CN111091071B

CN111091071B - Underground target detection method and system based on ground penetrating radar hyperbolic wave fitting

Info

Publication number: CN111091071B
Application number: CN201911185419.8A
Authority: CN
Inventors: 原达; 王崴; 陈飞凡; 李文生; 王冬雨; 苗翠; 崔嘉傲
Original assignee: Shandong Technology and Business University
Current assignee: Shandong Technology and Business University
Priority date: 2019-11-27
Filing date: 2019-11-27
Publication date: 2023-12-01
Anticipated expiration: 2039-11-27
Also published as: CN111091071A

Abstract

The application discloses an underground target detection method and system based on ground penetrating radar hyperbolic wave fitting, comprising the following steps: acquiring a ground penetrating radar GPR image; performing super-pixel segmentation on the GPR image to obtain a super-pixel segmented image containing a plurality of areas; each region is referred to as a super pixel block; the pixels of all points of each super pixel block are equal to the pixels of the center point of the current super pixel block; extracting the regions of interest from the super-pixel segmented image to extract a plurality of regions of interest; performing image binarization processing on each region of interest to obtain a binarized image; extracting a bone region from the binarized image; extracting hyperbolic waves from the bone region; and carrying out underground target detection according to the waveform, amplitude intensity and time change characteristics of the hyperbolic wave.

Description

Underground target detection method and system based on ground penetrating radar hyperbolic wave fitting

Technical Field

The disclosure relates to the technical field of underground target detection, in particular to an underground target detection method and system based on ground penetrating radar hyperbolic wave fitting.

Background

The statements in this section merely mention background art related to the present disclosure and do not necessarily constitute prior art.

Identification and fitting of hyperbolic waves in GPR images is very difficult. Ground Penetrating Radar (GPR) is used as a nondestructive detection device based on high-frequency electromagnetic waves, and when the electromagnetic waves propagate in an underground medium, hyperbolic-shaped reflection lines, called hyperbolic waves, are generated when the electromagnetic waves encounter an electrical difference interface of an underground target. Important information such as the spatial position, structure, morphology, burial depth and the like of the underground target can be deduced through characteristics such as waveform, amplitude intensity, time change and the like of the hyperbolic wave. This is therefore an important indicator for understanding the characteristics of the subsurface target. However, due to system noise and inhomogeneities of the subsurface medium, the resulting image is very complex, and hence it is necessary to extract and fit hyperbolic waves to the GPR image in real time.

Among the previous methods of identifying, fitting hyperbolic waves, there can be roughly classified into a mathematical calculation method other than machine learning and a calculation method by machine learning and a hybrid algorithm. The non-machine mathematical calculation method is distinguished according to the difference of the characteristics of the hyperbolic wave region and other regions, and the calculation method through machine learning is performed by training through a positive sample containing the hyperbolic wave and a negative sample containing the non-hyperbolic wave, and finally reducing the region containing the hyperbolic wave, and then selecting a reasonable region.

In the process of implementing the present disclosure, the inventor finds that the following technical problems exist in the prior art:

the prior art uses a method based on the hough transformation in a generalized mode, but is very time-consuming and has high parameter requirements, because the hough transformation needs to be performed in a four-dimensional space, and the precision of the hough transformation depends on discretization of parameters; the least square method is used in the prior art, but the least square method can be regarded as an extension of the generalized Hough transform, so that the method is not greatly improved. Chen and Cohn et al propose C3 sequences that are scanned horizontally and then judged by the sequence without taking into account the hyperbolic geometry. Collins L M and Torrione PA et al propose an idea of extraction of key points. The fritze et al first used an edge adaptation algorithm on the GPR image, but this algorithm had a certain requirement on the purity of the GPR image, so this algorithm value was applicable to GPR images with little noise. The prior art has detected the apex of the hyperbola but lost other parameters of the associated hyperbola. However, this is essential for identifying other properties of the object. zhou, chen et al propose a method of opening down the column segment. The gradient in the y-direction is calculated for the GPR image first, then a series of optimizations are performed to find the feature column segment that opens downward, and then judgment and recognition are performed.

Methods based on machine learning are used in the prior art. The method comprises the steps of extracting characteristics of a hyperbolic wave sample, combining a machine learning technology, so that the area where the hyperbolic wave is located is reduced, and then, applying a fitting method to find parameters of the hyperbolic wave, so as to judge whether the position containing the waveform in a GPR image has threat or not, wherein the result depends on the quality and the quantity of training data.

Disclosure of Invention

In order to solve the defects of the prior art, the disclosure provides an underground target detection method and system based on ground penetrating radar hyperbolic wave fitting;

in a first aspect, the present disclosure provides a method of subsurface target detection based on ground penetrating radar hyperbolic wave fitting;

the underground target detection method based on the ground penetrating radar hyperbolic wave fitting comprises the following steps:

acquiring a ground penetrating radar GPR image;

performing super-pixel segmentation on the GPR image to obtain a super-pixel segmented image containing a plurality of areas; each region is referred to as a super pixel block; the pixels of all points of each super pixel block are equal to the pixels of the center point of the current super pixel block;

extracting the regions of interest from the super-pixel segmented image to extract a plurality of regions of interest; performing image binarization processing on each region of interest to obtain a binarized image;

extracting a bone region from the binarized image;

extracting hyperbolic waves from the bone region;

and carrying out underground target detection according to the waveform, amplitude intensity and time change characteristics of the hyperbolic wave.

In a second aspect, the present disclosure also provides an underground target detection system based on ground penetrating radar hyperbolic wave fitting;

underground target detecting system based on ground penetrating radar hyperbolic wave fitting includes:

an acquisition module configured to: acquiring a ground penetrating radar GPR image;

a superpixel segmentation module configured to: performing super-pixel segmentation on the GPR image to obtain a super-pixel segmented image containing a plurality of areas; each region is referred to as a super pixel block; the pixels of all points of each super pixel block are equal to the pixels of the center point of the current super pixel block;

a region of interest extraction module configured to: extracting the regions of interest from the super-pixel segmented image to extract a plurality of regions of interest; performing image binarization processing on each region of interest to obtain a binarized image;

a bone region extraction module configured to: extracting a bone region from the binarized image;

a hyperbolic wave extraction module configured to: extracting hyperbolic waves from the bone region;

a subsurface target detection module configured to: and carrying out underground target detection according to the waveform, amplitude intensity and time change characteristics of the hyperbolic wave.

In a third aspect, the present disclosure also provides an electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the steps of the method of the first aspect.

In a fourth aspect, the present disclosure also provides a computer readable storage medium storing computer instructions which, when executed by a processor, perform the steps of the method of the first aspect.

Compared with the prior art, the beneficial effects of the present disclosure are:

firstly, super-pixel segmentation is carried out on a GPR image under a real condition, a multi-scale neighborhood calculation method is used for obtaining a salient value of each super-pixel block, and a plurality of salient images are obtained by setting different thresholds. And finally, respectively binarizing the images, extracting skeleton areas in the binary images, finally selecting areas possibly containing hyperbolic waves, and finally verifying to obtain an equation of the hyperbolic waves.

The result shows that the algorithm can well process the influence of various intersections and noises on extraction, and the algorithm can efficiently, accurately and quickly extract information in hyperbolic waves in the GPR image under the real condition. The effectiveness of the method is proved by the real verification, and the algorithm has stronger robustness than the prior algebraic distance algorithm.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application.

FIG. 1 is a flow chart of a method of a first embodiment;

FIG. 2 (a) is a graph showing a first embodiment of a gray-white-black-gray-white dominant hyperbolic wave;

FIG. 2 (b) is a graph showing a gray-black-white-gray-white dominant hyperbolic wave of the first embodiment;

FIG. 2 (c) shows a black dominant hyperbolic wave exhibiting gray-black-white-gray for the first embodiment;

FIG. 3 is a schematic diagram of an encoding mode of an 8 neighborhood in the bone region extraction module according to the first embodiment;

FIG. 4 shows the effect of a hyperbolic template in the underground object detection module when a is 10, 30, 50, and e is 1.05,1.5, and 2, respectively;

FIG. 5 shows 4 effective downwardly opening column sections;

FIG. 6 (a) is a schematic view showing the expansion of the effective point without waveform crossing in the first embodiment;

fig. 6 (b) is a schematic diagram showing the intersection of the hyperbolic wave of the first embodiment with other curved directions (the directions of the left wing and other curved directions of the hyperbolic wave are all from top right to bottom left);

fig. 6 (c) is a schematic diagram of the intersection of the hyperbolic wave of the first embodiment with other curves in the same direction (the left wing of the hyperbolic wave goes from top right to bottom left, and the other curves go from top left to bottom right);

FIG. 7 is a diagram showing the matching criteria between the active clusters and templates according to the first embodiment;

FIG. 8 (a) is an example of a hyperbolic wave to be identified and fitted for the first embodiment;

FIG. 8 (b) shows the location of the initially selected hyperbolic wave of the first embodiment;

fig. 8 (c) shows the fitting effect of the hyperbolic wave of the first embodiment.

Detailed Description

It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the present application. As used herein, the singular is also intended to include the plural unless the context clearly indicates otherwise, and furthermore, it is to be understood that the terms "comprises" and/or "comprising" when used in this specification are taken to specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof.

Introduction of technical terms:

SLIC, simple linear iterative clustering, is used in the present disclosure as a best super-pixel segmentation algorithm, and has the characteristics of few parameters, high speed, compact and orderly division.

The concept of significance is proposed by itti et al. Since primate vision systems are able to quickly identify salient regions in a scene in front of the eye, itti et al use this idea in computer images so that salient regions in the images can be located. This has the advantage that the "region of interest" can be selected instead of analysing the whole image. Later, the identification of salient regions is widely applied to image retrieval, adaptive content transfer, interested adaptive regions, and Achanta et al propose a method for local contrast comparison under a multi-scale neighborhood based on multiple aspects such as image compression and intelligent image adjustment, and can segment images well, so that the method is suitable for identification of hyperbolic waves in GPR images.

Skeleton extraction algorithm since waveforms in GPR images are all linear, skeleton extraction algorithm can effectively extract key information of waveforms or other noise areas in the region of interest, which simplifies the later judgment, and is very effective for further removing non-hyperbolic wave areas. Through the skeleton extraction algorithm, all the connected areas are finally changed into connected lines consisting of lines.

The classical algorithm of the opening and closing operation can enable edge points of the binary image to be smoother, so that the method is very suitable for the binary image obtained through saliency calculation.

An embodiment I provides an underground target detection method based on ground penetrating radar hyperbolic wave fitting;

as shown in fig. 1, the underground target detection method based on the ground penetrating radar hyperbolic wave fitting comprises the following steps:

s1: acquiring a ground penetrating radar GPR image;

s2: performing super-pixel segmentation on the GPR image to obtain a super-pixel segmented image containing a plurality of areas; each region is referred to as a super pixel block; the pixels of all points of each super pixel block are equal to the pixels of the center point of the current super pixel block;

s3: extracting the regions of interest from the super-pixel segmented image to extract a plurality of regions of interest; performing image binarization processing on each region of interest to obtain a binarized image;

s4: extracting a bone region from the binarized image;

s5: extracting hyperbolic waves from the bone region;

s6: and carrying out underground target detection according to the waveform, amplitude intensity and time change characteristics of the hyperbolic wave.

As one or more embodiments, the performing super-pixel segmentation on the GPR image to obtain a super-pixel segmented image including a plurality of regions; the method comprises the following specific steps:

performing super-pixel segmentation on the GPR image by using a SILC super-pixel segmentation algorithm to obtain a super-pixel segmentation image containing a plurality of areas; clustering discrete adjacent points in the segmented super-pixel segmented image, and finally realizing the effect of replacing the area by the center point.

As one or more embodiments, the extracting of the region of interest from the super-pixel segmented image extracts a plurality of regions of interest; the method comprises the following specific steps:

s31: calculating a salient value of each super pixel block of the super pixel segmentation image;

where τ represents the neighborhood at different scales, λ (τ) represents the weight of the saliency map computed under neighborhood τ, a represents the number of superpixels within the neighborhood under neighborhood τ,gray value representing the center of the ith cell, is->Gray value, ψ, representing the center of the jth cell _i Representing the saliency value of the ith super pixel area.

S32: setting a plurality of different thresholds;

the threshold values are +/-0.5, +/-0.65, +/-1.3, +/-1.5, +/-1.8 and +/-2.25 times of the average value of the significant values corresponding to all the super pixel blocks respectively;

s33: according to a plurality of different thresholds, a plurality of saliency maps are obtained:

for the positive number threshold, a super-pixel block with a significant value greater than the positive number threshold is changed to a foreground, and a super-pixel block with a significant value less than the positive number threshold is changed to a background;

for the negative threshold, super-pixel blocks with significance values smaller than the negative threshold are changed into foreground, and super-pixel blocks with significance values larger than the negative threshold are changed into background;

thereby obtaining a plurality of saliency maps.

As one or more embodiments, the neighborhood selection mode of the neighborhood τ includes a selection mode of a circular neighborhood and a selection mode of a column neighborhood; wherein,

the selection mode of the circular neighborhood is as follows: given a radius R, the center of each super pixel block is taken as a circle center, and for each center c _i Selecting all center sets { c } in the center sets in the circular neighborhood τ ₁ ,c ₂ ,c ₃ ...c _j }：

D(c _i ,c _j )<R；(8)

Wherein D (c) _i ,c _j ) Representation c _i And c _j The distance between the two at the location in the image.

The column neighborhood is selected in the following way: select center point to be in interval x _min ,x _max ]Wherein x is equal to the total number of super pixel blocks in (a) _min Representing the left boundary of a super pixel block, x _max Representing the right boundary of the super pixel block.

As one or more embodiments, the extracting bone regions from the binarized image; the method comprises the following specific steps: using skeleton extraction algorithm, all connected regions in the saliency map are represented by linear skeletons, and skeleton images are obtained.

As one or more embodiments, the extracting hyperbolic waves from the bone region; the method comprises the following specific steps:

defining a matching set, pre-screening hyperbolic wave vertexes in a skeleton image, judging and fitting the hyperbolic wave by using the matching set, and extracting the hyperbolic wave.

Further, pre-screening hyperbolic wave vertices in the bone image; the method comprises the following specific steps: and screening a line segment with a downward opening, and taking the midpoint of a flat area in the structure of ascending-flat-descending as a pre-screened hyperbolic wave vertex from the line segment with the downward opening.

Further, the matching set is used for judging and fitting the hyperbolic wave, and the hyperbolic wave is extracted; the method comprises the following specific steps:

for each vertex, in all branches, the one that most matches the matching set is found as the waveform for that vertex.

In practical situations, the waveforms in the generated GPR image will overlap in many ways, such as overlapping of hyperbolic waves and clutter, overlapping of clutter and clutter, etc., due to noise effects or due to too close a number of measured objects. However, these waveforms all have the potential to remain in the saliency calculation algorithm, exhibiting various intersections. The individual paths formed by these intersections are referred to as branches.

It should be appreciated that the SILC super pixel segmentation algorithm is applied to a single channel GPR image. The advantage of using super-pixels to segment the source image is that discrete adjacent points in the image can be clustered, and finally the effect of replacing the region with the center point is achieved. For example, a GPR image of 500 x 1000 contains 500000 ten thousand pixels, and after super-pixel segmentation with a step size of 20, only up to 25 x 50=1250 regions need to be processed, and the pixel values of the points in these regions are very similar, so that up to 1250 center points can be used to replace the 1250 regions, which obviously greatly reduces the calculation amount and greatly eliminates the influence of discrete noise points on image classification.

The idea of the SILC algorithm using KMEANS clustering is to initialize K centers for all data first, and then plan all data to the closest center point:

kmeans:minD(data ⁽ⁱ⁾ -μ _k )(1)

the above process is iterated successively to obtain corresponding super-pixel segmented images. Wherein, data ⁽ⁱ⁾ Represents the ith data, μ _k Representing the kth cluster center, D represents the distance between the two in a given algorithm, typically taking the euclidean distance. When SILC super pixel segmentation algorithm is applied to single-channel GPR image, data ⁽ⁱ⁾ Described as a three-dimensional vector: data ⁽ⁱ⁾ ＝[G _i -x _i -y _i ]The specific method comprises the following steps:

dividing the original image into a plurality of square blocks according to a certain step S by the GPR image, wherein the gray scale of the central point of each block and a vector consisting of two corresponding coordinates are used as the initialization center of the super pixel block. Thus a set of initialization centers is obtained: c= { C ₁ ,C ₂ ,...,C _{(m*n)/stride2} },C _i ＝[G _i ,x _i ,y _i ]Wherein G is _i Representing the pixel value at the center point, x _i ,y _i Representing coordinates.

Such a regular division method may cause the center point to be on an edge or noise point, and thus in order to prevent the point where the center is from being an unreasonable point such as an edge and noise, the center must be moved to a point where the gradient is minimum in a 3×3 neighborhood:

wherein,representing gradient values at pixel points (i, j) can be calculated using the sobel operator. The gradient is considered as a two-dimensional vector of gradients in both x, y directions:

defining gradient values:

wherein P (x, y) is the corresponding pixel value at the pixel point (x, y),is a matrix of 8-neighborhood pixel values at pixel points (x, y).

In the super-pixel algorithm, the distance between two points is defined by three-dimensional vectors [ G, x, y ]] ^T The calculation method is characterized by taking the value and the position of the pixel point into consideration:

where γ is the balance parameter, its size means the size of the influence of the position near or far on the cluster, and S is the step size.

For a general GPR image containing hyperbolic waves, the effect of 10 iterations is sufficient to meet the segmentation requirement. But after the process is finished, a plurality of isolated points or pixel blocks with small areas are generated, and the isolated points or the pixel blocks are required to be attached to the label with the largest communication.

Finally, obtaining a super-pixel picture, wherein the super-pixel picture comprises a plurality of super-pixel units:

U＝{U ₁ ,U ₂ ,...,U _n },N≤(m*n)/stride ² (6)

selecting a region of interest:

as can be seen from the GPR image, the region containing the waveform differs significantly from the other regions in that the region containing the waveform is color-significant in the y-direction (or one of the significant regions, as it may contain several waveforms), and in addition to that the region containing the waveform is in its circular neighborhood, it will also be significant. In general, hyperbolic waves in GPR images have two immediately following regions of black salience and white salience, but the distribution and position of black salience and white salience are different in different images. The hyperbolic wave regions in fig. 2 (a) and 2 (b) appear more distinct and more continuous as though they were white, but they are still distinct in that the hyperbolic wave regions in fig. 2 (a) appear to appear as gray-white-black-gray sequences from top to bottom on the y-axis with the background, while the gray-black-white-gray sequences are shown in fig. 2 (b); correspondingly, the image in fig. 2 (c) appears as a more distinct and more continuous black region and exhibits a gray-black-white-gray sequence. Therefore, it is not sufficient to consider only a significant region of white or a significant region of black in the GPR image, and thus the method in the present disclosure will be considered for both cases of black and white.

In the prior art, a color image is subjected to calculation of saliency maps with different scales in h/2, h/4 and h/8 neighborhood where pixel points are located, and summation is performed. The present disclosure therefore proposes a method for multi-scale neighborhood summation of GPR images, with respect to GPR images, with significance. Thus, hyperbolic wave division regions with different textures and colors in the GPR image can be extracted.

Therefore, in this section, a new and quick and effective method for multi-scale neighborhood calculation is proposed, which firstly proposes a way of combining a column neighborhood and a ring neighborhood for a GPR image subjected to super-pixel segmentation, calculates the significance of each super-pixel block, and then segments the image by setting thresholds with different positive and negative values and different sizes, wherein the different positive and negative values represent different black and white values. The method well solves the problem that the GPR image is subjected to binarization and then has a cavity due to the overlapping of noise and hyperbolic waves in the traditional threshold value calculation method, so that the image can be well segmented, and experiments prove the effectiveness of the algorithm. The area preliminarily selected by the algorithm has the characteristics of tidy edges, few redundant areas and continuous waveform without holes.

Among the previous algorithms, some, such as Xiren methou et al, first optimized the GPR image using a Matgpr or other tool, calculated the gradient in the y direction, and then calculated the gray value average of the set of pixels with all gradients other than 0 as the threshold to binarize the image. However, achieving binarization of an image by just depending on the pixel value level may cause image segmentation to be too rough; in a method proposed by Qingxu Dou et al, firstly, the GPR image is subjected to mean filtering denoising, then the mean value of the row set is subtracted from each row to eliminate ground reflection, an edge detection is used to binarize the image and finally the edge points are selected, specifically, for all edge points, after 10% of dark edge points are removed, the mean value of the gray values of the set of the remaining edge points is used as a threshold value, and finally all edge points meeting the above conditions are selected as the interested area of the image. The algorithm selects edge points, although not the entire image. However, it is worth noting that doing so may be effective for partitioning relatively pure GPR images. However, the actual GPR image may contain a lot of noise, so that the edge points are also numerous, which eventually leads to a possible excessive area of interest, which makes further operations difficult. Therefore, we introduce a view of local saliency computation, select the region of interest under the condition of multiple scales, and respectively take different thresholds for the black waveform and the white waveform. The problems are effectively avoided.

The preliminary selection of the region of interest is obtained by calculating the neighborhood distance of the pixel block formed by the super-pixel segmentation under the multi-scale. The specific method comprises the following steps: for a GPR image that has been superpixel segmented, for each unit U _i Defining a significance value:

where τ represents the neighborhood at different scales, λ (τ) represents the weight of the saliency map computed under neighborhood τ, a represents the number of superpixels within the neighborhood under neighborhood τ,gray value representing the center of the ith cell, is->The reason why the difference of pixel values is used for addition instead of taking absolute value is that the black and white waveforms can be separated, i.e. the calculated value ψ of the region where the black waveform is located _i >0, and the area ψ where the white waveform is located _i <0。

The reason for performing the distance calculation at different scales is to prevent the accidental of the image calculation at a single scale, resulting in the generation of an error.

However, the irregularities inherent to super-pixels, the selection of the neighborhood is very difficult, and is significant in terms of the columns in which the hyperbolic wave regions are located and the blocks in the adjacent regions, as mentioned above, so the present disclosure will use a combination of columns and circular neighbors.

Given a radius R, for each center ci, a set of centers { c1c2 … cj } within all the sets of centers within the circular neighborhood τ is selected:

D(c _i ,c _j )<R(8)

where D represents the distance between the two at the location in the image. And for the y-direction, all center points are selected to be in the interval [ y ] _min ,y _max ]Wherein y is _min Representing a super pixel block c _i Left boundary of y _max Representing a super pixel block c _i Right boundary of (c).

Converting into a binary image, and removing holes and noise points by using an opening and closing algorithm:

since the division of superpixels may result in the remaining region containing some holes or burrs, it is necessary to use an on-off algorithm to process the binary image. The principle of the open-close operation is to use the check image to perform an AND operation, wherein the open operation is to perform an AND operation on the image and the structural elements, then perform an OR operation, and the close operation is to perform an OR operation, then perform an AND operation to complete the open operation:

wherein,

extracting bone regions from the binarized image: since the fitting of the final hyperbolic wave is line-based, the binary image obtained in the above step is still an irregular area. To simplify the calculation, saliency maps obtained from different thresholds can be represented by linear bones using a skeleton extraction algorithm for all connected regions. This can simplify the calculation amount and can facilitate the fitting algorithm described later. A specific procedure is shown in fig. 3.

Setting and fitting of a matching set: in this section, a new method of judging and fitting hyperbolas is proposed. First, a definition of a matching set is given; then a method for pre-screening hyperbolic wave peaks in the skeleton image is provided; finally, a method for judging and fitting the hyperbolic wave by using the set is provided.

Definition of matching set:

for the hyperbolic wave fitting problem, hough transform is a common method. However, it is known that the corresponding equation for the hyperbolic wave appearing in the GPR image should be:

if the conventional hough transform method is used, the parameter x is obtained ₀ 、t ₀ It is required that the sequence of the sequences [ x, y, a, b ]] ^T The calculation is performed in the four-dimensional space where it is located. But if the vertex coordinates of the hyperbolic wave are known, the problem seems to be somewhat simplified. To match a known template, the vertex is first moved to the origin, as known from the definition of a hyperbolaWhen the vertex of (a) moves to the coordinate vertex, the expression is +.>Obviously, the hyperbolic equation with known vertex coordinates has only two unknown quantities-a and b.

The template is set in consideration of practical situations, and the speed can be determined within a certain range because the image of the GPR is obtained by manually pushing the GPR at a constant speed, however, the eccentricity e can be used for measuring the speed v well. If t is taken as a function of x, the hyperbolic expression moving to the origin is:

as shown in equation (12), the eccentricity e may reflect the "steep" condition of the hyperbola, and according to a mathematical definition, the derivative of each point in the curve, i.e. the velocity v, is set to e (1.05,2) in the matching set, with a step size of 0.05. The corresponding v e (0.81,4.42) represents the situation in practice. As shown in fig. 4, the eccentricity e is 1.05,1.5,2 when a is 10, 30, or 50. Essentially, the dataset is a corresponding set of pairs of parameters a and b in a hyperbolic equation, the hyperbola in the GPR image according to the real conditions in GHMS takes the value of a [10,50], the value of b is derived from the eccentricity, i.e

Thus, we have obtained 20×41=820 hyperbolic wave models in total.

FIG. 4 shows various results for eccentricity e of 1.05,1.5,2 when a has a value of 10, 30, 50. It can be seen that (1) when the eccentricity e, a is determined to be larger, the hyperbola is smoother (2) the eccentricity is better reflected in the advancing speed of the ground penetrating radar when the ground penetrating radar works, and the larger the eccentricity is, the slower the speed is.

It can be seen that when the value of a is determined, the eccentricity e can well reflect the shift velocity v, and when the eccentricity e is determined, the value of a can reflect the fullness of the hyperbolic wave waveform, so that the fitting set can well correspond to the real situation.

Pre-screening of hyperbolic wave vertexes of bone image seeds:

since each line segment with a downward opening will be easily determined after the bone image is generated, since a point segment match can be made at all points in the generated bone image, the successfully matched templates should be divided into 4 types, which will be shown in fig. 5: from a-d, 4 line segments with downward opening characteristics are shown, and by selecting all segments meeting the characteristics in the skeleton image, all hyperbolic wave vertices can be obtained, and the middle point of a flat area in 4 'up-flat-down' structures is taken as the pre-screened hyperbolic wave vertex.

Identification and fitting of hyperbolic waves: in practical situations, the waveforms in the generated GPR image will overlap in many ways, such as overlapping of hyperbolic waves and clutter, overlapping of clutter and clutter, etc., due to noise effects or due to too close a number of measured objects. However, these waveforms are likely to remain in the saliency calculation algorithm, so the advantage of extracting bone regions is very clear. Because each vertex is found, all branch conditions of the communication area corresponding to the vertex can be subjected to matching operation, then a threshold value P is set, and the area with the matching degree above the threshold value is taken as hyperbolic wave. Here, however, we need to consider three problems:

(1) Since the hyperbolic wave size varies among different GPR images, the threshold value should be related to the hyperbolic wave size, in other words, the matching ratio of the effective points should be not the number.

(2) A communication area may contain multiple hyperbolic waves, so several must be distinguished;

(3) Adjacent points may all have a high degree of matching and thus the fit cannot be repeated.

In the previous methods, it was quite difficult if not impossible to separate the regions.

A thresholding method is provided, which performs parameter matching on all branch path conditions corresponding to each vertex, specifically, for each vertex, one of the most matched branches in the fitting set is found as the waveform of the vertex, provided that the number of the points of the branch is enough, otherwise, the waveform cannot be represented well. And finding the corresponding template which is most suitable for the sub-path from all templates, and judging that the region is not a hyperbolic wave region if the matching degree of one template with the highest matching degree still cannot be suitable for the threshold value.

And carrying out a matching algorithm on the preliminarily screened connected areas with downward openings, wherein each point is subjected to the matching algorithm: firstly, the vertex of the hyperbolic wave is found, then, a set alpha of effective points is established, and the expansion is in pairs in the process of expanding to two wings of an effective cluster, so that the characteristics of axisymmetry of the hyperbolic wave are met.

As shown in fig. 6 (a) -6 (c), where gray represents a point that has been added to the set a, when it is about to be extended by point L toward N, the final set is (a-M) because point M on the same row in the right wing will not be able to continue to be extended. As shown in fig. 6 (b), when expanded to point K, there will be 3 expansion modes H, M, N but point H and point N obviously do not conform to the hyperbolic directional properties and are therefore excluded. I.e. depending on the nature of the hyperbolic wave, only a downward, leftward or downward-left expansion is possible. In fig. 6 (c), when the extension node is G, both points F and I will be selected, but if F is the extension node, it will only extend to point I, so in this case point F will be discarded, but a path G-I-K will be selected. But the problem does not seem as simple when point K is the extension node, since it can be extended to points M and N and they can each be extended to different paths. In this case, the M-branch and the N-branch are combined with the original effective points, respectively, and the path having the highest matching degree is determined.

Since branches may be generated during the expansion process and corresponding branches may also be generated, each vertex may have multiple effective paths { alpha }, respectively ₁ ,α ₁ ,…,α _n }. Therefore, matching degree operation is only needed to be carried out on paths with enough left and right tail parts and matching sets in all effective paths, and therefore:

where Δy refers to the difference in y direction for each column in the image when stacking the template vertices and the vertices of the downward opening line segments, and num refers to the number of points of the effective path.

In this section, experiments performed on data acquired under real conditions are presented, as well as some comparisons of the methods presented in this disclosure with other methods.

A. Experiment of real data set

The data tested are derived from very complex situations from cement grounds, sand beach, even grasslands, stone roads, etc. acquired under real conditions using GPR model GSSI-2000, from 80 GPR images acquired therefrom, which are all of a variety, with very weak distinction between black and white significantly different and with a very weak contrast between hyperbolic waves and background, and even with a large amount of noise and cross. By manual screening we obtained a total of 204 of the effective hyperbolic waveforms. Of these GPRs, a small part of the image is clear and the waveform is good, but most contain a large amount of hyperbolic wave crossing situations and noise, which are used as samples in the experiments.

According to the method, super pixels with the step length of 20 are adopted to divide the image, 12 saliency maps are obtained, the thresholds of the super pixels are respectively + -0.5, + -0.65, + -1.3, + -1.5, + -1.8 and+ -2.25 times of the average value of the saliency values corresponding to all super pixel blocks, and the initial screening is carried out until the tail length reaches 30, so that the template matching is carried out. A detailed result obtained by our method is shown in fig. 7.

In order to better describe the identification accuracy of the hyperbola by the method, euclidean distance in the y direction is used as an evaluation index. Previously, we have manually annotated hyperbolic waves in images in GPR and have obtained accurate equations. We define the similarity between the resulting hyperbola equation p1 and the standard hyperbola equation p2 as

Where distance d is defined as the average of the ratio of the distance between two points in the y-direction and the height of the hyperbola. See fig. 7 for a specific description. As shown in FIG. 7, if the fitted curve is a solid line in the graph and the artificially labeled curve is a broken line in the graph, it is assumed that the height of the effective hyperbolic wave region is defined as h and at the same x _m In the following, if the coordinates in the y direction are p, respectively ₁ :(x _m ,y _m ) And p _2: (x _m ,y _n ) I.e. the difference in height distance is Δh=y _m -y _n Then define:

as the OSCA algorithm and the C3 algorithm can obtain the hyperbolic wave-like clusters through graphic calculation, comparison is carried out among the three algorithms. One experimental result on an actual dataset using our method is shown in fig. 7, where 4 valid clusters are given. However, the use of OSCA and C3 algorithms yields many more effective clusters than this result. To more intuitively and effectively demonstrate the superiority of our method, we performed experiments with 80 images equally divided into 4 groups. Under the condition that each picture is ensured to contain at least one hyperbolic wave, the average fitting rate is calculated.

In all experiments, there were very few or no missed hyperbolic wave targets. However, although this algorithm contains few redundant targets, it is known that in practical situations this helps to reduce the risk and avoid accidents. It can thus be seen that the method we propose is very safe and reliable. B. Comparison of our approach with other approaches

To better verify the validity of the model, we have also performed some comparative experiments. We compared our method with the C3 column connected clustering algorithm and OSCA algorithm. The method, the C3 algorithm and the OSCA algorithm can perform unsupervised processing on the ground penetrating radar image to obtain a hyperbola. It is known that OSCA algorithm performs matching of the down-opening column segments, whereas C3 algorithm does not. In addition, our method also makes a judgment by matching templates, but OSCA and C3 algorithms do not make a similar judgment. We evaluated the above three methods on a real dataset after corresponding pre-processing. We count the number of active clusters and the CPU time, respectively, to measure the performance of the method. The results of some real datasets are shown in fig. 8 (a), 8 (b) and 8 (c).

It can be seen that OSCA processing time is better than C3 algorithm because C3 requires horizontal scanning of each connected region in the image, and both processing times are longer than our method because the number of regions to be processed is greatly reduced by our method after super-pixel segmentation, and the final screening and fitting processes are performed simultaneously, and only several tens of points per vertex are compared by extraction of bone regions, thus being very efficient.

Another important task is to compare the error rate of fitting between various methods, we use our method and Hough transform method and RADF method to compare, i.e. after passing our template matching step, the areas judged to be yes are fitted by Hough transform and RADF method, respectively. The correct recognition rate and the CPU time are given in graphs I and II.

In the case of preprocessing the point clusters, hyperbolic equations are obtained, and thus corresponding parameters of the object, such as depth and radius, are obtained. Wherein the Hough transform takes the longest time. Because the method only needs to match dozens of pixel points with the template after finding the vertex of the hyperbola, a good time effect can be obtained. While the other two methods have to process the region, require calculation of a large number of points and are not sufficiently noise-resistant, so the calculation time is more complicated and the recognition rate is low. Experiments show that the Hough transformation needs to be calculated in four dimensions, so that the time complexity is too high, and compared with an RADF method, the method is basically the same in recognition rate but higher than the RADF method and lower in time complexity, so that the hyperbolic wave can be better recognized and fitted in real time.

The disclosure provides a novel hyperbolic wave identification algorithm of a ground penetrating radar image, which comprises the steps of extracting significant regions of a GPR image, separating the significant regions under different thresholds after extraction, extracting connected skeleton regions, extracting vertexes, and then performing matching processing by using a matching set. The validity of the method is verified in practice.

In the proposed method, the region of interest is calculated by the neighborhood calculation under multiple scales and the final matching algorithm is the two core parts. The geometric characteristics of hyperbolic waves are considered by the neighborhood calculation under the early multi-scale and the final matching algorithm, so that the algorithm has stronger robustness.

The second embodiment also provides an underground target detection system based on the ground penetrating radar hyperbolic wave fitting;

In a third embodiment, the present embodiment further provides an electronic device including a memory and a processor, and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the steps of the method of the first embodiment.

In a fourth embodiment, the present embodiment further provides a computer readable storage medium storing computer instructions that, when executed by a processor, perform the steps of the method of the first embodiment.

The above description is only of the preferred embodiments of the present application and is not intended to limit the present application, but various modifications and variations can be made to the present application by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. The underground target detection method based on the ground penetrating radar hyperbolic wave fitting is characterized by comprising the following steps of:

acquiring a ground penetrating radar GPR image;

the GPR image is subjected to super-pixel segmentation to obtain a super-pixel segmented image comprising a plurality of areas; the method comprises the following specific steps: performing super-pixel segmentation on the GPR image by using a SILC super-pixel segmentation algorithm to obtain a super-pixel segmentation image containing a plurality of areas; clustering discrete adjacent points in the segmented super-pixel segmented image, and finally realizing the effect of replacing the area by the center point;

extracting the region of interest from the super-pixel segmented image to extract a plurality of regions of interest, wherein the method comprises the following specific steps:

wherein,representing neighborhoods at different scales, +.>Expressed in the neighborhood->Weights of the computed saliency map

The weight of the steel plate is increased,expressed in the neighborhood->The number of superpixels in the next neighborhood, +.>Represents->Gray value of the individual cell center, is->Represents->Gray value of the individual cell center, is->Indicate->Significant values for the individual superpixel regions;

s32: setting a plurality of different thresholds;

s33: obtaining a plurality of saliency maps according to a plurality of different thresholds;

performing image binarization processing on each region of interest to obtain a binarized image, and extracting a bone region from the binarized image, wherein the method comprises the following specific steps of: using a skeleton extraction algorithm to represent all connected areas in the saliency map by linear skeletons and acquiring skeleton images;

the method for extracting the hyperbolic wave from the bone region comprises the following specific steps:

defining a matching set, pre-screening hyperbolic wave vertexes in a skeleton image, judging and fitting the hyperbolic wave by using the matching set, and extracting the hyperbolic wave, wherein the specific steps comprise: for each vertex, finding the waveform which is most matched with the matching set in all branches as the vertex;

the pre-screening of hyperbolic wave vertexes in the skeleton image comprises the following specific steps: screening a line segment with a downward opening, and taking the midpoint of a flat area in an ascending-flat-descending structure as a pre-screened hyperbolic wave vertex from the line segment with the downward opening;

2. An underground target detection system based on ground penetrating radar hyperbolic wave fitting for realizing the method as set forth in claim 1, comprising:

3. An electronic device comprising a memory and a processor and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the steps of the method of claim 1.

4. A computer readable storage medium storing computer instructions which, when executed by a processor, perform the steps of the method of claim 1.