CN108960276B - Sample expansion and consistency discrimination method for improving spectral image supervision classification performance - Google Patents

Sample expansion and consistency discrimination method for improving spectral image supervision classification performance Download PDF

Info

Publication number
CN108960276B
CN108960276B CN201810432942.5A CN201810432942A CN108960276B CN 108960276 B CN108960276 B CN 108960276B CN 201810432942 A CN201810432942 A CN 201810432942A CN 108960276 B CN108960276 B CN 108960276B
Authority
CN
China
Prior art keywords
template
sample
matching
neighborhood
matrix
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
Application number
CN201810432942.5A
Other languages
Chinese (zh)
Other versions
CN108960276A (en
Inventor
肖亮
刘启超
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.)
Nanjing University of Science and Technology
Original Assignee
Nanjing University of Science and Technology
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 Nanjing University of Science and Technology filed Critical Nanjing University of Science and Technology
Priority to CN201810432942.5A priority Critical patent/CN108960276B/en
Publication of CN108960276A publication Critical patent/CN108960276A/en
Application granted granted Critical
Publication of CN108960276B publication Critical patent/CN108960276B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a sample expansion and consistency discrimination method for improving the supervision and classification performance of spectral images, which comprises the following steps: constructing a shape matching template library; calculating a neighborhood similarity matrix of a single training sample; matching the similarity matrix by using a template library, and selecting an optimal matching template according to the matching degree; expanding the training sample according to the best matching template; training a supervised classifier using the extended training set; calculating a neighborhood prediction matrix of the test sample; and matching the neighborhood prediction matrix by using a template library, and calculating the optimal discrimination result according to the consistency measurement. The invention effectively expands the supervised samples through shape template matching, improves the local aggregation of the classification result by utilizing the consistency of the template library and the classification prediction matrix, greatly improves the precision of the supervised classifier, improves the robustness of a small sample supervised classification algorithm, and is suitable for any supervised classifier.

Description

Sample expansion and consistency discrimination method for improving spectral image supervision classification performance
Technical Field
The invention relates to a preprocessing and post-processing technology in a hyperspectral image supervised classification algorithm, in particular to a sample expansion and consistency discrimination method for improving the supervised classification performance of a hyperspectral image.
Technical Field
The remote sensing hyperspectral image is a three-dimensional data set consisting of dozens or even hundreds of continuous spectral bands, and the channel number of each pixel far exceeds that of a common RGB (red, green and blue) color image, so that the remote sensing hyperspectral image has strong resolution performance in surface feature identification and classification application. Therefore, the method has wide application in fields such as agricultural monitoring, pollution monitoring, mineral identification and the like. The identification and classification by utilizing the hyperspectral images are the most important processing processes after the images are acquired, and the identification precision determines the final effect of the hyperspectral images on the identification of the ground objects. In hyperspectral image classification research, machine learning related methods are widely concerned.
Traditional machine learning methods such as extreme learning machine, support vector machine, Bayes and the like are applied to a certain extent in the field of hyperspectral image supervision and classification, but due to the fact that hyperspectral data set samples have strong similarity and the phenomena of same-spectrum foreign matters, same-object different spectrums and the like, the performance of a classifier is reduced to a certain extent, and the final classification accuracy is low. It becomes especially important to improve the classification performance by using the spatial information of the remote sensing image. However, the conventional space-spectrum combination method usually fuses the spatial information and the spectral information of the sample and generates a new sample, for example, the method of increasing the attribute dimension of the sample, such as Gabor-SVM, increases the input dimension of the model, thereby introducing an additional influence caused by dimension expansion. At present, a plurality of methods for improving classification effect by utilizing preprocessing and post-processing technologies on the premise of not changing data dimension are provided, for example, people such as Mucunhong and the like propose a method for optimizing a judgment result of a target sample by utilizing a judgment result of a high-similarity pixel in a neighborhood of a test sample on the basis of an active learning method (Mucunhong, Pyrola, Wang Yimna, and the like, a hyperspectral image classification method combining active learning and neighborhood information, CN 104182767A [ P ] 2014), and a better effect is obtained. However, the determination rule of the method is complex, and the pixel similarity parameter is fixed and cannot be adaptively adjusted according to data, so the application range of the method is limited.
Disclosure of Invention
The invention aims to provide a sample expansion and consistency judgment method for improving the spectral image supervision and classification performance.
The technical solution for realizing the purpose of the invention is as follows: a sample expansion and consistency discrimination method for improving spectral image supervision classification performance comprises the following steps:
firstly, constructing a shape matching template library, namely constructing a plurality of shape matching templates with the size of dxd as a template library;
secondly, calculating a neighborhood similarity matrix of a single training sample, namely intercepting a dxd neighborhood pixel block taking the training sample as a center, calculating the similarity between a neighborhood pixel and the center pixel, and arranging the similarity in sequence into a matrix form to be used as the neighborhood similarity matrix of the training sample;
thirdly, matching the similarity matrix by using a template library, and selecting an optimal matching template according to the matching degree;
fourthly, expanding the training sample according to the optimal matching template, namely classifying the pixels covered by the optimal matching template into the training set of the same type as the central pixels;
fifthly, training a supervision classifier by using the expanded training set;
sixthly, calculating a neighborhood prediction matrix of the test sample, namely intercepting a dXd neighborhood pixel block taking the test sample as a center, respectively predicting the category of each pixel, and sequentially arranging the prediction categories into a matrix form to be used as the neighborhood prediction matrix of the sample;
and seventhly, matching the neighborhood prediction matrix by using the template base, calculating an optimal discrimination result according to the consistency measurement, namely sequentially calculating the consistency measurement of each category corresponding to the matching of the neighborhood prediction matrix and each template in the template base, and selecting the category corresponding to the maximum value as a final discrimination result of the sample.
Compared with the prior art, the invention has the remarkable characteristics that: (1) the supervised sample can be effectively expanded through shape template matching, so that the supervised learning performance of the small sample is effectively improved; (2) the accuracy of the final classification result can be greatly improved by judging the optimal consistency of the neighborhood prediction result; (3) the method is suitable for any supervision classifier and has wide applicability; (4) the method is insensitive to the characteristics of the data, so that the method has better self-adaptability and expansibility, improves the problems of same-object different spectrum and same-spectrum foreign matter, and greatly improves the final classification precision of the algorithm.
Drawings
FIG. 1 is a flowchart of a sample expansion and consistency determination method for improving the supervised classification performance of spectral images according to the present invention.
FIG. 2 is an exemplary diagram of a 3 × 3 match template library.
FIG. 3 is an exemplary diagram of a matching template library of size 5 × 5.
FIG. 4(a) is a plot of the actual terrain profile for the Indian Pines dataset.
FIG. 4(b) is a diagram illustrating the classification effect of Indian Pines data set by ELM method.
FIG. 4(c) is a diagram showing the effect of classification of Indian Pines data sets using the ELM method in conjunction with the method of the present invention (using a matching template of size 3X 3).
FIG. 4(d) is a diagram showing the effect of classification of Indian Pines data sets using the ELM method in conjunction with the method of the present invention (using matching templates of size 5X 5).
FIG. 4(e) is a diagram illustrating the classification effect of Indian Pines data set by SVM method.
FIG. 4(f) is a graph of the classification effect of Indian Pines data sets using SVM with the method of the present invention (using matching templates of size 3X 3).
FIG. 4(g) is a graph of the classification effect of Indian Pines data sets using SVM with the method of the present invention (using matching templates of size 5X 5).
Fig. 5(a) is a true ground object distribution plot for the Pavia Center dataset.
Fig. 5(b) is a graph of the classification effect of the Pavia Center dataset by the ELM method.
Fig. 5(c) is a graph of the classification effect of Pavia Center dataset using the ELM method in combination with the method of the present invention (using a matching template of size 3 × 3).
FIG. 5(d) is a graph of the effect of classification of the Pavia Center dataset using the ELM method in conjunction with the method of the present invention (using a matching template of size 5X 5).
Fig. 5(e) is a classification effect diagram of Pavia Center dataset using SVM method.
Fig. 5(f) is a graph of the classification effect of Pavia Center dataset using SVM method in combination with the method of the present invention (using matching templates of size 3 × 3).
Fig. 5(g) is a graph of the classification effect of Pavia Center dataset using SVM method in combination with the method of the present invention (using matching templates of size 5 × 5).
Detailed Description
With reference to fig. 1, a sample expansion and consistency discrimination method for improving the supervised classification performance of spectral images includes the following steps:
firstly, constructing a shape matching template library, namely constructing a plurality of shape matching templates with the size of dxd as a template library;
secondly, calculating a neighborhood similarity matrix of a single training sample, namely intercepting a dxd neighborhood pixel block taking the training sample as a center, calculating the similarity between a neighborhood pixel and the center pixel, and arranging the similarity in sequence into a matrix form to be used as the neighborhood similarity matrix of the training sample;
thirdly, matching the similarity matrix by using a template library, and selecting an optimal matching template according to the matching degree;
fourthly, classifying the pixels covered by the best matching template into a training set of the same type as the central pixels;
fifthly, training a supervision classifier by using the expanded training set;
sixthly, calculating a neighborhood prediction matrix of the test sample, namely selecting a dxd neighborhood pixel block taking the test sample as a center, predicting the category of each pixel, and sequentially arranging the prediction categories into a matrix form to be used as the neighborhood prediction matrix of the sample;
and seventhly, matching the neighborhood prediction matrix by using the template base, calculating an optimal discrimination result according to the consistency measurement, namely sequentially calculating the consistency measurement of each category corresponding to the matching of the neighborhood prediction matrix and each template in the template base, and selecting the category corresponding to the maximum value as a final discrimination result of the sample.
Further, constructing L shape matching templates with the size of dxd as a template library, wherein L is more than or equal to 8, d is more than or equal to 3, and d is an odd number; the basic principle for constructing a matching template is as follows:
(1) dividing all pixels into covered pixels and uncovered pixels by taking pixel points as units in a template, wherein the covered pixels take 1 and the uncovered pixels take 0;
(2) the pixel regions covered in the template need to satisfy connectivity, i.e. pixels with a median value of 1 in the template can form a connected region satisfying 4-adjacency, and the connected region needs to cover a central point;
(3) the number n of coverage area points in the template must satisfy n ═ d-12。
Further, the second step calculates a neighborhood similarity matrix of a single training sample, and the specific process is as follows:
(1) for the ith training sample xiIntercepting pixel x in hyperspectral imageiThe D × D neighborhood block of pixels as the center is marked as DiExpressed as follows:
Figure BDA0001653901680000041
i is more than or equal to 1 and less than or equal to N, and N is the number of training samples;
(2) sequentially calculating central pixel xiAnd neighborhood pixel block DiIn (2)
Figure BDA0001653901680000042
The similarity of (2) is recorded as
Figure BDA0001653901680000043
Wherein e is more than or equal to 1 and less than or equal to d2The calculation method comprises the following steps:
Figure BDA0001653901680000044
(3) center pixel xiAnd neighborhood pixel
Figure BDA0001653901680000045
Degree of similarity of
Figure BDA0001653901680000046
Putting the training samples into a matrix form according to the sequence, taking the matrix as a neighborhood similarity matrix of the training samples and recording the matrix as SiExpressed as follows:
Figure BDA0001653901680000051
further, the third step matches the similarity matrix by using a template library, and selects an optimal matching template according to the matching degree, wherein the calculation method comprises the following steps:
(1) calculating the first template and training sample x in the template library in sequenceiNeighborhood similarity matrix S ofiDegree of matching of (D) is recorded as
Figure BDA0001653901680000052
The calculation formula is as follows:
Figure BDA0001653901680000053
wherein the operator
Figure BDA0001653901680000054
Representing a convolution operation, TlIs the first template in the template library, SiIs a sample xiL is more than or equal to 1 and less than or equal to L of the corresponding neighborhood similarity matrix.
(2) At sample xiMatching degree obtained by matching with L templates
Figure BDA0001653901680000055
In the method, the template corresponding to the maximum value is selected as the best matching template and recorded as the best matching template
Figure BDA0001653901680000056
The calculation formula is as follows:
Figure BDA0001653901680000057
further, the fourth step is to expand the training sample according to the best matching template, i.e. the training sample xiCorresponding best matching template
Figure BDA0001653901680000058
Pixels at the same position in the neighborhood pixel block corresponding to the covered area are classified to the central pixel xiThe training of the same category is centralized.
And further, a fifth step of training and supervising the classifier by using the expanded training set, namely, matching and expanding all original training samples according to the second step to the fourth step in sequence, and training the supervised classifier by using the expanded training samples.
Further, the sixth step calculates a neighborhood prediction matrix of the test sample, and the calculation method is as follows:
(1) for the jth test sample yjIntercepting pixel y in hyperspectral imagejA centered d × d neighborhood block, denoted as KjExpressed as follows:
Figure BDA0001653901680000059
j is more than or equal to 1 and less than or equal to M, and M is the number of test samples;
(2) sequentially aligning primitive blocks K using trained classifiersjIn (2)
Figure BDA00016539016800000510
The result of the discrimination is recorded as
Figure BDA00016539016800000511
Wherein e is more than or equal to 1 and less than or equal to d2And putting the discrimination results in a matrix form in sequence as the test sample yjIs the neighborhood prediction matrix of (1), denoted as QjExpressed as follows:
Figure BDA0001653901680000061
further, the seventh step matches the neighborhood prediction matrix by using a template library, and calculates the best discrimination result according to the consistency measurement, which comprises the following specific processes:
(1) calculating the test sample y in turnjWith the first template T in the template librarylConsistency measure under match belonging to class c, denoted
Figure BDA0001653901680000062
The calculation method comprises the following steps:
Figure BDA0001653901680000063
Figure BDA0001653901680000064
wherein the operator
Figure BDA0001653901680000065
Representing the Hadamard product, QjFor testing a sample yjThe neighborhood prediction matrix of (a) is,
Figure BDA0001653901680000066
representation matrix
Figure BDA0001653901680000067
The number of the middle elements is equal to the number of C, C is more than or equal to 1 and less than or equal to C, and C is the number of sample categories.
(2) In the sample yjSelecting the category corresponding to the maximum value of the consistency measurement as the final judgment result and recording the final judgment result as the final judgment result
Figure BDA0001653901680000068
The calculation formula is as follows:
Figure BDA0001653901680000069
wherein the content of the first and second substances,
Figure BDA00016539016800000610
is a sample yjAnd C is the number of sample classes, and L is the number of templates.
The present invention will be described in detail with reference to the following examples and drawings.
Examples
With reference to fig. 1, a sample expansion and consistency discrimination method for improving the supervised classification performance of spectral images includes the following steps:
firstly, a shape matching template library is constructed, namely L shape matching templates with the size of dxd are constructed as the template library, wherein L is more than or equal to 8, d is more than or equal to 3, and d is an odd number. In this embodiment, d is 3, L is 16, and 16 different matching templates satisfying the condition are designed as a template library, which is denoted as T1,T2,…,T16The template library is shown in FIG. 2. Each template is in the form of a matrixFirst template T depicted in FIG. 21Is represented as follows:
Figure BDA0001653901680000071
similarly, other templates are also in matrix representation form. The form of the matching template library with the size of 5 × 5 is shown in fig. 3, and the template design mode is designed according to the aforementioned rules.
Secondly, calculating a neighborhood similarity matrix of a single training sample, wherein the specific process is as follows under the conditions that d is 3 and L is 16:
(1) for the ith training sample xiIntercepting pixel x in hyperspectral imagei3 × 3 neighborhood pixel block as center, marked as DiExpressed as follows:
Figure BDA0001653901680000072
i is more than or equal to 1 and less than or equal to N, and N is the number of training samples;
(2) sequentially calculating central pixel xiAnd neighborhood pixel block DiIn (2)
Figure BDA0001653901680000073
The similarity of (2) is recorded as
Figure BDA0001653901680000074
Wherein e is more than or equal to 1 and less than or equal to 9. The calculation process is as follows:
Figure BDA0001653901680000075
(3) center pixel xiAnd neighborhood pixel
Figure BDA0001653901680000076
Degree of similarity of
Figure BDA0001653901680000077
Put in a matrix form in order to doThe neighborhood similarity matrix for the training sample, denoted SiExpressed as follows:
Figure BDA0001653901680000078
and thirdly, matching the similarity matrix by using a template library, and selecting the best matching template according to the matching degree. Under the condition that d is 3 and L is 16, the specific calculation process is as follows:
(1) calculating the first template and sample x in the template library in turniNeighborhood similarity matrix S ofiDegree of matching of (D) is recorded as
Figure BDA0001653901680000079
The calculation process is as follows:
Figure BDA00016539016800000710
wherein the operator
Figure BDA00016539016800000711
Representing a convolution operation, T1,T2,…,T16As templates in a template library, SiIs a sample xiAnd l is more than or equal to 1 and less than or equal to 16 of the corresponding neighborhood similarity matrix.
(2) At sample xiAll corresponding template matching degrees
Figure BDA00016539016800000712
Selecting the template corresponding to the maximum value as the best matching template, and recording as the best matching template
Figure BDA0001653901680000081
The calculation formula is as follows:
Figure BDA0001653901680000082
suppose that
Figure BDA0001653901680000083
Then sample xiThe best matching template is T1。
The fourth step, expand the training sample according to the best matching template, i.e. sample xiCorresponding best matching template
Figure BDA0001653901680000084
Pixels at the same position in the neighborhood pixel block corresponding to the covered (value 1) region are classified as central pixel xiThe training of the same category is centralized. Suppose a pixel xiThe best matching template of is T1Then will be
Figure BDA0001653901680000085
Is added toiThe training of the same kind is centralized.
And fifthly, training and supervising the classifier by using the expanded training set, namely sequentially matching and expanding all the training samples according to the second step to the fourth step, and training the supervised classifier by using the expanded training samples, wherein the scoring classifier is q (classify) (y), wherein y is the testing sample, and q is the class obtained by the testing sample y through the classifier.
And sixthly, calculating a neighborhood prediction matrix of the test sample, wherein under the assumption that d is 3 and L is 16, the specific process is as follows:
(1) for the jth test sample yjIntercepting pixel y in hyperspectral imagej3 × 3 neighborhood pixel block as center, marked as KjExpressed as follows:
Figure BDA0001653901680000086
j is more than or equal to 1 and less than or equal to M, and M is the number of test samples;
(2) sequentially aligning primitive blocks K using trained classifiersjIn (2)
Figure BDA0001653901680000087
The result of the discrimination is recorded as
Figure BDA0001653901680000088
Wherein e is more than or equal to 1 and less than or equal to 9, and the discrimination results are put into a matrix form according to the sequence and used as the test sample yjIs the neighborhood prediction matrix of (1), denoted as QjThe calculation method comprises the following steps:
Figure BDA0001653901680000089
Figure BDA00016539016800000810
seventhly, matching the neighborhood prediction matrix by using a template library, and calculating an optimal judgment result according to the consistency measurement, wherein under the conditions that d is 3, L is 16, and the sample category number C is 9, the specific process is as follows:
(1) calculating the test sample y in turnjWith the first template T in the template librarylConsistency measure under match belonging to class c, denoted
Figure BDA0001653901680000091
The calculation process is as follows:
Figure BDA0001653901680000092
Figure BDA0001653901680000093
Figure BDA0001653901680000094
Figure BDA0001653901680000095
Figure BDA0001653901680000096
wherein the operator ° represents the hadamard product,
Figure BDA0001653901680000097
representation matrix
Figure BDA0001653901680000098
The number of the medium elements is equal to c, and c is more than or equal to 1 and less than or equal to 9.
(2) In the sample yjSelecting the category corresponding to the maximum value of the consistency measurement as the final judgment result and recording the final judgment result as the final judgment result
Figure BDA0001653901680000099
The calculation formula is as follows:
Figure BDA00016539016800000910
wherein
Figure BDA00016539016800000911
Is a sample yjThe consistency metrics belong to 9 categories under 16 template matches respectively. Suppose that
Figure BDA00016539016800000912
Then the test sample yjIs the best matching template T, and belongs to class 71。
The effect of the present invention is illustrated by simulation experiments as follows:
the simulation experiment adopts two groups of real hyperspectral data: indian Pines dataset and Pavia Center dataset. The Indian Pines dataset is a hyperspectral remote sensing image acquired by an airborne visible infrared imaging spectrometer (AVIRIS) in an Indian Pines experimental area, indiana, usa. The image contains 220 bands in total, the spatial resolution is 20m, and the image size is 145 × 145. After removing 20 water vapor absorption and low signal-to-noise ratio bands (band numbers 104-. The area contains 10366 samples of 16 known land features. The Pavia Center dataset was acquired by a ross sensor in paviia, and included 115 bands in total, with an image size of 1096 × 490, and after removing the noise band, the remaining 102 bands were selected as the study subjects. 10% of each sample is randomly taken as a training sample, the other 90% of the samples are taken as a test sample, ten times of experiments are respectively carried out to calculate an average result, and OA (overhead accuracy) and AA (average accuracy) evaluation indexes are given. Before the experiment, both sets of data were smoothed using MH prediction method. The simulation experiments are all completed by adopting Python3.6 and MATLAB R2016a under a Windows 10 operating system.
The evaluation indexes adopted by the invention are conventional precision evaluation methods, including average precision (AA) and overall precision (OA). As the invention is used for improving the classification accuracy of any classifier, an ELM and an SVM are respectively adopted as basic classifiers for experiments in simulation experiments, in order to show the improvement of the method on the performance of the classifier, each group of data sets are subjected to experiments by using the ELM, the TM-ELM (3x3), the TM-ELM (5x5), the SVM, the TM-SVM (3x3) and the TM-SVM (5x5), and the results are compared. Wherein TM-ELM (3x3) represents the classification method using template matching (template matching) of 3x3 size in combination with ELM, TM-ELM (5x5) represents the classification method using template matching of 5x5 size in combination with ELM, and TM-SVM (3x3) and TM-SVM (5x5) represent similar meanings. A template library pattern of size 3x3 is shown in fig. 2, and a template library pattern of size 5x5 is shown in fig. 3.
Table 1 shows the classification accuracy (%) of simulation experiments performed on Indian Pines and Pavia Center datasets by the method of the present invention.
TABLE 1
Figure BDA0001653901680000101
From experimental results, on the premise of not changing the original ELM and SVM classification models, the classification precision can be greatly improved only by using the method to perform training sample expansion on data and judging the optimal consistency of the prediction result. On an Indian Pines data set, the method can improve the overall accuracy of 96.27% of an original ELM model to 98.85% (by using a 5x5 size template matching method) and improve the overall accuracy of 90.33% of the original SVM model to 98.99% (by using a 5x5 size template matching method). On the PaviaCenter data set, the method can improve the classification accuracy of the original ELM model to 98.08% (by using a 5x5 size template matching method) and improve the overall accuracy of the original SVM model to 99.79% (by using a 5x5 size template matching method). Obviously, after the method of the invention uses the 3x3 template and the 5x5 template for matching, the classification result is obviously improved, and the classification precision is improved along with the expansion of the size of the template, thus fully showing that the method of the invention has higher stability and expansibility. The result effect graphs of the method of the invention on two sets of data sets are shown in fig. 4 and fig. 5. The simulation experiment results of the two groups of real data sets show the effectiveness of the method.

Claims (7)

1. A sample expansion and consistency discrimination method for improving spectral image supervision classification performance is characterized by comprising the following steps:
firstly, constructing a shape matching template library, namely constructing a plurality of shape matching templates with the size of dxd as a template library;
secondly, calculating a neighborhood similarity matrix of a single training sample, namely intercepting a dxd neighborhood pixel block taking the training sample as a center, calculating the similarity between a neighborhood pixel and the center pixel, and arranging the similarity in sequence into a matrix form to be used as the neighborhood similarity matrix of the training sample;
thirdly, matching the similarity matrix by using a template library, and selecting an optimal matching template according to the matching degree; the calculation method comprises the following steps:
(1) calculating the first template and training sample x in the template library in sequenceiNeighborhood similarity matrix S ofiDegree of matching of (D) is recorded as
Figure FDA0003237626820000011
The calculation formula is as follows:
Figure FDA0003237626820000012
wherein the operator
Figure FDA0003237626820000013
Representing a convolution operation, TlIs the first template in the template library, SiIs a sample xiL is more than or equal to 1 and less than or equal to L of the corresponding neighborhood similarity matrix;
(2) at sample xiMatching degree obtained by matching with L templates
Figure FDA0003237626820000014
In the method, the template corresponding to the maximum value is selected as the best matching template and recorded as the best matching template
Figure FDA0003237626820000015
The calculation formula is as follows:
Figure FDA0003237626820000016
fourthly, classifying the pixels covered by the best matching template into a training set of the same type as the central pixels;
fifthly, training a supervision classifier by using the expanded training set;
sixthly, calculating a neighborhood prediction matrix of the test sample, namely selecting a dxd neighborhood pixel block taking the test sample as a center, predicting the category of each pixel, and sequentially arranging the prediction categories into a matrix form to be used as the neighborhood prediction matrix of the sample;
and seventhly, matching the neighborhood prediction matrix by using the template base, calculating an optimal discrimination result according to the consistency measurement, namely sequentially calculating the consistency measurement of each category corresponding to the matching of the neighborhood prediction matrix and each template in the template base, and selecting the category corresponding to the maximum value as a final discrimination result of the sample.
2. The method for sample expansion and consistency discrimination for improving the spectral image supervision and classification performance according to claim 1 is characterized in that in the first step, L shape matching templates with the size of dxd are constructed as a template library, wherein L is more than or equal to 8, d is more than or equal to 3, and d is an odd number; the basic principle for constructing a matching template is as follows:
(1) dividing all pixels into covered pixels and uncovered pixels by taking pixel points as units in a template, wherein the covered pixels take 1 and the uncovered pixels take 0;
(2) the pixel regions covered in the template need to satisfy connectivity, i.e. pixels with a median value of 1 in the template can form a connected region satisfying 4-adjacency, and the connected region needs to cover a central point;
(3) the number n of coverage area points in the template must satisfy n ═ d-12。
3. The method for sample expansion and consistency discrimination for improving the spectral image supervision and classification performance according to claim 1 is characterized in that a neighborhood similarity matrix of a single training sample is calculated in the second step, and the specific process is as follows:
(1) for the ith training sample xiIntercepting pixel x in hyperspectral imageiThe D × D neighborhood block of pixels as the center is marked as DiExpressed as follows:
Figure FDA0003237626820000021
wherein i is more than or equal to 1 and less than or equal to N, and N is the number of training samples;
(2) sequentially calculating central pixel xiAnd neighborhood pixel block DiIn (2)
Figure FDA0003237626820000022
The similarity of (2) is recorded as
Figure FDA0003237626820000023
Wherein e is more than or equal to 1 and less than or equal to d2The calculation method comprises the following steps:
Figure FDA0003237626820000024
(3) center pixel xiAnd neighborhood pixel
Figure FDA0003237626820000025
Degree of similarity of
Figure FDA0003237626820000026
Putting the training samples into a matrix form according to the sequence, taking the matrix as a neighborhood similarity matrix of the training samples and recording the matrix as SiExpressed as follows:
Figure FDA0003237626820000027
4. the method of claim 1, wherein the fourth step of expanding the training samples according to the best matching template, i.e. the training sample xiCorresponding best matching template
Figure FDA0003237626820000028
Pixels at the same position in the neighborhood pixel block corresponding to the covered area are classified to the central pixel xiThe training of the same category is centralized.
5. The method for sample expansion and consistency discrimination to improve the performance of spectral image supervised classification as recited in claim 1, wherein the fifth step is to train the supervised classifier using the expanded training set, i.e. all original training samples are subjected to matching expansion in sequence from the second step to the fourth step, and the supervised classifier is trained using the expanded training samples.
6. The method for sample expansion and consistency discrimination for improving the spectral image supervision and classification performance according to claim 1, wherein the sixth step is to calculate a neighborhood prediction matrix of the test sample, and the calculation method is as follows:
(1) for the jth test sample yjIntercepting pixel y in hyperspectral imagejA centered d × d neighborhood block, denoted as KjExpressed as follows:
Figure FDA0003237626820000031
wherein j is more than or equal to 1 and less than or equal to M, and M is the number of test samples;
(2) sequentially aligning primitive blocks K using trained classifiersjIn (2)
Figure FDA0003237626820000032
The result of the discrimination is recorded as
Figure FDA0003237626820000033
Wherein e is more than or equal to 1 and less than or equal to d2And putting the discrimination results in a matrix form in sequence as the test sample yjIs the neighborhood prediction matrix of (1), denoted as QjExpressed as follows:
Figure FDA0003237626820000034
7. the method for sample expansion and consistency discrimination for improving the spectral image supervision and classification performance according to claim 1, wherein in the seventh step, a template library is used for matching the neighborhood prediction matrix, and the optimal discrimination result is calculated according to consistency measurement, and the specific process is as follows:
(1) calculating the test sample y in turnjWith the first template T in the template librarylConsistency measure under match belonging to class c, denoted
Figure FDA0003237626820000035
The calculation method comprises the following steps:
Figure FDA0003237626820000036
Figure FDA0003237626820000037
wherein the operator ° represents the Hadamard product, QjFor testing a sample yjThe neighborhood prediction matrix of (a) is,
Figure FDA0003237626820000038
representation matrix
Figure FDA0003237626820000039
C is more than or equal to 1 and less than or equal to C, and C is the number of sample categories;
(2) in the sample yjSelecting the category corresponding to the maximum value of the consistency measurement as the final judgment result and recording the final judgment result as the final judgment result
Figure FDA00032376268200000310
The calculation formula is as follows:
Figure FDA00032376268200000311
wherein the content of the first and second substances,
Figure FDA0003237626820000041
is a sample yjAnd respectively belonging to the consistency measurement sets of the categories under the matching of the templates.
CN201810432942.5A 2018-05-08 2018-05-08 Sample expansion and consistency discrimination method for improving spectral image supervision classification performance Active CN108960276B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810432942.5A CN108960276B (en) 2018-05-08 2018-05-08 Sample expansion and consistency discrimination method for improving spectral image supervision classification performance

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810432942.5A CN108960276B (en) 2018-05-08 2018-05-08 Sample expansion and consistency discrimination method for improving spectral image supervision classification performance

Publications (2)

Publication Number Publication Date
CN108960276A CN108960276A (en) 2018-12-07
CN108960276B true CN108960276B (en) 2022-02-15

Family

ID=64499024

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810432942.5A Active CN108960276B (en) 2018-05-08 2018-05-08 Sample expansion and consistency discrimination method for improving spectral image supervision classification performance

Country Status (1)

Country Link
CN (1) CN108960276B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110008335A (en) * 2018-12-12 2019-07-12 阿里巴巴集团控股有限公司 The method and device of natural language processing
CN111241969A (en) * 2020-01-06 2020-06-05 北京三快在线科技有限公司 Target detection method and device and corresponding model training method and device
CN112906666A (en) * 2021-04-07 2021-06-04 中国农业大学 Remote sensing identification method for agricultural planting structure
CN113252585B (en) * 2021-04-21 2024-02-09 深圳市海谱纳米光学科技有限公司 Method and device for judging gold surface coating based on hyperspectral image

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103903007B (en) * 2014-03-10 2017-02-22 哈尔滨工程大学 Hyperspectral semi-supervised classification method based on space-spectral information
CN104318241A (en) * 2014-09-25 2015-01-28 东莞电子科技大学电子信息工程研究院 Local density spectral clustering similarity measurement algorithm based on Self-tuning
US10078780B2 (en) * 2015-03-27 2018-09-18 Intel Corporation Gesture recognition mechanism
CN107067413B (en) * 2016-12-27 2019-11-15 南京理工大学 A kind of moving target detecting method of time-space domain statistical match local feature
CN107491734B (en) * 2017-07-19 2021-05-07 苏州闻捷传感技术有限公司 Semi-supervised polarimetric SAR image classification method based on multi-core fusion and space Wishart LapSVM

Also Published As

Publication number Publication date
CN108960276A (en) 2018-12-07

Similar Documents

Publication Publication Date Title
CN110399909B (en) Hyperspectral image classification method based on label constraint elastic network graph model
CN108960276B (en) Sample expansion and consistency discrimination method for improving spectral image supervision classification performance
CN107292339B (en) Unmanned aerial vehicle low-altitude remote sensing image high-resolution landform classification method based on feature fusion
CN108491849B (en) Hyperspectral image classification method based on three-dimensional dense connection convolution neural network
Es-saady et al. Automatic recognition of plant leaves diseases based on serial combination of two SVM classifiers
US10984291B2 (en) End-member extraction method based on segmented vertex component analysis (VCA)
CN106503739A (en) The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics
Song et al. Hyperspectral image classification based on KNN sparse representation
CN107832797B (en) Multispectral image classification method based on depth fusion residual error network
Almogdady et al. A flower recognition system based on image processing and neural networks
CN108197650B (en) Hyperspectral image extreme learning machine clustering method with local similarity maintained
CN104299232B (en) SAR image segmentation method based on self-adaptive window directionlet domain and improved FCM
CN106023257A (en) Target tracking method based on rotor UAV platform
CN112990314B (en) Hyperspectral image anomaly detection method and device based on improved isolated forest algorithm
CN108427913A (en) The Hyperspectral Image Classification method of combined spectral, space and hierarchy information
KR101941043B1 (en) Method for Object Detection Using High-resolusion Aerial Image
CN105405138A (en) Water surface target tracking method based on saliency detection
CN110991493A (en) Hyperspectral anomaly detection method based on collaborative representation and anomaly elimination
CN103390170A (en) Surface feature type texture classification method based on multispectral remote sensing image texture elements
CN113673556A (en) Hyperspectral image classification method based on multi-scale dense convolution network
Singh et al. A hybrid approach for information extraction from high resolution satellite imagery
Elmannai et al. Classification using semantic feature and machine learning: Land-use case application
CN109472785B (en) Quick identification method for damaged building caused by earthquake damage based on multiple structural elements
Talasila et al. PLRSNet: a semantic segmentation network for segmenting plant leaf region under complex background
Shi et al. Urban feature shadow extraction based on high-resolution satellite remote sensing images

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