CN110580463A - Single spectrum driven high-spectrum image target detection method based on double-category sparse representation - Google Patents

Single spectrum driven high-spectrum image target detection method based on double-category sparse representation Download PDF

Info

Publication number
CN110580463A
CN110580463A CN201910811691.6A CN201910811691A CN110580463A CN 110580463 A CN110580463 A CN 110580463A CN 201910811691 A CN201910811691 A CN 201910811691A CN 110580463 A CN110580463 A CN 110580463A
Authority
CN
China
Prior art keywords
target
background
sparse
dictionary
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.)
Granted
Application number
CN201910811691.6A
Other languages
Chinese (zh)
Other versions
CN110580463B (en
Inventor
杜博
朱德辉
张良培
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.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
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 Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN201910811691.6A priority Critical patent/CN110580463B/en
Publication of CN110580463A publication Critical patent/CN110580463A/en
Application granted granted Critical
Publication of CN110580463B publication Critical patent/CN110580463B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB

Abstract

The invention discloses a single spectrum driven dual-class sparse representation hyperspectral image target detection method which comprises the steps of constructing a target dictionary and a background dictionary, wherein a given target spectrum is used as prior information when the target dictionary is constructed; when a background dictionary is constructed, classifying and selecting pixels with high representation frequency in each class as background training samples to obtain a global over-complete background dictionary; and for the target dictionary and the background dictionary, sparse representation is carried out on pixels by using a dual-category sparse representation model to obtain a target sparse vector and a background sparse vector, the pixels to be detected are detected one by one, and a target detection result of the hyperspectral remote sensing image X is extracted. The background dictionary is a global background dictionary, so that all background ground object types on a global image are well represented; and the target class and the background class are separated by the dual-class sparse representation mode to carry out sparse reconstruction, so that the separation of the target and the background in the hyperspectral image is efficiently realized, and the target of interest is detected.

Description

Single spectrum driven high-spectrum image target detection method based on double-category sparse representation
Technical Field
The invention belongs to the technical field of remote sensing image processing, and particularly relates to a single-spectrum-driven dual-class sparse representation hyperspectral image target detection method.
Background
The rapid development of Remote sensing earth observation technology and its application has changed the world model of human cognition to a great extent, and it has become an important technical means for acquiring earth surface information (Griffith, j. (1979). The value that the remote sensing earth observation technology can play in practical application depends to a great extent on whether the image obtained by satellite-transmitted remote sensing can provide detailed and rich earth surface information. Compared with multispectral images, the hyperspectral remote sensing images have the characteristics of large number of wave bands and extremely high spectral resolution, provide rich surface feature spectral information, particularly distinguish diagnostic spectral information of different surface features and distinguish spectral information of subtle differences among similar surface features, and bring unique advantages for acquiring surface information.
Hyperspectral image object detection is actually a binary problem, and is the process of separating objects on images from the background by the spectral difference of ground objects given a priori object information (nasrabdi, & Nasser, m. (2014.). Hyperspectral target detection: an overview of current and future. ieee Signal Processing Magazine,31(1), 34-44.). At present, hyperspectral image target detection is widely applied to the fields of environmental detection, urban investigation, mineral mapping, military reconnaissance and the like. How to quickly and accurately extract an interested target in an image is a difficult problem of target detection of a hyperspectral image.
Aiming at the problem of hyperspectral image target detection, scholars at home and abroad propose a plurality of methods, including a detection algorithm based on spectrum matching, a detection algorithm based on hypothesis testing, a detection algorithm based on subspace and the like. The algorithm is mostly proposed based on a linear mixed model, a pixel is assumed to be formed by mixing various ground objects, and noise is assumed to be subjected to multivariate normal distribution. This is not the case, and thus such algorithms do not achieve good detection results. The sparse representation does not need any assumption on the distribution of target pixels, background pixels or noise on the image, so that the sparse representation is widely applied to the problem of hyperspectral image target detection. It is considered that in the hyperspectral image, a background spectrum falls into a subspace spanned by a series of background training samples, and background pixels can be linearly represented by the series of background training samples; similarly, the target spectrum falls into a subspace spanned by the target training samples, and the target pixels can be linearly represented by a series of target training samples. And reconstructing pixels of two categories of a target and a background on the image through sparse representation to obtain a reconstructed residual, and comparing the residual to obtain an output value to obtain a target detection result. In the sparse representation-based method, a target dictionary is generally constructed by selecting a plurality of target pixels from a global image, and for a hyperspectral image, the target pixels are not known and cannot be acquired correctly. Meanwhile, the number of target pixels selected from the global image is very small, which is not enough to well represent all target pixels possibly existing on the image, and the recovery of sparse vectors and the acquisition of residual errors are influenced, so that the final detection effect is influenced. For the problem, in the previous work, a method for constructing an object dictionary through pre-detection is provided for improvement, and a better detection result is obtained. (Zhu, D., Du, B., and Zhang L. (2019). Target DirectionConstruction-Based spark repetition hyper spectral Target detection methods IEEE Journal of Selected Topics in Applied Earth Observation and remove Sensing 12(4): 1254-)
Although sparse representation has achieved some success in the hyperspectral target detection problem, there are still some disadvantages:
1) In the sparse representation-based method, the background dictionary is generally constructed by adopting a dual concentric window method. The double concentric windows construct the image elements of the outer window into a local background dictionary, and represent the central image element. In some parts of the image, when the ground object types of the central pixel and the external window pixel are not consistent, the mode is not reasonable, pixel reconstruction is affected, and a poor detection result is obtained. Meanwhile, the method for constructing the local background dictionary does not fully utilize the global information of the image.
2) When the existing sparse representation-based method is used for recovering sparse vectors, two sparse vectors of a target and a background are recovered simultaneously. The method can enable the two sparse vectors to mutually influence, and the recovered sparse vectors are poor in effect, so that the final detection effect is influenced.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides the single-spectrum-driven dual-category sparse representation hyperspectral image target detection method with high precision and high efficiency.
In order to solve the technical problems, the invention adopts the following technical scheme:
Finalization to be complemented according to the claims and temporary suspension of processing
The invention has the beneficial effects that:
(1) According to the method, a global background dictionary is constructed by fully utilizing the non-local similarity characteristics on the image through the classification idea. The construction of the global background dictionary avoids the defect that the traditional double concentric windows are unreasonable in local representation of the image, and meanwhile, the classified construction of the background dictionary enables atypical pixels, namely target pixels or anomalies in each category to be removed, so that the construction of the background dictionary is complete and pure, and all pixels of background ground object categories on the global image can be well represented.
(2) The invention provides a double-class sparse representation method, which separates a target class and a background class for sparse reconstruction, avoids mutual influence of target and background sparse vectors during reconstruction, and obtains sparse vectors with distinguishing force during reconstruction of two classes of pixels of the target and the background, thereby realizing separation of the target and the background to the maximum degree while efficiently detecting the target.
Drawings
FIG. 1 is a schematic flow chart of an embodiment of the present invention.
Detailed Description
For the convenience of those skilled in the art to understand and implement the technical solution of the present invention, the following detailed description of the present invention is provided in conjunction with the accompanying drawings and examples, it is to be understood that the embodiments described herein are only for illustrating and explaining the present invention and are not to be construed as limiting the present invention.
The invention discloses a single-spectrum-driven dual-category sparse representation hyperspectral image target detection method. In the target dictionary construction part: and a given target spectrum is used as prior information, a classical target detection operator is used for pre-detection, and a part of pixels with larger output values in the initial detection statistic values are selected to construct a target dictionary. In the background dictionary construction part: firstly, reducing the dimension of an original hyperspectral image by using a principal component analysis method, and then classifying the image by using K-means clustering; for each pixel category on the image, performing sparse representation by using other pixels except the current pixel in the category to obtain the use frequency of the current pixel in the category, wherein the more frequent the pixel has more typicality in the category, and the pixel with typicality is selected in each category to construct a global over-complete background dictionary. In the detection output section: and respectively detecting the pixels to be detected one by using a double-class sparse representation method for the constructed target dictionary and the background dictionary to obtain a target detection result. In the invention, the target dictionary is constructed, so that an unreasonable mode that a sparse representation model selects a target pixel from an image as prior information is avoided, and meanwhile, enough target samples can be obtained, so that the target class in the image is well represented; the background dictionary is a global background dictionary, so that the defect that the traditional double concentric windows are unreasonable in local representation of the image is avoided, and meanwhile, the pixels of all background ground object types on the global image can be well represented; the dual-category sparse representation method separates the target category and the background category for sparse reconstruction, avoids mutual influence of target and background sparse vectors during reconstruction, and obtains sparse vectors with distinguishing force during reconstruction of two categories of pixels of the target and the background, so that the target and the background in a hyperspectral image are separated, and an interested target is detected.
The key technology of the invention comprises three parts: the method comprises the steps of constructing a target dictionary, constructing a global background dictionary and performing double-category sparse representation. The construction of the target dictionary can obtain sufficient target training samples, target pixels in the hyperspectral image can be well expressed in a sparse mode, more appropriate sparse vectors can be recovered, and a better reconstruction effect can be obtained. The background dictionary is built in a classified manner, so that the dictionary can be built more cleanly (namely, objects or abnormal objects are removed), and meanwhile, all background ground object types possibly existing on the image can be included in the background dictionary, so that the background dictionary is built more completely, and pixels of all background ground object types on the global image can be well expressed. The dual-class sparse representation separates the target class and the background class for sparse reconstruction, avoids mutual influence of target and background sparse vectors during reconstruction, and the obtained sparse vectors have distinguishing capability during reconstruction of two classes of pixels of the target and the background. The three parts and the combination improve the target detection effect to a certain extent.
the embodiment is realized by adopting an MATLAB platform, and an MATLAB remote sensing image read-write function is taken as an implementation basis. Calling a remote sensing image reading function, inputting a file name of the remote sensing image, reading the hyperspectral remote sensing image X into a matrix SX with the size of P multiplied by N, wherein each element in the matrix is a pixel radiation value corresponding to each wave band, P is the wave band number of the hyperspectral remote sensing image, and N is the pixel number of the hyperspectral remote sensing image. The MATLAB remote sensing image read-write function is a well-known technology in the art and is not described herein.
In the embodiment, the following operations are performed on the hyperspectral remote sensing image X based on the matrix SX:
(1): pre-detecting an original hyperspectral remote sensing image by using a basic target detection operator to obtain an initial detection statistic value omega;
The specific operation of the step (1) is as follows: using a given target spectrum d as prior information, and for a hyperspectral remote sensing image X ═ { X ═ X1,x2,...,xNObtaining each pixel x on the image by using a basic target detection operator1,x2,...,xNInitial probe statistic ω1,ω2,...,ωNIt can be said that the set Ω ═ ω is set1,ω2,...,ωN}。
In an embodiment, the basic target detection operator preferably uses the constrained energy minimization operator CEM, i.e. the statistical valueWherein R is an autocorrelation matrix of the hyperspectral remote sensing image X, X is any pixel on the hyperspectral remote sensing image X, and XTIs the transpose of x, R-1Is the inverse of R, dTis the transpose of d. For the selection of the basic target detection operator, the present invention is not limited to using the CEM operator, and other various target detection operators such as the adaptive cosine estimation operator ACE, the spectrum matching filter operator SMF, etc. may be used.
(2): performing descending arrangement on the initial detection statistic value omega, and selecting pixels on the hyperspectral images corresponding to a plurality of values after descending arrangement as target training samples phi;
In the embodiment, the specific operation of step (2) is as follows: for the obtained initial detection statistic omega ═ omega { [ omega ]1,ω2,...,ωNAnd carrying out descending order arrangement, wherein the larger the output value is, the more probable the pixel corresponding to the hyperspectral image is to be a target pixel. Introducing a parameter t, wherein t is a multiple of the total pixel number of the hyperspectral image, considering that the number of interested targets in the hyperspectral image target detection problem only accounts for a small part, preferably, the value of the suggested t can be controlled to be 10-3Left and right (or adjusted according to multiple of target number or multiple of image pixel total number possibly existing in the hyperspectral image), and t is adjustable as a parameter according to different hyperspectral images. Selecting the output value of the t multiplied by N after descending order as a threshold value, selecting the pixel on the hyperspectral image corresponding to the element larger than the threshold value in the initial detection statistic value as a target training sample, and thus obtaining the target training sampleWherein N istThe number of samples is trained for the target.
(3): directly taking the target training sample phi obtained in the step (2) as a target dictionary At，
(4): and reducing the dimension of the hyperspectral remote sensing image X by using a principal component analysis method, and classifying the image on the image subjected to dimension reduction by using K-means clustering.
In the embodiment, the specific operation of step (4) is as follows: and (4) performing dimensionality reduction on the image X by using a principal component analysis method, preferably recommending to reduce the dimensionality to 5 dimensionalities, namely selecting the first 5 principal components after the principal component analysis for subsequent classification. And dividing the reduced images into p categories by using K-means clustering, and preferably controlling the value of the recommended p to be about 10 (the optimal value of p may be different according to different hyperspectral images, and the value of p can be set according to the rough judgment of the number of surface feature categories possibly existing on the images by visual interpretation). Set to obtain the classthe set of categories may be represented as1≤i≤p。
The specific implementation of the principal component analysis method and the K-means clustering is the prior art, and the detailed description of the invention is omitted.
(5) Sparse representation is carried out on the pixels of each category in the step (4), the representation frequency of each pixel is obtained, and the pixel with high representation frequency in each category is selected as a background training sample delta;
In the embodiment, the specific operation of step (5) is as follows: for the s picture element e in the i categoryisUsing sets in the current class other than the pelI is more than or equal to 1 and less than or equal to p to eisPerforming sparse representation, obtaining corresponding sparse vector thetais，θisIs the pixel eisIs indicative of frequency. Sparse representation is prior art, and the present invention is not described in detail. The higher the representation frequency is, the more typical the image element is in the ith category, the image elements with the typical characteristics are selected as background training samples in the category, and j is selectediA typical pixel is recorded asthen a background training sample is obtainedThen, a part of pixels with typicality are taken out from each category, and the whole background training sample delta is formed1,δ2,…,δpthe background training samples can also be expressed asWherein N isbTo the total number of background training samples,for the sample of picture elements therein, Nb＝j1+j2+...+jpPreferably, suggest NbThe value of (A) can be controlled to be about 300, and N is a high spectral image according to different high spectral imagesbIs adjustable as a parameter. In addition, when the pixels in each category are represented, the target pixels and the noise anomalies only occupy a very small part of the image, so that the target pixels or the noise anomalies cannot become typical pixels in any category, and the background training samples cannot have the target pixels or the noise anomalies, so that the construction is purer.
(6): directly taking the background training sample delta obtained in the step (5) as a background dictionary AbTo obtain a global overcomplete background dictionary
(7): for the target dictionary A obtained in step (3)tAnd the background dictionary A obtained in the step (6)bCarrying out sparse representation on pixels by using a two-category sparse representation method to obtain a target sparse vector alphatAnd the background sparse vector alphab。
The specific operation of the step (7) is as follows: each pixel in the hyperspectral image can be represented by a target and background dictionary, and when the pixel x is a background pixel:
Wherein N isbNumber of training samples for background, αbIn the case of a background sparse vector,When pixel x is the target pixel:
Wherein N istFor the number of target training samples, αtIn order to be the target sparse vector,In the existing sparse representation model, because samples in a target dictionary are insufficient, the target dictionary and a background dictionary are combined to obtain a combined dictionary A, and an unknown sample x is supposed to fall into the combined dictionary of the target and the background
x≈Abαb+Atαt＝Aα
Wherein A ═ Ab At]，The sparse vector α can be obtained using a sparse representation solution method and decomposed into target and background sparse vectors. However, in this way, when the sparse vector is recovered, the target and background sparse vectors will affect each other, and the recovered sparse vector cannot well perform sparse reconstruction on the pixel. Therefore, the invention provides that the target sparse vector and the background sparse vector are recovered separately, only the sample in the target dictionary is used for recovering the target sparse vector, and the sample in the background dictionary is used for recovering the background sparse vector. Because the method for constructing the target dictionary adopted by the invention fully considers the good representation of all possible target pixels on the image, the obtained target dictionary contains enough samplesTherefore, the recovered target sparse vector can well carry out sparse reconstruction on the pixel.
The two-class sparse representation model provided by the invention can be expressed as
x≈Abαb subject to||αb||0Kb target is absent
x≈Atαt subject to||αt||0Presence of Kt target ≦
where Kb is the background sparse level, Kt is the target sparse level, | | αb||0denotes alphabZero norm, | | αt||0Denotes alphatZero norm of (d). In the present invention, preferably, the value of the recommended Kb is controlled to be 20, and preferably, the value of the recommended Kt is controlled to be 10. Solving and obtaining the target sparse vector alpha of the pixel element through the double-category sparse representation model, the target dictionary and the background dictionarytand the background sparse vector alphab. The method for solving and obtaining the sparse vector for the sparse representation is a well-known technique in the technical field, and is not described herein again.
(8): using a target sparse vector alphatAnd the background sparse vector alphabreconstructing the image element to obtain reconstructed residual errors under two categories
rb(x)＝||x-Abαb||
rt(x)＝||x-Atαt||
Wherein r isb(x) Residual error, r, obtained for reconstructing the background picture elementt(x) The residual error obtained for reconstructing the target pixel.
After obtaining the residues under both categories, according to D (x) rb(x)-rt(x) And obtaining an output value D (X) of each pixel X on the hyperspectral remote sensing image X, wherein the larger the output value is, the more likely the pixel is a target pixel, and thus a target detection result of the hyperspectral remote sensing image X is obtained.
In specific implementation, the automatic operation of the process can be realized by adopting a software mode. The apparatus for operating the process should also be within the scope of the present invention.
The advantageous effects of the present invention are verified by comparative experiments as follows.
The data adopted in the test are AVIRIS data and Viaregio data, the AVIRIS data has 189 wave bands, the size of the image is 100 pixels multiplied by 100 pixels, and the image has 58 target pixels to be detected in total for 3 airplane targets; the Viareggio data has 511 wave bands in total, the image size is 100 pixels multiplied by 100 pixels, and 3 panel targets in the image have 24 target pixels to be detected in total; . The method adopts a classical adaptive cosine estimation operator ACE (method 1), a sparse representation detection method STD (method 2), a sparse representation detection method SRBBH (method 3) based on binary hypothesis and the method for detecting the target, and takes the method of a specific implementation mode as an example.
Hyperspectral image target detection evaluation indexes: AUC (area under ROC curve) values.
The AUC value is calculated from the area under the ROC (receiver operating characteristic curve) curve. According to a general detection processing process, a detection decision is related to a threshold, under a certain threshold, some real targets are detected, some real targets can be missed to be detected, and some real background pixels are judged as targets, namely false alarms. Therefore, the setting of the threshold value in the detector design process is very important, and it is usually required to achieve a high detection rate while keeping the detection error (including false detection and false alarm) low. Detection rate PdAnd false alarm rate PfThe definition of (A) is:
and
Wherein N isdetectedRepresenting the actual target pixel detected at a given threshold, NtRepresenting real target pixels in the image, NmisBackground pixels, N, representing objects mistakenly divided in the detection resultallAll picture elements in the image are represented. And drawing an ROC curve by taking the detection rate as a vertical coordinate and the false alarm rate as a horizontal coordinate, and obtaining an AUC value of the area under the curve through integration.
TABLE 1 comparative test results
The method of the invention Method 1 Method 2 Method 3
AVIRIS 0.9934 0.4808 0.9479 0.8650
Viareggio 0.9998 0.9805 0.9882 0.9259
As can be seen from Table 1, the method of the present invention can obtain higher AUC values on both sets of data tested, indicating that the method of the present invention has stronger target detection capability. Compared with a basic classical operator such as the method 1, the AUC value of the method is greatly improved, and the target detection precision of the method is much higher than that of the classical operator; compared with methods based on sparse representation such as methods 2 and 3, the AUC value of the method is also higher, which shows that the method can indeed improve the problems existing in the existing methods based on sparse representation.
Therefore, the method provided by the invention has higher target detection precision compared with the existing hyperspectral image target detection method. The invention fully considers the problem of insufficient target training samples in the sparse representation model, and the target pixel on the image can be accurately reconstructed by acquiring the sufficient target training samples. Meanwhile, the construction of the global background dictionary avoids the defect that the traditional double concentric windows are unreasonable in local representation of the image, and can well represent the pixels of all background ground object types on the global image; the dual-category sparse representation method separates the target category and the background category for sparse reconstruction, avoids mutual influence of target and background sparse vectors during reconstruction, and obtains sparse vectors with distinguishing force during reconstruction of two categories of pixels of the target and the background, so that separation of the target and the background in a hyperspectral image is realized, and the target detection effect is improved.
It should be understood that parts of the specification not set forth in detail are well within the prior art.
It should be understood that the above description of the preferred embodiments is given for clarity and not for any purpose of limitation, and that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (5)

1. A single spectrum-driven dual-class sparse representation hyperspectral image target detection method comprises the steps of constructing a target dictionary and a background dictionary, when the target dictionary is constructed, a given target spectrum is used as prior information to pre-detect an original hyperspectral remote sensing image to obtain an initial detection statistical value, the initial detection statistical value is arranged in a descending order, pixels on the hyperspectral image corresponding to a plurality of previous values after the initial detection statistical value is selected as a target training sample, and a target dictionary A is obtainedtThe method is characterized in that: construction of a backDuring the scene dictionary, a principal component analysis method is used for reducing the dimension of the hyperspectral remote sensing image X, K-means clustering is used for classifying the image after dimension reduction, the image elements of each category are sparsely represented, the representing frequency of each image element is obtained, the image element with high representing frequency in each category is selected as a background training sample, and a global over-complete background dictionary A is obtainedb；
For the target dictionary AtAnd background dictionary AbCarrying out sparse representation on pixels by using a two-category sparse representation model to obtain a target sparse vector alphatAnd the background sparse vector alphab(ii) a For the pixel X on the hyperspectral remote sensing image X, the two-class sparse representation model is as follows,
x≈Abαb subject to||αb||0Kb target is absent
x≈Atαt subject to||αt||0Presence of Kt target ≦
kb is a background sparse level, and Kt is a target sparse level; the model separately recovers a target background sparse vector and a background sparse vector, only uses a sample in a target dictionary to recover the target sparse vector, and uses a sample in a background dictionary to recover the background sparse vector;
From a target sparse vector alpha obtained using a two-class sparse representation modeltAnd the background sparse vector alphabAnd detecting the pixels to be detected one by one, and extracting to obtain a target detection result of the hyperspectral remote sensing image X.
2. The single spectrum driven dual-class sparse representation hyperspectral image target detection method of claim 1, wherein: the target sparse vector alpha obtained according to the model using the dual-category sparse representationtAnd the background sparse vector alphabAnd extracting the target detection result of the hyperspectral remote sensing image X, the implementation mode is as follows,
using a target sparse vector alphatAnd the background sparse vector alphabReconstructing the image element to obtain the reconstructed residual errors and the reconstructed weights under two categoriesResidual r obtained by building background pixelb(x) And a residual r obtained by reconstructing the target pixelt(x) (ii) a According to D (x) rb(x)-rt(x) And obtaining an output value D (X) of each pixel X on the hyperspectral remote sensing image X, wherein the larger the output value is, the more likely the pixel is a target pixel, and thus a target detection result of the hyperspectral remote sensing image X is obtained.
3. The single spectrum driven dual-class sparse representation hyperspectral image target detection method of claim 1 or 2, wherein: when the principal component analysis method is used for reducing the dimension of the hyperspectral remote sensing image X, the dimension is preferably reduced to 5 dimensions.
4. The single spectrum driven dual-class sparse representation hyperspectral image target detection method of claim 1 or 2, wherein: and classifying the images after dimensionality reduction by using K-means clustering, wherein the value of p is preferably about 10.
5. The single spectrum driven dual-class sparse representation hyperspectral image target detection method of claim 1 or 2, wherein: when the image elements with high representation frequency in each category are selected as background training samples, the total number N of the suggested background training samples is preferably selectedbThe value of (c) is controlled to be around 300.
CN201910811691.6A 2019-08-30 2019-08-30 Single spectrum driven high-spectrum image target detection method based on double-category sparse representation Active CN110580463B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910811691.6A CN110580463B (en) 2019-08-30 2019-08-30 Single spectrum driven high-spectrum image target detection method based on double-category sparse representation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910811691.6A CN110580463B (en) 2019-08-30 2019-08-30 Single spectrum driven high-spectrum image target detection method based on double-category sparse representation

Publications (2)

Publication Number Publication Date
CN110580463A true CN110580463A (en) 2019-12-17
CN110580463B CN110580463B (en) 2021-07-16

Family

ID=68812256

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910811691.6A Active CN110580463B (en) 2019-08-30 2019-08-30 Single spectrum driven high-spectrum image target detection method based on double-category sparse representation

Country Status (1)

Country Link
CN (1) CN110580463B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111027509A (en) * 2019-12-23 2020-04-17 武汉大学 Hyperspectral image target detection method based on double-current convolution neural network
CN112365490A (en) * 2020-11-25 2021-02-12 重庆邮电大学 Sparse representation hyperspectral image target detection method based on sample oversampling

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103971123A (en) * 2014-05-04 2014-08-06 南京师范大学 Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL)
CN104318243A (en) * 2014-10-14 2015-01-28 西安电子科技大学 Sparse representation and empty spectrum Laplace figure based hyperspectral data dimension reduction method
US20150269441A1 (en) * 2014-03-24 2015-09-24 International Business Machines Corporation Context-aware tracking of a video object using a sparse representation framework
CN106056097A (en) * 2016-08-17 2016-10-26 西华大学 Millimeter wave weak small target detection method
CN107194936A (en) * 2017-05-24 2017-09-22 哈尔滨工业大学 The high spectrum image object detection method represented based on super-pixel joint sparse
CN108229551A (en) * 2017-12-28 2018-06-29 湘潭大学 A kind of Classification of hyperspectral remote sensing image method based on compact dictionary rarefaction representation
CN108564107A (en) * 2018-03-21 2018-09-21 温州大学苍南研究院 The sample class classifying method of semi-supervised dictionary learning based on atom Laplce's figure regularization
CN109584270A (en) * 2018-11-13 2019-04-05 大连大学 Based on the visual tracking method for differentiating dictionary learning

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150269441A1 (en) * 2014-03-24 2015-09-24 International Business Machines Corporation Context-aware tracking of a video object using a sparse representation framework
CN103971123A (en) * 2014-05-04 2014-08-06 南京师范大学 Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL)
CN104318243A (en) * 2014-10-14 2015-01-28 西安电子科技大学 Sparse representation and empty spectrum Laplace figure based hyperspectral data dimension reduction method
CN106056097A (en) * 2016-08-17 2016-10-26 西华大学 Millimeter wave weak small target detection method
CN107194936A (en) * 2017-05-24 2017-09-22 哈尔滨工业大学 The high spectrum image object detection method represented based on super-pixel joint sparse
CN108229551A (en) * 2017-12-28 2018-06-29 湘潭大学 A kind of Classification of hyperspectral remote sensing image method based on compact dictionary rarefaction representation
CN108564107A (en) * 2018-03-21 2018-09-21 温州大学苍南研究院 The sample class classifying method of semi-supervised dictionary learning based on atom Laplce's figure regularization
CN109584270A (en) * 2018-11-13 2019-04-05 大连大学 Based on the visual tracking method for differentiating dictionary learning

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
ZHANG, YUXIANG, ET AL，ET.AL: "A nonlinear sparse representation-based binary hypothesis model for hyperspectral target detection", 《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》 *
ZHU, D.,ET.AL: "Target dictionary construction-based sparse representation hyperspectral target detection methods", 《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》 *
李非燕,等: "基于稀疏表示和自适应模型的高光谱目标检测", 《光学学报》 *
赵春晖，等: "基于字典重构的高光谱图像亚像元目标检测", 《 哈尔滨工程大学学报》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111027509A (en) * 2019-12-23 2020-04-17 武汉大学 Hyperspectral image target detection method based on double-current convolution neural network
CN112365490A (en) * 2020-11-25 2021-02-12 重庆邮电大学 Sparse representation hyperspectral image target detection method based on sample oversampling

Also Published As

Publication number Publication date
CN110580463B (en) 2021-07-16

Similar Documents

Publication Publication Date Title
Liu et al. Tensor matched subspace detector for hyperspectral target detection
Zhang et al. A low-rank and sparse matrix decomposition-based Mahalanobis distance method for hyperspectral anomaly detection
Chang Hyperspectral target detection: Hypothesis testing, signal-to-noise ratio, and spectral angle theories
Du et al. Random-selection-based anomaly detector for hyperspectral imagery
CN107992891B (en) Multispectral remote sensing image change detection method based on spectral vector analysis
CN108564026B (en) Network construction method and system for thyroid tumor cytology smear image classification
Hu et al. Accurate automatic quantification of taxa-specific plankton abundance using dual classification with correction
CN109766858A (en) Three-dimensional convolution neural network hyperspectral image classification method combined with bilateral filtering
CN111027509B (en) Hyperspectral image target detection method based on double-current convolution neural network
Fang et al. A visual attention model combining top-down and bottom-up mechanisms for salient object detection
Wang et al. A sparse representation-based method for infrared dim target detection under sea–sky background
CN110580463B (en) Single spectrum driven high-spectrum image target detection method based on double-category sparse representation
WO2017009812A1 (en) System and method for structures detection and multi-class image categorization in medical imaging
CN106844739B (en) Remote sensing image change information retrieval method based on neural network collaborative training
Jiang et al. LREN: Low-rank embedded network for sample-free hyperspectral anomaly detection
Zhao et al. Center attention network for hyperspectral image classification
CN114862838A (en) Unsupervised learning-based defect detection method and equipment
Yang et al. Dynamic fractal texture analysis for PolSAR land cover classification
Liu et al. Training data assisted anomaly detection of multi-pixel targets in hyperspectral imagery
Chang et al. IBRS: An iterative background reconstruction and suppression framework for hyperspectral target detection
CN115661069A (en) Hyperspectral anomaly detection method and computer device
CN115272861A (en) Subspace sparse representation hyperspectral target detection method based on spectral correlation
CN112883895B (en) Illegal electromagnetic signal detection method based on self-adaptive weighted PCA and realization system thereof
CN107203779A (en) The EO-1 hyperion dimension reduction method kept based on empty spectrum information
CN107273919A (en) A kind of EO-1 hyperion unsupervised segmentation method that generic dictionary is constructed based on confidence level

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