point and the response values of the characteristic points of all the points in the neighborhood of the point;
a feature point determination module 5, configured to determine a certain point p in the hyperspectral image phi (x, y, lambda')0When the response value of the characteristic point is larger than the response values of the characteristic points of all the points in the neighborhood, the point p is judged0Namely the characteristic points of the hyperspectral image phi (x, y, lambda').
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.
Claims (8)
1. A hyperspectral image feature detection method based on spectral curve fitting is characterized by comprising the following steps:
step S1, fitting the hyperspectral image f (x, y, lambda) in the n order in the spectral direction to obtain a new hyperspectral image phi (x, y, lambda');
wherein x and y represent space domain coordinates, and lambda represents spectrum domain coordinates, wherein lambda is more than or equal to 1 and less than or equal to L, and lambda' is more than or equal to 1 and less than or equal to n + 1; λ represents a band index value before fitting, λ' represents a band index value after fitting, L represents the maximum band number before fitting, and n +1 represents the maximum band number after fitting;
step S2, constructing a new hyperspectral image phi (x, y, lambda') about a certain point p0And points in its neighborhoodp1A weighted correlation function of;
step S3, constructing a characteristic point response function according to the weighted correlation function;
step S4, calculating a certain point p in the hyperspectral image phi (x, y, lambda') according to the characteristic point response function0The response value of the characteristic point of the point and the response values of the characteristic points of all the points in the neighborhood of the point;
step S5, if a certain point p in the hyperspectral image phi (x, y, lambda')0If the response value of the feature point is greater than the response values of the feature points of all the points in the neighborhood, the point p is determined0The characteristic points of the hyperspectral image phi (x, y, lambda') are obtained;
the step S1 specifically includes:
step S11, constructing a spectral curve expression for all the band components with spatial coordinates (x, y) contained in the hyperspectral image f (x, y, λ) as follows:
g(x,y)＝[f(x,y,1),f(x,y,2),...,f(x,y,λ),...,f(x,y,L)]
wherein, λ is more than or equal to 1 and less than or equal to L, λ represents a band index value before fitting, L represents the maximum band number before fitting, and g (x, y) is a spectral curve expression of any spatial point before fitting;
step S12, for a specific point (x) in space in the spectral direction according to the above formula0,y0) G (x) of0,y0) Performing polynomial fitting of each term f (x)0,y0λ) the fitting result is:
wherein x is0,y0Is a constant, phi (x)0,y0,1),φ(x0,y0,2),φ(x0,y0,λ′),φ(x0,y0,n),φ(x0,y0N +1) are each corresponding term coefficient of fitting, λ 'is a band index value after representing fitting, which is an integer and 1. ltoreq. λ'. ltoreq.n +1,
is to f (x)0,y0λ) fitting estimates;
step S13, extracting coefficients of the polynomial in the fitting result to reconstruct a new spectral curve:
G(x0,y0)＝[φ(x0,y0,1),φ(x0,y0,2),...,φ(x0,y0,λ′),...,φ(x0,y0,n+1)]
wherein λ' is a band index value after fitting; g (x)0,y0) For the fitted specific point (x) in space0,y0) The spectral curve expression of (a);
step S14, performing curve fitting on a spectrum curve g (x, y) formed by the pixels (x, y) at any position of the whole hyperspectral image by combining the new spectrum curve to obtain:
G(x,y)＝[φ(x,y,1),φ(x,y,2),...,φ(x,y,λ′),...,φ(x,y,n+1)]
wherein phi (x, y, lambda ') is a new hyperspectral image expression, lambda ' is more than or equal to 1 and less than or equal to n +1, and lambda ' is a waveband index value after fitting; n +1 represents the maximum number of bands after fitting, and G (x, y) is the spectral curve expression of the arbitrary point in space after fitting.
2. The hyperspectral image feature detection method of claim 1, wherein the weighted correlation function is:
wherein, the point p0Is a pixel in the hyperspectral image phi (x, y, lambda ') with coordinates (x, y, lambda ') and phi (x, y, lambda ') as a point p0The DN value of the corresponding hyperspectral image; point p1Coordinates are (x + Δ x, y + Δ y, λ '+ Δ λ), φ (x + Δ x, y + Δ y, λ' + Δ λ) is a point p1The corresponding DN value;
the window function ω (x, y, λ') employs a gaussian weighting function, as follows:
where σ is a scale factor of the Gaussian function and Δ x is a point p0To point p1Offset value in x direction, Δ y being point p0To point p1Offset value in y direction, Δ λ is point p0To point p1An offset value in the λ' direction;
wherein the content of the first and second substances,
for the convolution operation symbol, l is the length of the window function moving in the x direction, m is the length of the window function moving in the y direction, r is the length of the window function moving in the λ' direction, l is 1, m is 1, r is 1, i.e., the window size is 3 × 3.
3. The hyperspectral image feature detection method of claim 2, wherein in the weighted correlation function
and, instead,
then the process of the first step is carried out,
wherein the content of the first and second substances,
in the formula, phix,φy,φλ′Respectively, the gradient of the image phi (x, y, lambda ') in three directions x, y, lambda', namely:
in the above formula, ω represents a Gaussian weighting function ω (x, y, λ'),
for the convolution symbols A, B, C, D, E, F correspond to each element of the matrix M, phi, respectivelyx 2,φy 2,φλ′ 2Respectively represents the gradient phi of the multispectral image in three directions of x, y and lambdax,φy,φλ′Square of (d), phixφyIs indicative of phixPhi and phiyProduct of (a), phiyφλ′Is indicative of phiyPhi and phiλ′Product of (a), phixφλ′Is indicative of phixPhi and phiλ′A, B, C, D, E, F correspond to the elements of the matrix M, respectively.
4. The hyperspectral image feature detection method according to claim 3, wherein the feature point response function is:
R＝det(M)-k(trace(M))3＝(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3；
wherein k is 0.001, and k is an empirical constant; det (M) represents the determinant of matrix M, trace (M) represents the traces of matrix M, and the expression is as follows:
det(M)＝λ1λ2λ3＝ABC+2DEF-BE2-AF2-CD2
trace(M)＝λ1+λ2+λ3＝A+B+C
wherein λ is1、λ2、λ3Respectively, the eigenvalues of the matrix M.
5. A hyperspectral image feature detection device based on spectral curve fitting is characterized by comprising:
the fitting module is used for performing n-order fitting on the hyperspectral image f (x, y, lambda) in the spectral direction to obtain a new hyperspectral image phi (x, y, lambda');
wherein x and y represent space domain coordinates, and lambda represents spectrum domain coordinates, wherein lambda is more than or equal to 1 and less than or equal to L, and lambda' is more than or equal to 1 and less than or equal to n + 1; λ represents a band index value before fitting, λ' represents a band index value after fitting, L represents the maximum band number before fitting, and n +1 represents the maximum band number after fitting;
a weighted correlation function construction module for constructing a new hyperspectral image phi (x, y, lambda') about a certain point p0And point p in its neighborhood1A weighted correlation function of;
the characteristic point response function constructing module is used for constructing a characteristic point response function according to the weighted correlation function;
a characteristic point response value calculation module for calculating a certain point p in the hyperspectral image phi (x, y, lambda') according to the characteristic point response function0The response value of the characteristic point of the point and the response values of the characteristic points of all the points in the neighborhood of the point;
a characteristic point judging module for judging a certain point p in the hyperspectral image phi (x, y, lambda')0When the response value of the characteristic point is larger than the response values of the characteristic points of all the points in the neighborhood, the point p is judged0The characteristic points of the hyperspectral image phi (x, y, lambda') are obtained;
the fitting module includes:
a first construction submodule, configured to construct a spectral curve expression for all band components with spatial coordinates (x, y) contained in the hyperspectral image f (x, y, λ) as follows:
g(x,y)＝[f(x,y,1),f(x,y,2),...,f(x,y,λ),...,f(x,y,L)]
wherein, λ is more than or equal to 1 and less than or equal to L, λ represents a band index value before fitting, L represents the maximum band number before fitting, and g (x, y) is a spectral curve expression of any spatial point before fitting;
a first fitting submodule for fitting a specific point (x) in space in the spectral direction according to the above formula0,y0) G (x) of0,y0) Performing polynomial fitting of each term f (x)0,y0λ) the fitting result is:
wherein x is0,y0Is a constant, phi (x)0,y0,1),φ(x0,y0,2),φ(x0,y0,λ′),φ(x0,y0,n),φ(x0,y0N +1) are each corresponding term coefficient of fitting, λ 'is a band index value after fitting, which is an integer and λ' is not less than 1 and not more than n +1,
is to f (x)0,y0λ) fitting estimates;
a second construction submodule, configured to extract coefficients of the polynomial in the fitting result to reconstruct a new spectral curve:
G(x0,y0)＝[φ(x0,y0,1),φ(x0,y0,2),...,φ(x0,y0,λ′),...,φ(x0,y0,n+1)]
wherein λ' is the band index value after fitting; g (x)0,y0) For the fitted specific point (x) in space0,y0) The spectral curve expression of (a);
and the second fitting submodule is used for performing curve fitting on a spectral curve g (x, y) formed by pixels (x, y) at any position of the whole hyperspectral image by combining the new spectral curve to obtain:
G(x,y)＝[φ(x,y,1),φ(x,y,2),...,φ(x,y,λ′),...,φ(x,y,n+1)]
wherein phi (x, y, lambda ') is a new hyperspectral image expression, lambda ' is more than or equal to 1 and less than or equal to n +1, and lambda ' is a wave band index value after fitting; n +1 represents the maximum number of bands after fitting, and G (x, y) is the spectral curve expression of the arbitrary point in space after fitting.
6. The hyperspectral image feature detection apparatus according to claim 5, wherein the weighted correlation function is:
wherein, the point p0Is a pixel in the hyperspectral image phi (x, y, lambda ') with coordinates (x, y, lambda ') and phi (x, y, lambda ') as a point p0The DN value of the corresponding hyperspectral image; point p1Coordinates are (x + Δ x, y + Δ y, λ '+ Δ λ), φ (x + Δ x, y + Δ y, λ' + Δ λ) is a point p1The corresponding DN value;
the window function ω (x, y, λ') employs a gaussian weighting function, as follows:
where σ is a scale factor of the Gaussian function and Δ x is a point p0To pointp1Offset value in x direction, Δ y being point p0To point p1Offset value in y direction, Δ λ is point p0To point p1An offset value in the λ' direction;
wherein the content of the first and second substances,
for the convolution operation symbol, l is the length of the window function moving in the x direction, m is the length of the window function moving in the y direction, r is the length of the window function moving in the λ' direction, l is 1, m is 1, r is 1, i.e., the window size is 3 × 3.
7. The hyperspectral image feature detection apparatus of claim 6, wherein in the weighted correlation function
and, instead,
then the process of the first step is carried out,
wherein the content of the first and second substances,
in the formula, phix,φy,φλ′Respectively, the gradient of the image phi (x, y, lambda ') in three directions x, y, lambda', namely:
in the above formula, ω represents a Gaussian weighting function ω (x, y, λ'),
for the convolution symbols A, B, C, D, E, F correspond to each element of the matrix M, phi, respectivelyx 2,φy 2,φλ′ 2Respectively represents the gradient phi of the multispectral image in three directions of x, y and lambdax,φy,φλ′Square of (d), phixφyIs indicative of phixPhi and phiyProduct of (a), phiyφλ′Is indicative of phiyPhi and phiλ′Product of (a), phixφλ′Is indicative of phixPhi and phiλ′A, B, C, D, E, F correspond to the elements of the matrix M, respectively.
8. The hyperspectral image feature detection apparatus according to claim 7, wherein the feature point response function is:
R＝det(M)-k(trace(M))3＝(ABC+2DEF-BE2-AF2-CD2)-k(A+B+C)3；
wherein k is 0.001, and k is an empirical constant; det (M) represents the determinant of matrix M, trace (M) represents the traces of matrix M, and the expression is as follows:
det(M)＝λ1λ2λ3＝ABC+2DEF-BE2-AF2-CD2
trace(M)＝λ1+λ2+λ3＝A+B+C
wherein λ is1、λ2、λ3Respectively, are characteristic of the matrix M.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711097228.7A CN108009550B (en) | 2017-11-09 | 2017-11-09 | Hyperspectral image characteristic detection method and device based on spectral curve fitting |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711097228.7A CN108009550B (en) | 2017-11-09 | 2017-11-09 | Hyperspectral image characteristic detection method and device based on spectral curve fitting |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108009550A CN108009550A (en) | 2018-05-08 |
| CN108009550B true CN108009550B (en) | 2021-01-22 |
Family
ID=62052283
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711097228.7A Active CN108009550B (en) | 2017-11-09 | 2017-11-09 | Hyperspectral image characteristic detection method and device based on spectral curve fitting |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108009550B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108876711B (en) * | 2018-06-20 | 2023-01-31 | 山东师范大学 | Sketch generation method, server and system based on image feature points |
| CN109360264B (en) * | 2018-08-30 | 2023-05-26 | 深圳大学 | Method and device for establishing unified image model |
| CN111199251B (en) * | 2019-12-27 | 2020-11-27 | 中国地质大学（北京） | Multi-scale hyperspectral image classification method based on weighted neighborhood |
| CN113962904B (en) * | 2021-11-26 | 2023-02-10 | 江苏云脑数据科技有限公司 | Method for filtering and denoising hyperspectral image |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2005000110A3 (en) * | 2003-06-23 | 2005-06-16 | Microvision Inc | Scanning endoscope |
| CN101299237A (en) * | 2008-06-05 | 2008-11-05 | 北京航空航天大学 | High spectroscopic data supervision classifying method based on information quantity dimensionality sequence |
| CN101598798A (en) * | 2008-12-31 | 2009-12-09 | 中国资源卫星应用中心 | A kind of system and method to rebuilding spectrum of high spectrum intervention data |
| WO2013109966A1 (en) * | 2012-01-20 | 2013-07-25 | The Trustees Of Dartmouth College | Method and apparatus for quantitative hyperspectral fluorescence and reflectance imaging for surgical guidance |
| CN103940511A (en) * | 2014-04-03 | 2014-07-23 | 清华大学 | Spectrum line calibration method and device for hyper-spectrum acquisition system |
| CN104766282A (en) * | 2015-04-13 | 2015-07-08 | 清华大学深圳研究生院 | Repairing method of hyperspectral image |
| CN105139412A (en) * | 2015-09-25 | 2015-12-09 | 深圳大学 | Hyperspectral image corner detection method and system |
| CN105426585A (en) * | 2015-11-05 | 2016-03-23 | 浙江大学 | Potato sprouting early warning method based on sine function fitting method |
-
2017
- 2017-11-09 CN CN201711097228.7A patent/CN108009550B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2005000110A3 (en) * | 2003-06-23 | 2005-06-16 | Microvision Inc | Scanning endoscope |
| CN101299237A (en) * | 2008-06-05 | 2008-11-05 | 北京航空航天大学 | High spectroscopic data supervision classifying method based on information quantity dimensionality sequence |
| CN101598798A (en) * | 2008-12-31 | 2009-12-09 | 中国资源卫星应用中心 | A kind of system and method to rebuilding spectrum of high spectrum intervention data |
| WO2013109966A1 (en) * | 2012-01-20 | 2013-07-25 | The Trustees Of Dartmouth College | Method and apparatus for quantitative hyperspectral fluorescence and reflectance imaging for surgical guidance |
| CN103940511A (en) * | 2014-04-03 | 2014-07-23 | 清华大学 | Spectrum line calibration method and device for hyper-spectrum acquisition system |
| CN104766282A (en) * | 2015-04-13 | 2015-07-08 | 清华大学深圳研究生院 | Repairing method of hyperspectral image |
| CN105139412A (en) * | 2015-09-25 | 2015-12-09 | 深圳大学 | Hyperspectral image corner detection method and system |
| CN105426585A (en) * | 2015-11-05 | 2016-03-23 | 浙江大学 | Potato sprouting early warning method based on sine function fitting method |
Non-Patent Citations (3)
| Title |
|---|
| "实践九号"A卫星光学遥感图像杂散光噪声去除;严明等;《航天返回与遥感》;20141031;第35卷(第5期);第72-80页 * |
| 《镇江香醋固态发酵参数的智能在线监测及其分布研究》;朱瑶迪;《中国博士学位论文全文数据库 工程科技Ⅰ辑》;20160815;第2016年卷(第8期);B024-47 * |
| Using Curve Fitting for Spectral Reflectance Curves Intervals in Order to Hyperspectral Data Compression;Mersedeh Beitollahi等;《2016 10th International Symposium on Communication Systems, Networks and Digital Signal Processing (CSNDSP)》;20160922;第1-5页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108009550A (en) | 2018-05-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108009550B (en) | Hyperspectral image characteristic detection method and device based on spectral curve fitting | |
| CN109376804B (en) | Hyperspectral remote sensing image classification method based on attention mechanism and convolutional neural network | |
| Du et al. | On the performance evaluation of pan-sharpening techniques | |
| Bendoumi et al. | Hyperspectral image resolution enhancement using high-resolution multispectral image based on spectral unmixing | |
| CN106557784B (en) | Rapid target identification method and system based on compressed sensing | |
| Chen et al. | Sparse hyperspectral unmixing based on constrained lp-l 2 optimization | |
| Bourennane et al. | Improvement of classification for hyperspectral images based on tensor modeling | |
| CN109102469B (en) | Remote sensing image panchromatic sharpening method based on convolutional neural network | |
| CN109859110B (en) | Hyperspectral image panchromatic sharpening method based on spectrum dimension control convolutional neural network | |
| JPS63118889A (en) | Change detection system by picture | |
| CN110427997B (en) | Improved CVA change detection method for complex remote sensing image background | |
| CN109255358B (en) | 3D image quality evaluation method based on visual saliency and depth map | |
| Mignotte | A bicriteria-optimization-approach-based dimensionality-reduction model for the color display of hyperspectral images | |
| CN110570395B (en) | Hyperspectral anomaly detection method based on spatial-spectral combined collaborative representation | |
| Li et al. | Denoising of hyperspectral images employing two-phase matrix decomposition | |
| Zhang et al. | Regularization framework for target detection in hyperspectral imagery | |
| Zhang et al. | Data-driven transform-based compressed image quality assessment | |
| CN112633202A (en) | Hyperspectral image classification algorithm based on dual denoising combined multi-scale superpixel dimension reduction | |
| Wang et al. | Semi-NMF-based reconstruction for hyperspectral compressed sensing | |
| CN111260655A (en) | Image generation method and device based on deep neural network model | |
| CN107742114B (en) | Hyperspectral image feature detection method and device | |
| CN113129300A (en) | Drainage pipeline defect detection method, device, equipment and medium for reducing false detection rate | |
| Yang et al. | Joint collaborative representation with shape adaptive region and locally adaptive dictionary for hyperspectral image classification | |
| US8897378B2 (en) | Selective perceptual masking via scale separation in the spatial and temporal domains using intrinsic images for use in data compression | |
| Wang et al. | Image recovery and recognition: a combining method of matrix norm regularisation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |