CN107679453A - Weather radar electromagnetic interference echo recognition methods based on SVMs - Google Patents

Abstract

The invention relates to image information processing technology. In order to fully utilize the different characteristics of the three types of electromagnetic interference echoes and the superiority of the SVM method, the three types of electromagnetic interference echoes can be identified simultaneously, and the accuracy of electromagnetic interference echo identification is improved. In the present invention, the weather radar electromagnetic interference echo recognition method based on a support vector machine, the steps are as follows: step 1, establish a training sample set step 2, load positive and negative samples and assign labels; step 3, extract positive and negative sample features step 4, Train the SVM classifier, and generate the SVM classifier corresponding to the training sample data; step 5, save the classifier step 6, detect the accuracy of the classifier classification step 7, load the test image; step 8, extract the test image feature step 9, Using the trained SVM classifier to classify the test pattern; step 10, realize the identification of the spiral electromagnetic interference echo and the pitted electromagnetic interference echo. The invention is mainly applied to image information processing occasions.

Description

Echo Recognition Method of Weather Radar Electromagnetic Interference Based on Support Vector Machine

technical field

The invention relates to the technical field of image information processing, in particular to a support vector machine-based identification method for electromagnetic interference echoes of weather radars.

Background technique

Weather radar echoes are divided into meteorological echoes and non-meteorological echoes, and non-meteorological echoes will interfere with the identification of meteorological echoes. Non-meteorological echoes are mainly divided into surface object echoes, biological echoes, electromagnetic interference echoes and abnormal radar fault echoes. Among them, the electromagnetic interference echo refers to the interference echo phenomenon that occurs when the radar echo is interfered by external equipment with the same frequency or adjacent frequency. These interference echoes can be divided into three categories, namely, radial interference echo and spiral interference echo. Waves and pitted interference echoes. For short-range electromagnetic interference, the radar screen will be full of interference pits; for long-distance electromagnetic interference, there will be fixed interference in a certain direction. If it is a single-frequency point electromagnetic interference, the interference is in a straight line; and for interference with a certain bandwidth, it will produce a full screen of interference pits on the radar screen.

Electromagnetic interference echo is an important factor affecting the quality of the new generation of weather radar products. Its appearance will reduce the quality of radar-generated products, mainly polluting the reflectivity data, which brings great troubles to the application of radar data. Although radar operators or forecasters can artificially observe and mark electromagnetic interference echoes, if there is no automatic identification method, it will have a great impact on precipitation estimation and radar data assimilation in real-time operations. For this reason, people hope to propose a method to eliminate electromagnetic interference from the aspects of hardware and software, so as to improve the quality of radar base data and other secondary products. Solving this problem through hardware has not yet been realized. In terms of software, it is mainly to identify electromagnetic interference echoes by designing different algorithms. At present, there are few identification algorithms, mainly including isolated echo elimination algorithms and path detection algorithms. Algorithm for identifying interference. The isolated echo algorithm performs filtering processing on the isolated interference echo. First, a moving window is established on the observation map. If the effective value points around the center point in the window are less than a certain threshold, the The central point is removed, that is, the point is regarded as an isolated electromagnetic interference echo. Therefore, this method is suitable for identifying or removing helical and pitted EMI echoes. However, due to the uncertain distance, random azimuth, different pulse width and working repetition frequency of the interference source, the shape of the spiral interference echo is not the same, that is, the continuity in the radial direction and the azimuth is different, so the isolated echo cancellation algorithm The spiral electromagnetic interference echoes of different forms cannot be completely filtered out. Similarly, this problem also exists for pitted interference echoes; the radial interference identification method is established based on the shape of electromagnetic interference echoes. First, the range bins at the same distance are divided into echo bands along the tangential direction. Then, count the effective total distance library number R _T of each radial echo intensity and the distance library number R _N whose width of the echo segment where the distance library is located is between 0 and 5, and calculate R _D (R _D =R _N / R _T ×100%), if R _D ≥ 30%, it is determined that this radial direction is an electromagnetic interference echo. This method is only suitable for identifying radial electromagnetic interference echoes, and can only identify narrow electromagnetic interference echoes with less than 5 radial directions. Radial electromagnetic interference echoes beyond this range cannot be identified by this method. Therefore, the current electromagnetic interference echo identification methods need to be improved and enhanced.

Image classification can be used as a method of information extraction, which has been widely used in various fields. The focus of image classification research mainly focuses on the selection and extraction of image features and the selection of classification models. HOG (Histogram of Oriented Gradient) is a histogram of oriented gradients, which is a feature descriptor used for object detection in computer vision and image processing. The HOG feature can maintain good invariance to the geometric and optical deformation of the image, has good robustness to the light intensity, and is widely used in outdoor detection. Support Vector Machine (SVM) is a data mining method based on statistical learning theory, which has been proved to be ideal for processing problems such as classification, regression and pattern recognition. SVM is essentially a feed-forward neural network. According to the structural risk minimization criterion, the generalization ability of the classifier is improved as much as possible on the premise of minimizing the classification error of the training samples. The basic principle is to find an optimal classification hyperplane that meets the classification requirements, so that the hyperplane can maximize the classification interval on both sides of the hyperplane while ensuring the classification accuracy.

Contents of the invention

In order to overcome the deficiencies of the prior art, the present invention aims to propose a method for identifying echoes of weather radar electromagnetic interference. By making full use of the different characteristics of the three types of electromagnetic interference echoes and the superiority of the SVM method, the three types of electromagnetic interference echoes can be identified at the same time, and the accuracy of electromagnetic interference echo identification can be improved. For this reason, the technical solution that the present invention adopts is, based on the weather radar electromagnetic interference echo recognition method of support vector machine, the steps are as follows:

Step 1. Create a training sample set

The positive sample set is established as the radial electromagnetic interference echo image, and the negative sample set is the spiral electromagnetic interference echo and pitted electromagnetic interference echo image;

Step 2. Load positive and negative samples and assign labels;

Step 3. Extract positive and negative sample features

Read in the positive sample image, convert the color image to a grayscale image, and then extract its HOG features; read in the negative sample image, convert the color image to a grayscale image, and then extract its HOG features;

Step 4, train the SVM classifier, and generate the SVM classifier corresponding to the training sample data;

Step 5. Save the classifier

Step 6. Check the classification accuracy of the classifier

Mix the HOG features corresponding to the positive and negative sample sets, and randomly divide them into two groups, one group is a new training set, and the other group is a verification set. Use the verification set to self-test the classification results of the classifier and test the classification accuracy of the classifier. ;

Step 7, load the test image;

Step 8. Extract test image features

Read in the test image, convert the color image to a grayscale image, and then extract its HOG features;

Step 9, using the trained SVM classifier to classify the test pattern, obtain the classification results of radial electromagnetic interference, spiral electromagnetic interference and pitted electromagnetic interference echo, and realize the identification of radial electromagnetic interference echo; If the test result is radial electromagnetic interference echo, the program ends; if the test result is non-radial electromagnetic interference echo, proceed to the next step;

Step 10. Establish a positive sample set as a spiral electromagnetic interference echo image, and a negative sample set as a pitted electromagnetic interference echo image, repeat steps 2 to 9 to obtain the classification of spiral electromagnetic interference and pitted electromagnetic interference echo As a result, the identification of spiral electromagnetic interference echoes and pockmarked electromagnetic interference echoes is realized.

A further technical solution is that in steps 3 and 8, the specific method of HOG feature extraction is: first, calculate the gradient magnitude and gradient direction of each pixel in the sample image; then, divide the image into several connected areas, These connected areas are called cell units, and then the cell units are divided into several blocks, and the 0°~180° are divided into several channels on average; secondly, the direction histogram of the gradient of each pixel in the cell unit is collected; finally, the different The gradient size of the pixels of the channels is accumulated, and a set of vectors composed of the sum of the gradients of the pixels of each channel is obtained, and then the vectors are normalized in units of blocks, and the normalized vectors are linked to form a HOG sample. feature descriptor;

The calculation gradient is to use two discrete differential templates (-1, 0, 1) and (-1, 0, 1) ^T , T means transpose, and process the horizontal and vertical directions of the image respectively to get each pixel gradient.

In step 4, the specific process of obtaining the SVM classifier is: the process of obtaining the optimal SVM classification function based on the known sample set can solve the classification problem. For nonlinear problems, first transform the nonlinear problem into a high-dimensional space The linear problem in , and then solve the optimal classification surface in the transformed space, set the given sample set as s={( _xi ,y _i )|i=1,2…n}, x _i ∈ R ^d is The vector in the d-dimensional space, y _i ={+1,-1} is the category label corresponding to x _i , i is the sample set number, assuming that the nonlinear mapping is in As the mapping function, the support vector machine classification surface is written as: w is the weight vector, b is the offset constant, is the classification interval, the optimization problem of obtaining the maximum classification interval can be expressed as The minimum value problem of , the constraints are According to the KKT (Karush-Kuhn-Tucker) condition and the Lagrangian function, the classification function is:

Among them, sgn( ) is a sign function, a _i is the optimal Lagrangian multiplier, x is the sample to be tested, The dot product operation ( _xi x) in high-dimensional space is replaced by the kernel function k( _xi , x _j ) in low-dimensional space, and the classification function can be rewritten as

The commonly used kernel functions are as follows:

Polynomial kernel function k(x, x _i )=[(x x _i )+1] ^q q=1,2,3...;

Gaussian Radial Basis Function σ is a parameter;

sigmoid function k(x, _xi )=tanh(b(x, _xi )+c) b, c is a constant;

Fourier series where N is a constant;

B-spline kernel function k(x, _xi )=B _2p+1 (xx _i ) where B _2p+1 (x) is B of order 2p+1;

Selecting different kernel functions can generate different SVM classification functions, thereby obtaining different SVM classifiers, so it can be extended to different SVM classification methods through the selection of kernel functions.

Features and beneficial effects of the present invention are:

Compared with the prior art, the method of the present invention has the beneficial effects that: the present invention can identify three types of weather radar electromagnetic interference echoes, the algorithm principle is simple, has real-time performance, and can self-test the classification accuracy at the same time, ensuring Accuracy of EMI echo identification.

Description of drawings:

Fig. 1 basic flow chart of the method of the present invention.

Fig. 2 Example classification results.

detailed description

The weather radar electromagnetic interference echo identification method of the present invention specifically comprises the following steps:

Step 1. Establish a training sample set.

The positive sample set (positive sample) is the radial electromagnetic interference echo image, and the negative sample set (negative sample) is the spiral electromagnetic interference echo and pitted electromagnetic interference echo image.

Step 2. Load positive and negative samples and assign labels.

Step 3. Extract positive and negative sample features.

Read in the positive sample image, convert the color image to a grayscale image, and then extract its HOG features; read in the negative sample image, convert the color image to a grayscale image, and then extract its HOG features.

Step 4, train the SVM classifier, and generate the SVM classifier corresponding to the training sample data.

Step 5. Save the classifier.

Step 6. Detect the classification accuracy of the classifier.

Mix the HOG features corresponding to the positive and negative sample sets, and randomly divide them into two groups. One group is a new training set, and the other group is a verification set. Use the detection set to self-test the classification results of the classifier to test the classification accuracy of the classifier. .

Step 7. Load the test image.

Step 8, extracting test image features.

Read in a test image, convert the color image to a grayscale image, and then extract its HOG features.

Step 9. Use the trained SVM classifier to classify the test pattern to obtain the classification results of radial electromagnetic interference, spiral electromagnetic interference and pitted electromagnetic interference echoes, and realize the identification of radial electromagnetic interference echoes. If the test result is radial electromagnetic interference echo, the program ends. If the test result is non-radial electromagnetic interference echo, proceed to the next step.

Step 10. Establish a positive sample set as a spiral electromagnetic interference echo image, and a negative sample set as a pitted electromagnetic interference echo image, repeat steps 2 to 9 to obtain the classification of spiral electromagnetic interference and pitted electromagnetic interference echo As a result, the identification of spiral electromagnetic interference echoes and pockmarked electromagnetic interference echoes is realized.

A further technical solution is that in steps 3 and 8, the specific method of HOG feature extraction is as follows: first, calculate the gradient magnitude and gradient direction of each pixel in the sample image. Then, the image is divided into several connected regions, which are called cell units. Then divide the cell unit into several blocks, and divide 0°～180° into several channels on average. Secondly, the direction histogram of the gradient of each pixel point in the cell unit is collected. Finally, the gradients of the pixels of different channels are accumulated to obtain a set of vectors composed of the cumulative sum of the gradients of the pixels of each channel, and then the vectors are normalized in units of blocks, and the normalized vectors are linked together , constituting the HOG sample feature descriptor.

The calculation of the gradient is to use two discrete differential templates (-1, 0, 1) and (-1, 0, 1) ^T (T means transpose) to process the horizontal and vertical directions of the image respectively, and obtain each pixel gradient.

In step 4, the specific process of obtaining the SVM classifier is: the process of obtaining the optimal SVM classification function according to the known sample set can solve the classification problem. For nonlinear problems, first transform the nonlinear problem into a linear problem in a high-dimensional space, and then solve the optimal classification surface in the transformed space. Suppose the given sample set is s={( _xi ,y _i )|i=1,2n}, and x _i ∈ R ^d is a vector in d-dimensional space. y _i ={+1,-1} is the category label corresponding to _xi . i is the sample set number. Suppose the nonlinear mapping is in As a mapping function, the support vector machine classification surface can be written as: w is the weight vector and b is the offset constant. is the classification interval, the optimization problem of obtaining the maximum classification interval can be expressed as minimum value problem. The constraints are According to the Karush-Kuhn-Tucker (KKT) condition and the Lagrangian function, the classification function can be obtained as Among them, sgn(·) is a symbolic function. a _i is the optimal Lagrangian multiplier. x is the sample to be tested. The dot product operation ( _xi x) in high-dimensional space can be replaced by the kernel function k( _xi , x _j ) in low-dimensional space, and the classification function can be rewritten as

Commonly used kernel functions are as shown in Table 1. Selecting different kernel functions can produce different SVM classification functions, thereby obtaining different SVM classifiers. Therefore, the method of the present invention can be extended to different SVM classification methods by selecting kernel functions .

Table 1 Commonly used kernel functions

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention. The basic process of the inventive method as shown in Figure 1, specifically comprises the following steps:

Step 1. Establish a training sample set. The positive sample set (positive sample) is 20 radial electromagnetic interference echo images, and the positive sample file list pos_list.txt is established; the negative sample set (negative sample) is 20 spiral electromagnetic interference echoes and pitted electromagnetic interference Echo image, create negative sample file list neg_list.txt.

Step 2. Load positive and negative samples and assign labels. Load the positive sample list, label it as label1, load the negative sample list, label it as label2, and record the label summary as label.

Step 3. Extract positive and negative sample features. Read in the positive sample image, convert the color image to a grayscale image, and then use the hogcalculator function to extract its HOG features; read in the negative sample image, convert the color image to a grayscale image, and then use the hogcalculator function to extract its HOG features. Here, the specific method for extracting HOG features by the hogcalculator function is as follows: First, calculate the gradient magnitude and gradient direction of each pixel in the sample image. Then, the image is divided into several connected regions of 6*6 pixels, and these connected regions are called cell units. 3*3 cell units constitute a block, which is divided into 9 channels evenly from 0° to 180°. Secondly, the direction histogram of the gradient of each pixel point in the cell unit is collected. Finally, the gradients of the pixels of 9 different channels are accumulated to obtain a set of vectors composed of the cumulative sum of the gradients of the pixels of each channel, and then the vectors are normalized in units of blocks, and the normalized vectors Linked to form a HOG sample feature descriptor.

Step 4. Use the positive sample set and the negative sample set to train the SVM classifier respectively (the positive sample set and the negative sample set here refer to their respective corresponding HOG features), and generate an SVM classifier corresponding to the training sample data. The specific process of generating the SVM classifier is: Given a known sample set as m={( _xi ,y _i )|i=1,2n}, _xi is a positive and negative training sample set. y _i ={+1,-1} is the category label corresponding to _xi . i is the sample set number. The process of finding the optimal SVM classification function based on the known sample set can solve the classification problem. The process of obtaining the classification function has been described in detail in Step 4 of the Summary of the Invention. The classification function is where sgn(·) is a symbolic function. a _i is the optimal Lagrangian multiplier. x is the sample to be tested. k(x _i , x _j ) is a kernel function, and in this embodiment, the kernel function is a Gaussian radial basis kernel function: σ is the kernel function parameter.

Step 5. Save the classifier.

Step 6. Detect classification accuracy. Mix the HOG features corresponding to the positive and negative sample sets, and randomly divide them into two groups. One group is a new training set, and the other group is a verification set. accuracy. When using the 15th picture in the positive sample set for self-verification test, the classification result is a positive sample, and the self-verification detection value is 0.8947. When using the 15th picture in the negative sample set for self-verification test, the classification result is a negative sample, and the self-verification detection value is 0.9474. It can be seen that the classification result of the verification classifier is correct and the confidence is high.

Step 7. Load the test image. The test image sample set includes 10 radial EMI echo images, 10 helical EMI images and 10 pitted EMI images, all of which are 64×64 in size.

Step 8, extracting test image features. Read in the test image, convert the color image to a grayscale image, and then use the hogcalculator function to extract its HOG features. The HOG feature extraction is the same as step 3.

Step 9. Use the trained SVM classifier to classify the test pattern to obtain the classification results of radial electromagnetic interference, spiral electromagnetic interference and pitted electromagnetic interference echoes, and realize the identification of radial electromagnetic interference echoes. If the test result is radial electromagnetic interference echo, the program ends. If the test result is non-radial electromagnetic interference echo, proceed to the next step.

Step 10. Establish a positive sample set (positive sample) as a list of spiral electromagnetic interference echo image files, and a negative sample set (negative sample) as a list of pitted electromagnetic interference echo image files. Repeat steps 2 to 9 to obtain a spiral The classification results of electromagnetic interference and pitted electromagnetic interference echoes realize the identification of spiral electromagnetic interference echoes and pitted electromagnetic interference echoes.

The classification result of this embodiment is shown in FIG. 2 . "class" in the figure indicates the classification result of the classifier, "true" indicates the type of the picture itself, "1" indicates radial interference electromagnetic echo, "2" indicates spiral electromagnetic interference echo, and "3" indicates pitted electromagnetic interference echo. Among them, the identification results of radial interference electromagnetic echoes in the first and second rows, the identification accuracy rate of radial interference echoes is 90%, and one spiral interference echo is included in the 10 identification results of radial interference echoes; the third and fourth Behavioral spiral electromagnetic interference echo recognition results, the accuracy rate of spiral interference echo recognition is 90%, 10 spiral interference echo recognition results include one pitting interference echo; the 5th to 6th behavior pitting electromagnetic interference echo Wave identification results, pitting interference echo identification accuracy rate of 90%, 10 pitting interference echo identification results include a spiral interference echo. An overall accuracy of 90% can be calculated.

Although the present invention has been described here with reference to the embodiments of the present invention, it should be understood that those skilled in the art can design many other modifications and implementations, and these modifications and implementations will fall within the scope and spirit of the principles disclosed in the application within. More specifically, within the scope of the disclosure and claims of the present application, various modifications and improvements can be made to the components and or layout of the subject combination layout.

Claims (4)

1. a kind of weather radar electromagnetic interference echo recognition methods based on SVMs, it is characterized in that, step is as follows：

Step 1, establish training sample set

Establish positive sample to integrate as radial direction electromagnetic interference echo, negative sample integrates as helical form electromagnetic interference echo and pit shape electricity Magnetic disturbance echo；

Step 2, it is loaded into positive negative sample and distributes label；

Step 3, the positive and negative sample characteristics of extraction

Positive sample image is read in, coloured image is converted into gray level image, then extracts its HOG feature；Negative sample image is read in, Coloured image is converted into gray level image, then extracts its HOG feature；

Step 4, training SVM classifier, and generate the SVM classifier of corresponding training sample data；

Step 5, preserve grader

Step 6, the accuracy of detection grader classification

HOG features corresponding to positive and negative sample set are mixed, are randomly divided into two groups, one group is new training set, and one group is checking collection, Autonomous test, the accuracy rate of testing classification device classification are carried out using checking set pair grader classification results；

Step 7, it is loaded into test image；

Step 8, extraction test image feature

Test image is read in, coloured image is converted into gray level image, then extracts its HOG feature；

Step 9, using the SVM classifier trained test chart is classified, obtain radial direction electromagnetic interference and spiral electricity Magnetic disturbance and the classification results of pit shape electromagnetic interference echo, realize the identification to radial direction electromagnetic interference echo；Such as test result For radial direction electromagnetic interference echo, then EP (end of program)；If test result is non-radial electromagnetic interference echo, then carry out in next step；

Step 10, establish positive sample and integrate as helical form electromagnetic interference echo, negative sample integrates as pit shape electromagnetic interference echo Image, repeat step 2 arrive step 9, obtain spiral electromagnetic interference and the classification results of pit shape electromagnetic interference echo, realize to spiral shell Revolve the identification of electromagnetic interference echo and pit shape electromagnetic interference echo.

2. the weather radar electromagnetic interference echo recognition methods based on SVMs as claimed in claim 1, it is characterized in that, Further technical scheme is in the step 3 and step 8, and the specific method of HOG feature extractions is：First, sample is calculated The gradient magnitude and gradient direction of each pixel in image；Then, some UNICOM regions, these connected regions are divided the image into It is called cell factory, then cell factory is divided into several blocks, several passages is equally divided into by 0 °~180 °；Secondly, adopt Collect the direction histogram of the gradient of each pixel in cell factory；Finally, the gradient magnitude of the pixel of different passages is added up, One group is obtained by the vector that each passage pixel gradient is cumulative and forms, then in units of block, vector is normalized, Vector after normalized is chained up, forms HOG sample characteristics descriptors；

It is to use two discrete differential templates (- 1,0,1) and (- 1,0,1) to calculate gradient^T, T represents transposition, respectively to image level Direction and vertical direction are handled, and obtain the gradient of each pixel.

3. the weather radar electromagnetic interference echo recognition methods based on SVMs as claimed in claim 1, it is characterized in that, In step 4, obtaining SVM classifier detailed process is：The process of optimal svm classifier function is asked for according to known sample collection to be solved Certainly classification problem, it is the linear problem in some higher dimensional space first nonlinear problem shift conversion for nonlinear problem, Then optimal classification surface is solved in space after transformation, if given sample set is s={ (x_i,y_i) | i=1,2n }, x_i∈R^d For the vector in d dimension spaces, y_i={+1, -1 } is x_iCorresponding category label, i are sample set numberings, it is assumed that Nonlinear Mapping isWhereinFor mapping function, support vector cassification face is write as：W is weight vector, b For constant offset,For class interval, the optimization problem that class interval obtains maximum is represented by askingMinimum value ask Topic, constraints areAccording to KKT (Karush-Kuhn-Tucker) conditions and drawing Ge Lang functions obtain classification function：

Wherein, sgn () is sign function, a_iFor optimal Lagrange multiplier, x is sample to be tested, Point multiplication operation (the x of higher dimensional space_iX) by the kernel function k (x of lower dimensional space_i,x_j) replace, classification function is rewritable to be

4. the weather radar electromagnetic interference echo recognition methods based on SVMs as claimed in claim 1, conventional core Function is as follows：

Polynomial kernel function k (x, x_i)=[(xx_i)+1]^qQ=1,2,3 ...；

Gaussian radial basis functionσ is parameter；

Sigmoid function k (x, x_i)=tanh (b (x, x_i)+c) b, c be constant；

Fourier spaceWherein N is constant；

B-spline kernel function k (x, x_i)=B_2p+1(x-x_i) wherein B_2p+1(x) it is 2p+1 ranks B；

Different svm classifier functions can be produced by choosing different kernel functions, so as to obtain different SVM classifiers, therefore can be with Different svm classifier methods is expanded to by the selection of kernel function.