CN114373079A - Rapid and accurate ground penetrating radar target detection method - Google Patents

Abstract

The invention discloses a rapid and accurate ground penetrating radar target detection method, which comprises the following steps: in the model training phase: acquiring a positive and negative label sample set of the ground penetrating radar, extracting HOG characteristics, performing dimensionality reduction on the original high-dimensional HOG characteristics by using PCA, training and storing a classifier; in the target detection stage: preprocessing an original ground penetrating radar image, acquiring target candidate frames with different scales and length-width ratios by using a selective search algorithm, extracting HOG characteristics of images in the candidate frames, reducing dimensions of the original high-dimensional HOG characteristics by using PCA, and inputting the original high-dimensional HOG characteristics into a trained classifier to obtain a target detection result. The original HOG characteristics are subjected to dimensionality reduction through PCA, and the problems of overlong classifier identification time and poor generalization capability caused by high-dimensional HOG characteristics in a traditional ground penetrating radar target detection algorithm are solved; meanwhile, the invention obtains the target candidate frames with different scales and length-width ratios by using a selective search algorithm, thereby greatly improving the detection precision and efficiency.

Description

Rapid and accurate ground penetrating radar target detection method

Technical Field

The invention relates to the technical field of ground penetrating radar target detection, in particular to a rapid and accurate ground penetrating radar target detection method.

Background

In recent years, investigation of shallow underground has become of increasing interest. The ground penetrating radar is a nondestructive underground detection method, which utilizes a receiving antenna to obtain electromagnetic waves reflected by a target to explain underground data, and is widely applied to road pavement measurement, pipeline detection, surface mining and land mine detection at present.

At present, a lot of target detection algorithm researches based on ground penetrating radar images exist. The use of curve fitting in combination with an edge detection algorithm is a low complexity and easy to implement method. However, in a complex environment, the accuracy of the edge detection algorithm may not be satisfied due to the influence of clutter and image noise. Moreover, the time complexity of the curve fitting algorithm is too large, so that the curve fitting algorithm is difficult to be applied to the detection of the ground penetrating radar target in a large range. Recently, machine learning and deep learning have attracted great attention and show good performance in the ground penetrating radar image target detection scheme. In most cases, deep learning requires a very large amount of training data to achieve good performance. However, the actual data acquisition of the ground penetrating radar is difficult. Therefore, the machine learning method is more suitable for the current practical scene, and a plurality of ground penetrating radar target detection algorithms based on machine learning exist at present.

At present, Haar, EHD, SIFT, SURF, HOG and other features are used for detecting the ground penetrating radar target based on machine learning. Among these common features, HOG has proven to be a more suitable feature descriptor for ground penetrating radar target detection. The HOG features can extract the distribution of the magnitude and direction of the gradient, are robust to contrast and illumination, and have great success in detecting pedestrians and vehicles in video monitoring.

The existing ground penetrating radar target detection algorithm based on the HOG characteristics usually slides a fixed window in a ground penetrating radar image to obtain a target candidate frame, and the target candidate frame is sent to a classifier for target detection. However, the HOG features are not scale invariant and are very sensitive to the aspect ratio of the target in the ground penetrating radar image, which makes it difficult to detect the target of all scales and aspect ratios by the conventional fixed-size sliding window method. And the optimal sizes of the sliding windows of different ground penetrating radar images are manually determined, which is time-consuming and labor-consuming. And the sliding window method has high time complexity as an exhaustive search method.

In addition, the original HOG features typically have hundreds of dimensions, which weakens the generalization capability of the classifier and greatly increases the recognition time of the classifier.

In summary, the traditional ground penetrating radar target detection algorithm based on the HOG features has two defects:

1. too high HOG characteristic dimensionality leads to poor generalization capability and too long identification time of the classifier;

2. the conventional sliding window method is too time-complex and cannot detect targets having various scales and aspect ratios.

Disclosure of Invention

In view of the above, the present invention provides a method for detecting a ground penetrating radar target quickly and accurately, so as to solve the technical problems mentioned in the background art; aiming at the technical problem, the invention provides a method for reducing the dimension of the original high-dimensional HOG characteristic by using PCA, improving the generalization capability of a classifier and reducing the target identification time. Meanwhile, a selective search algorithm is used for replacing a traditional sliding window method to obtain target candidate frames with different scales and aspect ratios, and target detection precision and detection efficiency are greatly improved.

In order to achieve the purpose, the invention adopts the following technical scheme:

a fast and accurate ground penetrating radar target detection method comprises the following steps:

s1, acquiring a ground penetrating radar target positive and negative label data set for training;

step S2, extracting HOG characteristics in the data set acquired in the step S1;

step S3, performing dimensionality reduction processing on the HOG characteristics acquired in the step S2 by using a PCA method;

step S4, inputting the HOG features subjected to dimensionality reduction into a classifier network for training to obtain a trained detection model;

step S5, preprocessing an original ground penetrating radar image for target detection;

s6, aiming at the preprocessed ground penetrating radar image, obtaining target candidate frames with different scales and length-width ratios by using a selective search algorithm;

step S7, extracting HOG characteristics of the images in the candidate frame;

step S8, using PCA to perform dimension reduction processing on the HOG characteristics;

and S9, inputting the HOG characteristics subjected to dimension reduction into the detection model obtained in the step S4 for target detection, and obtaining a target detection result.

Further, in step S1, the ground penetrating radar target positive and negative label data set includes a training set and a test set, where the training set is an image actually acquired by the ground penetrating radar, and the test set is simulation data.

Further, the step S2 specifically includes:

step S201, calculating the horizontal gradient P of each pixel point_h(x, y) and a vertical gradient P_v(x, y), the expression is:

in formula (1) and formula (2), G (x, y) represents a gradation value of an image at an (x, y) pixel point;

step S202, calculating a gradient amplitude D (x, y) and a direction θ (x, y) of each pixel point according to the horizontal direction gradient and the vertical direction gradient obtained in step S201, where the expression is:

step S203, dividing each sample image into a plurality of cell units, and counting a gradient histogram of each cell unit, wherein the counting of the gradient histogram of each cell unit specifically includes: uniformly dividing the gradient direction of 0-360 degrees into 9 channels, selecting a corresponding channel by a pixel in each cell unit according to the gradient direction of the pixel, voting the corresponding channel based on the gradient amplitude of the pixel, and finally counting a 9-channel gradient histogram of the cell unit;

step S204, forming a plurality of cell units into a cell block, and counting a gradient histogram of each cell block, wherein the forming of the plurality of cell units into the cell block specifically includes: connecting the gradient histograms of the plurality of cell units in series and normalizing;

step S205, connecting all the gradient histograms of all the cell blocks of each sample image in series to obtain the final HOG feature x of the whole image_i；

Step S206, HOG characteristic x of each training sample_iThe composition feature matrix X ═ X₁,x₂,...,x_N]。

Further, the step S3 specifically includes:

step S301, centralizing the feature matrix obtained in step S206, which includes: let X be ═ X₁,x₂,...,x_N]Is a feature matrix composed of N M-dimensional HOG features and used for training, and then centralizing the feature matrix

Expressed as:

step S302, aiming at the characteristic matrix obtained in step S301

And calculating a covariance matrix, wherein the expression is as follows:

step S303, the covariance matrix C_XPerforming characteristic decomposition to obtain: c_XXi ═ λ xi, where λ ═ diag { λ₁,λ₂,...,λ_M}，λ_iIs a characteristic value in descending order, xi ═ xi₁,ξ₂,...,ξ_M]Is λ_iCorresponding feature vector xi_iA matrix of compositions;

step S304, feature dimension reduction, comprising: forming a projection matrix P ([ xi ]) by the eigenvectors corresponding to the large eigenvalue of front k₁,ξ₂,...,ξ_k]Then, the HOG feature matrix Y reduced to k dimensions is expressed as

Y＝P^TX (5)

In equation (5), X represents the HOG feature X of all training samples_iAnd forming a feature matrix.

Further, the classifier network is an SVM, and the SVM selects a gaussian kernel function.

Further, the step S5 specifically includes:

step S501, performing direct current filtering processing on the original image, wherein the expression is as follows:

in formula (6), I' (x, y) is represented as an image subjected to dc filtering, I (x, y) is represented as an original image of the ground penetrating radar, and 2w₁+1 denotes the sliding window length, k denotes the kth column of the image to be processed;

step S502, clutter suppression is carried out on the image I' (x, y) after direct current filtering by using an averaging method, and the expression is as follows:

in formula (7), I ″ (x, y) is expressed as an image after clutter suppression, 2w₂+1 denotes the sliding window length, k denotes the kth column of the image to be processed.

Further, the step S6 specifically includes:

step S601, obtaining an initial segmentation region R ═ { R ═ R from the preprocessed ground penetrating radar image₁,r₂,...,r_nAnd initialize similarity sets

Step S602, calculating two adjacent areas r_i,r_jSimilarity in color, including:

counting a certain region r of the image_iObtaining a color vector normalized by L1 norm for each color channel bins of 25

Two adjacent regions r_i,r_jThe color similarity between the two formulas is as follows:

step S603, calculating a texture similarity between the two regions, including:

local binary pattern is adopted to represent the texture characteristics, namely the region r_iTexture feature representation of

Then two adjacent regions r_i,r_jThe texture similarity calculation formula is as follows:

step S604, calculating two adjacent areas r_i,r_jThe size similarity between the two is expressed as:

in formula (8), size (·) represents the number of pixels in the region, and size (im) represents the number of pixels in the whole image;

step S605, calculating the matching similarity as follows:

in formula (9), size (BB)_ij) Indicating the region r_i,r_jThe size of the circumscribed rectangle of (1);

step S606, similarity S according to color aspect_color(r_i,r_j) Similarity of texture s_texture(r_i,r_j) Size similarity s_size(r_i,r_j) And the similarity of anastomosis s_fill(r_i,r_j) Calculating the total similarity, wherein the expression is as follows:

s(r_i,r_j)＝ε₁s_color(r_i,r_j)+ε₂s_texture(r_i,r_j)+ε₃s_size(r_i,r_j)+ε₄s_fill(r_i,r_j) (10)

in the formula (10), ε₁，ε₂，ε₃，ε₄Weighting coefficients for different degrees of similarity;

step S607, similarity S (r) of all two adjacent areas_i,r_j) Calculating and adding a similarity set S;

step S608, two areas r corresponding to the maximum value in the similarity set S_i,r_jAre merged into the same region r_iS (r)_i,r_j) Removing and dividing the region r from the similarity set S_iAdding R to the region set R_iAdding the similarity with the adjacent region into a similarity set S;

step S609, repeating the step S608 until the similarity set

The subset in the region set R is then the final candidate region of the image.

Further, the step S7 specifically includes:

s701, zooming the image in the candidate frame area obtained by the selective search algorithm in the step S6 to make the size of the image consistent with that of the training sample;

step S702, adopting the method in the step S2 to extract the HOG characteristics of the image and then form an original high-dimensional HOG characteristic matrix X_detect。

Further, the step S8 specifically includes:

using the projection matrix P obtained in the training stage to perform the comparison on the original high-dimensional HOG feature matrix X obtained in step S7_detectReducing dimension of the feature, and reducing dimension of the feature matrix Y_detectThe expression is as follows: y is_detect＝P^TX_detect。

Further, the step S9 specifically includes:

step S901, reducing the dimension of the HOG feature matrix Y_detectInputting the target candidate frames into the trained SVM to judge which target candidate frames are target areas;

and step S902, performing non-maximum suppression on all the candidate frames judged as the targets, removing redundant candidate frames, and finally outputting a target detection result.

The invention has the beneficial effects that:

1. in general, the original HOG features are very high in dimensionality, resulting in degraded classifier performance. The invention uses the PCA method to reduce the dimension of the original HOG characteristics, greatly improves the generalization capability of the classifier and simultaneously can improve the target detection efficiency.

2. According to the traditional ground penetrating radar target detection algorithm based on the HOG characteristics, a target candidate region is obtained through a sliding window and is sent to a classifier for detection and identification, but the time complexity of the sliding window method is high, and due to the fact that the HOG characteristics are very sensitive to the size of a target, the sliding window with the fixed size cannot detect the target with different sizes in a ground penetrating radar image, and the detection performance is poor. The invention uses the selective search algorithm to replace the traditional sliding window, adaptively selects the candidate frames with different scales and length-width ratios, and sends the candidate frames into the classifier for detection and identification, thereby greatly improving the target detection efficiency and the detection precision.

Drawings

Fig. 1 is a schematic flowchart of a method for detecting a target of a ground penetrating radar in a fast and accurate manner provided in embodiment 1;

FIG. 2 is a schematic diagram of a training positive sample provided in example 1;

FIG. 3 is a schematic diagram of a training negative example provided in example 1;

fig. 4 is a visualized image of the training positive sample HOG feature provided in example 1;

FIG. 5 is a visual image of the training negative sample HOG feature provided in example 1;

FIG. 6 is a test image 1 used in the method of embodiment 1 for verifying the target detection performance, where the test image 1 is preprocessed;

FIG. 7 is a candidate box obtained by the selective search algorithm of the test image 1 provided in example 1;

fig. 8 is the target detection result finally obtained from the test image 1 provided in example 1.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example 1

Referring to fig. 1 to 8, the embodiment provides a method for detecting a ground penetrating radar target quickly and accurately, which includes two stages of training and detecting, specifically, the training stage includes the following steps:

s1, acquiring a ground penetrating radar target positive and negative label data set for training;

specifically, in the present embodiment, the positive and negative label data sets include a training set and a test set, wherein the training set is derived from images acquired by an actual ground penetrating radar, and the sample size is 33 × 52 pixels. The test set used to evaluate the target test performance is derived from the simulation data.

Step S2, extracting HOG characteristics of the positive and negative label data sets;

specifically, in this embodiment, the HOG feature extraction in step S2 specifically includes:

step S201, calculating the horizontal direction gradient P of each pixel point by using the following formula_h(x, y) and a vertical gradient P_v(x,y)：

P_h(x,y)＝G(x-1,y+1)+2G(x,y+1)

+G(x+1,y+1)-G(x-1,y-1)

-2G(x,y-1)-G(x+1,y-1)

P_v(x,y)＝G(x-1,y-1)+2G(x-1,y)

+G(x-1,y+1)-G(x+1,y-1)

-2G(x+1,y)-G(x+1,y+1)

Wherein G (x, y) represents the gray value of the image at the (x, y) pixel point;

step S202, the gradient amplitude D (x, y) and the direction theta (x, y) of each pixel point are respectively calculated by using the following formulas, and the expression is as follows:

step S203, dividing each sample image into a plurality of cell units, and counting a gradient histogram of each cell unit. The statistical method of the gradient histogram of each cell unit comprises the following steps: uniformly dividing the gradient direction of 0-360 degrees into 9 channels, selecting a corresponding channel by a pixel in each cell unit according to the gradient direction of the pixel, voting the corresponding channel based on the gradient amplitude of the pixel, and finally counting a 9-channel gradient histogram of the cell unit;

step S204, a plurality of cell units are combined into cell blocks, and the gradient histogram of each cell block is counted. The specific method is to connect the gradient histograms of each cell unit in series and normalize the histograms;

step S205, connecting all the gradient histograms of all the cell blocks of each training sample image in series to obtain the final HOG feature x of the whole image_i；

Step S206, HOG characteristic x of each training sample_iThe composition feature matrix X ═ X₁,x₂,...,x_N]。

Step S3, performing dimensionality reduction on the original HOG characteristic by using a PCA method;

specifically, in this embodiment, the specific step of step S3 is:

step S301, centralizing the feature matrix for training, including: let X be ═ X₁,x₂,...,x_N]N M-dimensional HOG features form a feature matrix for training. Feature matrix after centering

Can be calculated as

Step S302, calculateA covariance matrix. Post-centering feature matrix

Can be expressed as

And step S303, characteristic decomposition. Covariance matrix C_XCan be expressed as

C_Xξ＝λξ

Wherein λ ═ diag { λ₁,λ₂,...,λ_M}，λ_iIs a characteristic value in descending order, xi ═ xi₁,ξ₂,...,ξ_M]Is λ_iCorresponding feature vector xi_iA matrix of components.

Step S304, feature dimension reduction, which comprises the following steps: the eigenvectors corresponding to the large eigenvalues of the front k form a projection matrix P ═ xi₁,ξ₂,...,ξ_k]. The HOG feature matrix Y reduced to k dimensions can be expressed as:

Y＝P^TX

step S4, inputting the HOG features subjected to dimensionality reduction into a classifier network for training, and storing the trained model for subsequent target detection;

specifically, in the embodiment, the classifier network selects an SVM, and the SVM selects a gaussian kernel function, where C and gamma select appropriate parameter values according to actual conditions.

More specifically, in this example, C is set to 1 and gamma is set to 0.5. And inputting HOG characteristics of all sample images subjected to dimensionality reduction into an SVM for training, and finally storing the trained model for subsequent target detection.

Specifically, in this embodiment, the target detection stage includes the following steps:

step S5, preprocessing an original ground penetrating radar image for target detection;

specifically, in this embodiment, the step S5 specifically includes the following steps:

step S501, performing direct current filtering processing on the original image, wherein the expression is as follows:

wherein, I' (x, y) is an image which is processed by direct current filtering, I (x, y) is an original image of the ground penetrating radar, and 2w₁+1 denotes the sliding window length;

step S502, clutter suppression is carried out on the image I' (x, y) after direct current filtering by using an averaging method, and the expression is as follows:

where I "(x, y) is represented as the clutter suppressed image, 2w₂+1 denotes the sliding window length.

S6, applying a selective search algorithm to the preprocessed ground penetrating radar image to obtain target candidate frames with different scales and aspect ratios;

specifically, in this embodiment, the step S6 specifically includes the following steps:

step S601, acquiring an initial segmentation region R ═ { R ═ R in the ground penetrating radar image preprocessed by the image segmentation algorithm based on the graph theory₁,r₂,...,r_nAnd initialize similarity sets

Step S602, calculating two adjacent areas r_i,r_jSimilarity in color s_color(r_i,r_j). The color histogram of each color channel bins of the statistical image is 25. For a certain area r of a color RGB image_iThe L1 norm normalized color vector can be obtained

The color similarity calculation formula is as follows:

step S603, calculating a texture similarity between the two regions. There are many ways to extract the local features of the image to represent the texture similarity between two adjacent regions, in this example, the Local Binary Pattern (LBP) is used to represent the texture features, and the region r_iTexture feature representation of

Then two adjacent regions r_i,r_jThe texture similarity calculation formula is as follows:

step S604, calculating two adjacent areas r_i,r_jThe size similarity between them is:

wherein size (·) represents the number of pixels in the region, and size (im) represents the number of pixels in the whole image;

step S605, calculating the matching similarity as:

wherein size (BB)_ij) Indicating the region r_i,r_jThe size of the circumscribed rectangle of (1);

step S606, similarity S (r)_i,r_j) Obtained by weighting the four similarity degrees, and the calculation process is

s(r_i,r_j)＝ε₁s_color(r_i,r_j)+ε₂s_texture(r_i,r_j)+ε₃s_size(r_i,r_j)+ε₄s_fill(r_i,r_j)

Wherein epsilon₁，ε₂，ε₃，ε₄Weighting coefficients for different degrees of similarity, in this example, ε₁，ε₂，ε₃And ε₄All are set to be 1;

step S607, similarity S (r) of all two adjacent areas_i,r_j) Calculating and adding a similarity set S;

step S608, two areas r corresponding to the maximum value in the similarity set S_i,r_jAre merged into the same region r_iS (r)_i,r_j) Removing and dividing the region r from the similarity set S_iAdded to the region set R. Will r is_iAdding the similarity with the adjacent region into a similarity set S;

step S609, repeating the step S608 until the similarity set

The subset in the region set R is then the final candidate region of the image.

Step S7, extracting HOG characteristics of the images in the candidate frame;

specifically, in this embodiment, the step S7 specifically includes:

step S701, scaling the size of the image in the candidate frame with various dimensions and aspect ratios obtained by the selective search algorithm to be consistent with the training sample, that is, to 33 × 52 pixels in this example;

step S702, extracting HOG characteristics like step S2 to form an original high-dimensional HOG characteristic matrix X_detect。

Step S8, using PCA to perform dimension reduction processing on the HOG characteristics;

specifically, in this embodiment, the step S8 specifically includes:

using the projection matrix P obtained in the training phase, forOriginal high-dimensional HOG feature matrix X obtained in step S7_detectReducing the dimension of the feature into a k-dimension feature matrix Y_detectThe calculation formula is as follows:

Y_detect＝P^TX_detect

step S9, inputting the HOG features after dimension reduction into the SVM which is trained in the training stage to obtain a target detection result;

specifically, in this embodiment, the step S9 specifically includes:

step S901, reducing the dimension of the HOG feature matrix Y_detectInputting the result into the SVM trained in the step S4 to judge which target candidate boxes are target areas;

s902, sorting all the candidate frames judged as targets according to the confidence coefficient from large to small;

step S903, selecting the candidate frame with the maximum confidence, and suppressing, namely deleting, all the other candidate frames with the intersection ratio with the candidate frame being greater than the threshold (set to be 0.01 in the example), if the intersection ratio with the candidate frame is smaller than the threshold, keeping, and adding the candidate frame with the maximum confidence into the final detection frame set;

step S904, continuously finding the candidate frame with the maximum confidence coefficient from the rest candidate frames, and repeating the step S903 until all the frames are processed;

in step S905, the output detection results, that is, all the detection frames included in the final detection frame set.

Specifically, fig. 2 and 3 are positive and negative samples of training, fig. 4 and 5 are HOG feature visualization results extracted from the training samples, fig. 6 is a test image after preprocessing for target detection, fig. 7 is a candidate box obtained by the test image through a selective search algorithm, and fig. 8 is a final target detection result.

The invention is not described in detail, but is well known to those skilled in the art.

The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations could be devised by those skilled in the art in light of the present teachings without departing from the inventive concepts. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.

Claims (10)

1. A fast and accurate ground penetrating radar target detection method is characterized by comprising the following steps:

s1, acquiring a ground penetrating radar target positive and negative label data set for training;

step S2, extracting HOG characteristics in the data set acquired in the step S1;

step S3, performing dimensionality reduction processing on the HOG characteristics acquired in the step S2 by using a PCA method;

step S4, inputting the HOG features subjected to dimensionality reduction into a classifier network for training to obtain a trained detection model;

step S5, preprocessing an original ground penetrating radar image for target detection;

s6, aiming at the preprocessed ground penetrating radar image, obtaining target candidate frames with different scales and length-width ratios by using a selective search algorithm;

step S7, extracting HOG characteristics of the images in the candidate frame;

step S8, using PCA to perform dimension reduction processing on the HOG characteristics;

and S9, inputting the HOG characteristics subjected to dimension reduction into the detection model obtained in the step S4 for target detection, and obtaining a target detection result.

2. The method for detecting the target of the ground penetrating radar as claimed in claim 1, wherein in the step S1, the data set of positive and negative labels of the target of the ground penetrating radar includes a training set and a test set, wherein the training set is an image actually acquired by the ground penetrating radar, and the test set is simulation data.

3. The method for detecting the target of the georadar according to claim 1, wherein the step S2 specifically includes:

step S201, calculating the horizontal gradient P of each pixel point_h(x, y) and a vertical gradient P_v(x, y), the expression is:

in formula (1) and formula (2), G (x, y) represents a gradation value of an image at an (x, y) pixel point;

step S202, calculating a gradient amplitude D (x, y) and a direction θ (x, y) of each pixel point according to the horizontal direction gradient and the vertical direction gradient obtained in step S201, where the expression is:

step S203, dividing each sample image into a plurality of cell units, and counting a gradient histogram of each cell unit, wherein the counting of the gradient histogram of each cell unit specifically includes: uniformly dividing the gradient direction of 0-360 degrees into 9 channels, selecting a corresponding channel by a pixel in each cell unit according to the gradient direction of the pixel, voting the corresponding channel based on the gradient amplitude of the pixel, and finally counting a 9-channel gradient histogram of the cell unit;

step S204, forming a plurality of cell units into a cell block, and counting a gradient histogram of each cell block, wherein the forming of the plurality of cell units into the cell block specifically includes: connecting the gradient histograms of the plurality of cell units in series and normalizing;

step S205, connecting all the gradient histograms of all the cell blocks of each sample image in series to obtain the final HOG feature x of the whole image_i；

Step S206, HOG characteristic x of each training sample_iThe composition feature matrix X ═ X₁,x₂,...,x_N]。

4. The method for detecting the target of the georadar as claimed in claim 3, wherein the step S3 specifically includes:

step S301, centralizing the feature matrix obtained in step S206, which includes: let X be ═ X₁,x₂,...,x_N]Is a feature matrix composed of N M-dimensional HOG features and used for training, the feature matrix X after centering is represented as:

step S302, aiming at the characteristic matrix obtained in step S301

And calculating a covariance matrix, wherein the expression is as follows:

step S303, the covariance matrix C_XPerforming characteristic decomposition to obtain: c_XXi ═ λ xi, where λ ═ diag { λ₁,λ₂,...,λ_M}，λ_iIs a characteristic value in descending order, xi ═ xi₁,ξ₂,...,ξ_M]Is λ_iCorresponding feature vector xi_iA matrix of compositions;

step S304, feature dimension reduction, comprising: forming a projection matrix P ([ xi ]) by the eigenvectors corresponding to the large eigenvalue of front k₁,ξ₂,...,ξ_k]Then, the HOG feature matrix Y reduced to k dimensions is expressed as

Y＝P^TX (5)

In equation (5), X represents the HOG feature X of all training samples_iAnd forming a feature matrix.

5. The method of claim 4, wherein the classifier network is an SVM, and the SVM selects a Gaussian kernel function.

6. The method for detecting the target of the georadar as claimed in claim 5, wherein the step S5 specifically includes:

step S501, performing direct current filtering processing on the original image, wherein the expression is as follows:

in formula (6), I' (x, y) is represented as an image subjected to dc filtering, I (x, y) is represented as an original image of the ground penetrating radar, and 2w₁+1 denotes the sliding window length, k denotes the kth column of the image to be processed;

step S502, clutter suppression is carried out on the image I' (x, y) after direct current filtering by using an averaging method, and the expression is as follows:

in formula (7), I ″ (x, y) is expressed as an image after clutter suppression, 2w₂+1 denotes the sliding window length, k denotes the kth column of the image to be processed.

7. The method for detecting the target of the georadar according to claim 6, wherein the step S6 specifically includes:

step S601, obtaining an initial segmentation region R ═ { R ═ R from the preprocessed ground penetrating radar image₁,r₂,...,r_nAnd initialize similarity sets

Step S602, calculating two adjacent areas r_i,r_jSimilarity in color, including:

counting a certain region r of the image_iObtaining a color vector normalized by L1 norm for each color channel bins of 25

Two adjacent regions r_i,r_jThe color similarity between the two formulas is as follows:

step S603, calculating a texture similarity between the two regions, including:

local binary pattern is adopted to represent the texture characteristics, namely the region r_iTexture feature representation of

Then two adjacent regions r_i,r_jThe texture similarity calculation formula is as follows:

step S604, calculating two adjacent areas r_i,r_jThe size similarity between the two is expressed as:

in formula (8), size (·) represents the number of pixels in the region, and size (im) represents the number of pixels in the whole image;

step S605, calculating the matching similarity as follows:

in formula (9), size (BB)_ij) Indicating the region r_i,r_jThe size of the circumscribed rectangle of (1);

step S606, similarity S according to color aspect_color(r_i,r_j) Similarity of texture s_texture(r_i,r_j) Size similarity s_size(r_i,r_j) And the similarity of anastomosis s_fill(r_i,r_j) Calculating the total similarity, wherein the expression is as follows:

s(r_i,r_j)＝ε₁s_color(r_i,r_j)+ε₂s_texture(r_i,r_j)+ε₃s_size(r_i,r_j)+ε₄s_fill(r_i,r_j) (10)

in the formula (10), ε₁，ε₂，ε₃，ε₄Weighting coefficients for different degrees of similarity;

step S607, similarity S (r) of all two adjacent areas_i,r_j) Calculating and adding a similarity set S;

step S608, two areas r corresponding to the maximum value in the similarity set S_i,r_jAre merged into the same region r_iS (r)_i,r_j) Removing and dividing the region r from the similarity set S_iAdding R to the region set R_iAdding the similarity with the adjacent region into a similarity set S;

step S609, repeating the step S608 until the similarity set

The subset in the region set R is then the final candidate region of the image.

8. The method for detecting the target of the georadar according to claim 7, wherein the step S7 specifically includes:

s701, zooming the image in the candidate frame area obtained by the selective search algorithm in the step S6 to make the size of the image consistent with that of the training sample;

step S702, adopting the method in the step S2 to extract the HOG characteristics of the image and then form an original high-dimensional HOG characteristic matrix X_detect。

9. The method for detecting the target of the georadar according to claim 8, wherein the step S8 specifically includes:

using the projection matrix P obtained in the training stage to perform the comparison on the original high-dimensional HOG feature matrix X obtained in step S7_detectReducing dimension of the feature, and reducing dimension of the feature matrix Y_detectThe expression is as follows: y is_detect＝P^TX_detect。

10. The method for detecting the target of the georadar according to claim 9, wherein the step S9 specifically includes:

step S901, reducing the dimension of the HOG feature matrix Y_detectInputting the target candidate frames into the trained SVM to judge which target candidate frames are target areas;

and step S902, performing non-maximum suppression on all the candidate frames judged as the targets, removing redundant candidate frames, and finally outputting a target detection result.

Images