CN104268579A - Hyperspectral remote sensing image classifying method based on hierarchy ensemble learning - Google Patents

Abstract

The invention discloses a hyperspectral remote sensing image classification method based on layered ensemble learning, which belongs to the technical field of hyperspectral remote sensing image classification. The invention aims to solve the problem of low classification accuracy of hyperspectral remote sensing image data. It mainly uses a two-layer integrated structure to classify hyperspectral images, which are the inner layer structure and the outer layer structure; the inner layer structure is to form a spectral set with differences through random band selection; then use the spectral set as a unit, respectively use The integrated method of Adaboost is used to train, and then classify the test samples; the outer structure is to integrate the classification results of each spectral collection in the inner layer integration, and use the method of weight voting to determine the final category of the sample; finally, the whole image is used as Test the sample and realize the classification of the whole image to obtain the classified theme map. The invention is used for classifying hyperspectral remote sensing images.

Description

Hyperspectral Remote Sensing Image Classification Method Based on Hierarchical Ensemble Learning

technical field

The invention relates to a hyperspectral remote sensing image classification method based on hierarchical integrated learning, and belongs to the technical field of hyperspectral remote sensing image classification.

Background technique

With the improvement of remote sensing data acquisition methods, hyperspectral remote sensing images have become the key research content due to their unique advantages. Hyperspectral remote sensing can acquire many very narrow and spectrally continuous image data in the ultraviolet, visible, near-infrared and mid-infrared regions of the electromagnetic spectrum, and has the characteristics of combining images and spectra. This has important research value and application significance for using remote sensing images for target classification, target recognition, and target tracking. At present, the classification of target objects using hyperspectral remote sensing data is one of the hot applications of hyperspectral imaging. Therefore, the research on hyperspectral data classification methods has profound and long-term significance.

The current classification methods for hyperspectral remote sensing images cannot overcome the small number of samples and the influence of noise and singular values, which makes the classification accuracy of image data low.

Contents of the invention

The object of the present invention is to solve the problem of low classification accuracy of hyperspectral remote sensing image data, and provide a hyperspectral remote sensing image classification method based on layered ensemble learning.

The hyperspectral remote sensing image classification method based on layered integrated learning of the present invention, it comprises the following steps:

Step 1: Read the hyperspectral raw data, preprocess it, obtain supervised data, then determine the labeled samples according to the supervised data, and select training samples and test samples from the labeled samples;

Step 2: Perform random band selection for all spectral bands of the hyperspectral raw data, and then extract a pixel vector combination from each marked sample, the number of bands contained in each pixel vector in the pixel vector combination is the same as that of the random The number of bands corresponds one by one to form a spectral set; repeat the random band selection process and form a corresponding spectral set to form multiple spectral sets with differences;

Step 3: Using each spectral collection as a unit, use the training samples to train the inner layer structure of the Adaboost integration framework, and then classify the test samples to obtain the inner layer integration composed of the Adaboost framework;

Step 4: Construct the outer structure of the Adaboost integration framework from the classification results and classification accuracy of the test samples in step 3;

Step 5: Use the Adaboost integration framework with the above-mentioned inner structure and outer structure to classify the hyperspectral raw data in a traversal manner, and obtain the classification theme map.

In the first step, the marked samples are determined, and the specific method of selecting training samples and test samples from the marked samples is as follows:

Read the hyperspectral raw data stored in the form of a three-dimensional matrix, the three-dimensional matrix form includes two-dimensional spatial position information and one-dimensional spectral information; The object categories in the marker map are respectively marked in the form of integer values, and the supervision data in the form of a two-dimensional matrix is obtained. composition;

From the hyperspectral raw data and supervisory data, determine the number of feature categories C in the real feature marker map, the number of pixels in the two-dimensional spatial position information M rows × N columns, and the number of effective bands in the one-dimensional spectral information B;

The spatial position data of pixels in the supervisory data is extracted from top to bottom and from left to right in the original hyperspectral data according to the position coordinates, and the pixel vector of spectral information is extracted as a marked sample, and the marked samples are arranged in two-dimensional rows In matrix form, the number of rows of the two-dimensional matrix is the number of marked samples, and the number of columns is the number of bands B contained in the pixel vector;

The odd-numbered rows of the marked samples expressed in a two-dimensional matrix form are selected as training samples, and the even-numbered rows are selected as test samples, and the ratio of the number of training samples to the number of test samples is 1:1;

Then extract the feature category data corresponding to the marked samples expressed in the two-dimensional matrix form from the supervised data, and arrange them into a column vector. The number of feature category data in the column vector is the number of marked samples, and each The value of the feature category data is the category label of a corresponding labeled sample, the odd-numbered element in the column vector is used as the category label of the training sample, and the even-numbered element is used as the category label of the test sample.

The specific method for forming multiple spectral sets with differences in step 2 is as follows:

From 1 to B, a group of b random numbers {i ₁ , i ₂ ,…,i _b } is drawn, where 1≤i _k ≤B, k=1,2,…,b), and then In each marked sample, the i _kth vector element is extracted in the order of k from 1 to b to form a pixel vector combination containing b elements respectively, and the number of the pixel vector combination is the same as the number of marked samples , all pixel vectors containing b elements are combined as a spectral set;

The process of forming a spectrum set is repeated multiple times, and b random numbers are selected in the repeated process, and each combination of random numbers is different, so as to obtain multiple spectrum sets with differences.

The specific method to obtain the inner layer integration composed of the Adaboost framework in step 3 is:

Taking each spectral set as a unit, select the support vector machine SVM as the weak classifier of the Adaboost integration framework; for the first spectral set, iterate the combination of pixel vectors corresponding to the training samples in the spectral set according to the integration method of Adaboost , set the number of iterations as F, then form a series of weak classifiers f(f=1,2,...,

First, assign the same weight to each pixel vector combination corresponding to the training sample in the first spectral set, and then use the weak classifier 1 to train and test the current pixel vector weight combination, for the misclassified pixel vector Increase its weight, and reduce its weight for the pixel vector of correct decision; the pixel vector weight combination after weight adjustment is used to train the weak classifier 2 on the current pixel vector weight combination, and so on, iterate F times repeatedly, and obtain the focus on F weak classifiers with different combinations of pixel vector weights;

Taking each spectral set as a unit, first use F weak classifiers to classify each pixel vector corresponding to the test sample in the first spectral set, and obtain F classification results for each pixel vector, thus obtaining F classification results of the corresponding test samples, and then the final classification results of the corresponding test samples are determined by majority voting;

Set a total of T spectral sets, repeat the above classification process for the second spectral set to the Tth spectral set, and the final classification results of the test samples constitute T internal integrations composed of the Adaboost framework.

The specific method of constructing the outer structure of the Adaboost integrated framework in step 4 is:

Integrate the final classification results of the test samples obtained by T internal integrations, and then use the method of weight voting to determine the final category of the integration results, and construct the outer structure of the Adaboost integration framework.

The specific method of obtaining the classification theme map in step five is:

Using the traversal method from top to bottom and from left to right, the pixel vectors in the same method as above are extracted for each pixel in the hyperspectral original data spatial position information, and M×N pixel vectors are obtained, and then the above-mentioned The Adaboost integrated framework of the inner structure and the outer structure classifies M×N pixel vectors one by one to obtain M×N category labels, and then converts the M×N category labels into corresponding two-dimensional arrays of M rows×N columns Matrix, the two-dimensional matrix of M rows×N columns is displayed as a two-dimensional image to obtain a classification theme map.

Advantages of the present invention: the method of the present invention aims at the high dimensionality of hyperspectral images and the data characteristics of noise and outliers, and proposes a layered-based integrated learning method, which effectively reduces the number of samples and the impact of noise and singular values on hyperspectral images. Influenced by the accuracy of image classification, this method can not only obtain higher classification accuracy, but also provide a more intuitive classification theme map, which can better serve subsequent image processing applications.

According to the inherent characteristics of hyperspectral data, the method of the present invention randomly selects bands for hyperspectral data, and at the same time classifies hyperspectral data using a layered integrated learning method, thereby establishing a data classification algorithm with theoretical basis and practical application value .

Description of drawings

Fig. 1 is a schematic diagram of obtaining T spectral sets in the hyperspectral remote sensing image classification method based on hierarchical ensemble learning in the present invention.

Detailed ways

Specific embodiment one: below in conjunction with Fig. 1, illustrate this embodiment, the hyperspectral remote sensing image classification method based on layered integrated learning described in this embodiment, it comprises the following steps:

Step 1: Read the hyperspectral raw data, preprocess it, obtain supervised data, then determine the labeled samples according to the supervised data, and select training samples and test samples from the labeled samples;

Step 2: Perform random band selection for all spectral bands of the hyperspectral raw data, and then extract a pixel vector combination from each marked sample, the number of bands contained in each pixel vector in the pixel vector combination is the same as that of the random The number of bands corresponds one by one to form a spectral set; repeat the random band selection process and form a corresponding spectral set to form multiple spectral sets with differences;

Step 3: Using each spectral collection as a unit, use the training samples to train the inner layer structure of the Adaboost integration framework, and then classify the test samples to obtain the inner layer integration composed of the Adaboost framework;

Step 4: Construct the outer structure of the Adaboost integration framework from the classification results and classification accuracy of the test samples in step 3;

Step 5: Use the Adaboost integration framework with the above-mentioned inner structure and outer structure to classify the hyperspectral raw data in a traversal manner, and obtain the classification theme map.

This embodiment mainly uses a two-layer integrated structure to classify hyperspectral images, namely an inner layer structure and an outer layer structure. The inner structure is a collection of different spectra formed by random band selection. After that, the spectrum collection is used as the unit, and the Adaboost integration method is used to train respectively, and then the test samples are classified. The outer structure integrates the classification results of each spectral set in the inner integration, and uses the method of weight voting to determine the final category of the sample. Finally, the whole image is used as a test sample to realize the classification of the whole image to obtain the classified theme map.

Specific implementation mode two: this implementation mode further explains implementation mode one, the specific method of determining the marked samples in step one, and selecting training samples and test samples from the marked samples is as follows:

Read the hyperspectral raw data stored in the form of a three-dimensional matrix, the three-dimensional matrix form includes two-dimensional spatial position information and one-dimensional spectral information; The object categories in the marker map are respectively marked in the form of integer values, and the supervision data in the form of a two-dimensional matrix is obtained. composition;

From the hyperspectral raw data and supervisory data, determine the number of feature categories C in the real feature marker map, the number of pixels in the two-dimensional spatial position information M rows × N columns, and the number of effective bands in the one-dimensional spectral information B;

The spatial position data of pixels in the supervisory data is extracted from top to bottom and from left to right in the original hyperspectral data according to the position coordinates, and the pixel vector of spectral information is extracted as a marked sample, and the marked samples are arranged in two-dimensional rows In matrix form, the number of rows of the two-dimensional matrix is the number of marked samples, and the number of columns is the number of bands B contained in the pixel vector;

The odd-numbered rows of the marked samples expressed in a two-dimensional matrix form are selected as training samples, and the even-numbered rows are selected as test samples, and the ratio of the number of training samples to the number of test samples is 1:1;

Then extract the feature category data corresponding to the marked samples expressed in the two-dimensional matrix form from the supervised data, and arrange them into a column vector. The number of feature category data in the column vector is the number of marked samples, and each The value of the feature category data is the category label of a corresponding labeled sample, the odd-numbered element in the column vector is used as the category label of the training sample, and the even-numbered element is used as the category label of the test sample.

Preprocess the hyperspectral raw data to prepare data for subsequent feature extraction and classification algorithms. It mainly includes two parts: reading hyperspectral data, determining labeled samples, and selecting training samples and test samples. In the supervised data, the feature category of each pixel in the spatial position information is marked in the form of an integer value, and the pixels of the same category have the same value. Labeled samples obtained from supervised data are used to train and test classification methods.

The vector formed by different spectral positions of the same spatial position in the hyperspectral raw data is called a pixel vector. The first pixel vector of the labeled samples is arranged in the first row, the second pixel vector is arranged in the second row, and so on, finally forming a two-dimensional matrix. There is a one-to-one correspondence between the pixel vectors of labeled samples and the class labels.

Specific embodiment three: the present embodiment will be described below in conjunction with FIG. 1. This embodiment will further describe embodiment two. The specific method for forming a plurality of different spectral sets in step two is:

From 1 to B, a group of b random numbers {i ₁ , i ₂ ,…,i _b } is drawn, where 1≤i _k ≤B, k=1,2,…,b), and then In each marked sample, the i _kth vector element is extracted in the order of k from 1 to b to form a pixel vector combination containing b elements respectively, and the number of the pixel vector combination is the same as the number of marked samples , all pixel vectors containing b elements are combined as a spectral set;

The process of forming a spectrum set is repeated multiple times, and b random numbers are selected in the repeated process, and each combination of random numbers is different, so as to obtain multiple spectrum sets with differences.

The hyperspectral raw data contains dozens or even hundreds of bands, and the method of constructing the spectral collection adopts the method of random band extraction. A part of the bands is randomly selected from all the bands of the hyperspectral data as the spectral members of the spectral set. Such random extraction is repeated for a certain number of times to reach the corresponding number of spectral sets. Specific steps include:

1. Random band extraction. In a hyperspectral image, each pixel has a pixel vector of length B, where B is the number of bands contained in the original data. Generate a set of combinations {i ₁ ,i ₂ ,…,i _b } containing b random numbers between 1-B, from the pixel vector of each labeled sample (including training samples and test samples), according to k from The sequence from 1 to b extracts the i _kth vector element to form a pixel vector containing b elements. In this way, for all labeled samples, there is a pixel vector of length b after random band extraction. The labeled samples extracted from these random bands are regarded as a spectrum set. Experience shows that when b is about 0.3 times of B, b takes an integer value, and the effect of subsequent classification is better.

2. Construct a collection of spectra. After the above random band extraction, only one spectrum set is formed. Repeat the above random extraction process T times, and the parameters are consistent during the repetition process, but there are differences in the combinations of generated random numbers, and then T spectral sets can be obtained. Experience shows that when T is about 20-30, the effect of subsequent classification is better.

Figure 1 can illustrate the above process more intuitively, where gray is the band extracted during the T repetition process.

Specific implementation mode four: this implementation mode further explains the implementation mode three, and the specific method for obtaining the inner layer integration composed of the Adaboost framework in step three is:

Taking each spectral set as a unit, select the support vector machine SVM as the weak classifier of the Adaboost integration framework; for the first spectral set, iterate the combination of pixel vectors corresponding to the training samples in the spectral set according to the integration method of Adaboost , set the number of iterations as F, then form a series of weak classifiers f(f=1,2,...,F) with differences;

First, assign the same weight to each pixel vector combination corresponding to the training sample in the first spectral set, and then use the weak classifier 1 to train and test the current pixel vector weight combination, for the misclassified pixel vector Increase its weight, and reduce its weight for the pixel vector of correct decision; the pixel vector weight combination after weight adjustment is used to train the weak classifier 2 on the current pixel vector weight combination, and so on, iterate F times repeatedly, and obtain the focus on F weak classifiers with different combinations of pixel vector weights;

Taking each spectral set as a unit, first use F weak classifiers to classify each pixel vector corresponding to the test sample in the first spectral set, and obtain F classification results for each pixel vector, thus obtaining F classification results of the corresponding test samples, and then the final classification results of the corresponding test samples are determined by majority voting;

Set a total of T spectral sets, repeat the above classification process for the second spectral set to the Tth spectral set, and the final classification results of the test samples constitute T internal integrations composed of the Adaboost framework.

The inner layer structure is composed of different spectral sets by random band selection, and then the integrated method of Adaboost is used to train the spectral sets, and then the test samples are classified. Specific steps are as follows:

1. Design and use training samples to train the Adaboost ensemble framework for each spectral collection. The support vector machine (SVM) is chosen as the weak classifier of the Adaboost ensemble framework. For spectral set 1, according to the integration method of Adaboost, the training samples in the set are iterated, and the number of iterations is set as F, which will form a series of classifiers f (f=1,2,...,F) with differences. First, assign the same weight to each training sample; then train and test the training samples through the weak classifier 1, increase the weight of the misclassified samples, and reduce the weight of those samples that are correctly judged; finally according to the weight The adjusted training set trains the next round of classifier 2, and this process can be repeated for F times to obtain F weak classifiers focusing on different training samples.

The weak classifiers integrated with Adaboost can be replaced by SVM classifiers such as K-nearest neighbors, decision trees, and neural networks. Experience shows that SVMs perform better than other classifiers.

2. Take each spectral set as a unit to classify the test samples. Use F weak classifiers to classify test samples, and there will be F classification results for each sample, and the final class is determined by majority voting. The basic idea of the majority voting method: Multiple weak classifiers in the integrated framework perform classification predictions and give their respective classification results, and then vote according to the classification results using the principle of voting, and the minority obeys the majority. For each test sample, let F weak classifiers perform class voting, and the class with the most votes is used as the classification result of the corresponding sample.

3. For the spectrum set 2-T, repeat the above steps 1-2, and T inner layer integrations composed of the Adaboost framework will be formed.

Specific implementation mode five: this implementation mode further explains implementation mode four, and the specific method of constructing the outer layer structure of Adaboost integrated framework in step four is:

Integrate the final classification results of the test samples obtained by T internal integrations, and then use the method of weight voting to determine the final category of the integration results, and construct the outer structure of the Adaboost integration framework.

The outer layer structure integrates the classification results obtained from T inner layer integrations, and uses the method of weight voting to determine the final category of the sample.

For each internal integration in the external integration structure, a judgment category will be given for the test samples. In order to obtain the final integration decision, the output of these individual classifiers needs to be integrated with a certain strategy to obtain hierarchical integrated learning. classification results. In the face of multiple classification results, it is necessary to seek an integration strategy to obtain the best decision. Select the weight voting method here. The weight of each internal integration structure is determined by its classification accuracy for training samples. The category of a certain pixel is judged by the sum of weights after voting, and the category with the largest sum of weights is the final category output.

Specific implementation mode six: This implementation mode further explains implementation mode five, and the specific method for obtaining the classification theme map in step five is:

Using the traversal method from top to bottom and from left to right, the pixel vectors in the same method as above are extracted for each pixel in the hyperspectral original data spatial position information, and M×N pixel vectors are obtained, and then the above-mentioned The Adaboost integrated framework of the inner structure and the outer structure classifies M×N pixel vectors one by one to obtain M×N category labels, and then converts the M×N category labels into corresponding two-dimensional arrays of M rows×N columns Matrix, the two-dimensional matrix of M rows×N columns is displayed as a two-dimensional image to obtain a classification theme map.

By traversing from top to bottom and from left to right, pixel vectors are extracted from all pixels in the original hyperspectral image. M×N pixel vectors will be obtained, and all pixel vectors will be classified using a hierarchical integrated structure, and M×N category labels will be obtained, which are one-dimensional vectors, and the category labels will be converted into M rows×N columns The two-dimensional matrix of , displaying the two-dimensional matrix as a two-dimensional image will give an intuitive evaluation of the classification, that is, the classification theme map.

Claims (6)

1., based on a Classification of hyperspectral remote sensing image method for layering integrated study, it is characterized in that, it comprises the following steps:

Step one: read EO-1 hyperion raw data, carry out pre-service to it, obtains monitoring data, then according to monitoring data determination marker samples, and by selecting training sample and test sample book in marker samples;

Step 2: to all spectral bands of EO-1 hyperion raw data, enter row stochastic band selection, then from each marker samples, extract pixel vectors combination, the wave band number that in the combination of this pixel vectors, each pixel vectors comprises and described random wave band number one_to_one corresponding, form a spectrum set; Repeat random band selection process and form corresponding spectrum set, forming multiple spectrum set that there are differences;

Step 3: be combined into unit with each spectra collection, uses the endothecium structure of training sample training Adaboost integrated framework, then classifies to test sample book, obtain the internal layer be made up of Adaboost framework integrated;

Step 4: by classification results and the nicety of grading of test sample book in step 3, the layer structure of structure Adaboost integrated framework;

Step 5: adopt the Adaboost integrated framework with above-mentioned endothecium structure and layer structure to adopt traversal mode to classify to EO-1 hyperion raw data, obtain classification scheme figure.

2. the Classification of hyperspectral remote sensing image method based on layering integrated study according to claim 1, is characterized in that, determine marker samples in step one, and by selecting the concrete grammar of training sample and test sample book to be in marker samples:

Read the EO-1 hyperion raw data stored with three-dimensional matrice form, described three-dimensional matrice form comprises the spatial positional information of two dimension and the spectral information of one dimension; By the atural object classification in the figure of substance markers truly corresponding for pixel locus each in the spatial positional information of two dimension respectively with integer-valued formal notation out, obtain the monitoring data of two-dimensional matrix form, this monitoring data is by pixel spatial position data and the atural object categorical data with integer-valued formal notation form one to one;

By EO-1 hyperion raw data and monitoring data, determine pixel number M in the spatial positional information of the atural object class number C in substance markers figure truly, two dimension capable × spectral information of N row and one dimension in effective wave band number B;

By pixel spatial position data in monitoring data according to position coordinates mode from the top down and from left to right in EO-1 hyperion raw data, extract the pixel vectors of spectral information as marker samples, and marker samples is become two-dimensional matrix form by rows, the line number of this two-dimensional matrix is the number of marker samples, and columns is the wave band number B that pixel vectors comprises;

Select the odd-numbered line of the marker samples represented using two-dimensional matrix form as training sample, even number line is chosen as test sample book, and makes training sample be 1:1 with the ratio of test sample book number;

Again by extracting the marker samples atural object categorical data one to one represented with two-dimensional matrix form in monitoring data, and arrange vector in column, in this column vector, the number of atural object categorical data is the number of marker samples, the numerical value of each atural object categorical data is the category label of a corresponding marker samples, in this column vector, odd number element is as the category label of training sample, and even number element is as the category label of test sample book.

3. the Classification of hyperspectral remote sensing image method based on layering integrated study according to claim 2, is characterized in that, the concrete grammar forming multiple spectrum set that there are differences in step 2 is:

Combination { the i that one group contains b random number is extracted between 1 to B ₁, i ₂..., i _b, wherein 1≤i _k≤ B, k=1,2 ..., b), then by each marker samples, according to the order of k from 1 to b by i-th _kindividual vector element extracts, and form a pixel vectors containing b element respectively and combine, the number of this pixel vectors combination is identical with the number of marker samples, and all pixel vectors combinations containing b element are as a spectrum set;

By the process of a formation spectrum set repeatedly, in repeated process, random number is all chosen as b, and the combination of random number each time be there are differences, and obtains multiple spectrum set that there are differences.

4. the Classification of hyperspectral remote sensing image method based on layering integrated study according to claim 3, is characterized in that, the concrete grammar obtaining the internal layer that is made up of Adaboost framework in step 3 integrated is:

Be combined into unit with each spectra collection, select support vector machines as the Weak Classifier of Adaboost integrated framework; For first spectrum set, the integrated approach according to Adaboost is combined into row iteration to the pixel vectors corresponding with training sample in this spectrum set, and setting iterations is F, then form the serial Weak Classifier f (f=1 that there are differences, 2 ..., F);

First be that each pixel vectors corresponding with training sample combines and give identical weights in first spectrum set, then training and testing is carried out by the combination of Weak Classifier 1 pair of current pixel vector weight, for wherein being improved its weight by the pixel vectors that mistake is divided, the pixel vectors for correct judgement reduces its weight; Pixel vectors set of weights after weight adjusting is combined in described current pixel vector set of weights close training Weak Classifier 2, the rest may be inferred, iterates F time, obtain lay particular emphasis on different pixels vector weight combine F Weak Classifier;

Unit is combined into again with each spectra collection, first combinationally use F Weak Classifier to each pixel vectors corresponding with test sample book in first spectrum set to classify, each pixel vectors obtains F classification results, obtain F classification results of corresponding test sample book thus, then the mode of being voted by majority determines the final classification results of corresponding test sample book;

Total T the spectrum set of setting, is bonded to T spectrum set for second spectra collection, repeats above-mentioned assorting process, and final classification results formation T internal layer be made up of Adaboost framework of test sample book is integrated.

5. the Classification of hyperspectral remote sensing image method based on layering integrated study according to claim 4, is characterized in that, the concrete grammar constructing the layer structure of Adaboost integrated framework in step 4 is:

The final classification results of the test sample book that T internal layer integrates is integrated, then adopts the method for weight votes to carry out the determination of final classification to integrated results, construct the layer structure of Adaboost integrated framework.

6. the Classification of hyperspectral remote sensing image method based on layering integrated study according to claim 5, is characterized in that, the concrete grammar obtaining classification scheme figure in step 5 is:

Adopt from top to bottom, traversal mode from left to right, the extraction of the pixel vectors of same method in aforesaid way is carried out to each pixel in EO-1 hyperion original data space positional information, obtain M × N number of pixel vectors, the Adaboost integrated framework with above-mentioned endothecium structure and layer structure is adopted to classify one by one to M × N number of pixel vectors again, obtain M × N number of class label, again M × N number of class label is converted to corresponding M capable × two-dimensional matrix of N row, this M is capable × two-dimensional matrix of N row obtains classification scheme figure as two dimensional image display.