A hyperbolic linearization identification method based on lower opening scanning

Abstract

The application discloses a hyperbolic linearization identification method based on a lower opening scanning. The application converts a hyperbola into an equidistant model, converts whether a target point cluster is a hyperbola into whether the point cluster presents an equidistant characteristic, reduces a to-be-estimated parameter of the hyperbola to one, and meanwhile, retains all information of the original hyperbola shape, and has better accuracy compared with a traditional least square method.

Description

Technical Field

[0001] This invention relates to the field of radar image processing technology, and more specifically to a hyperbola linearization recognition method based on lower opening scanning. Background Technology

[0002] Ground penetrating radar (GPR) technology is a geophysical detection method used to detect underground targets. Traditional crack monitoring methods include visual inspection, hammering, and coring, which are either destructive or inefficient and costly. Compared to traditional methods, non-destructive testing (NDT) can reduce maintenance costs and risks. Besides GPR, NDT methods also include resistivity methods and low-frequency electromagnetic induction methods. Compared to other NDT methods, GPR offers advantages such as fast detection speed, continuous detection process, high resolution, convenient and flexible operation, and low cost, and is considered the most effective and reliable NDT method in terms of continuity, safety, and interference resistance.

[0003] Urban road pipes have a different dielectric constant than the surrounding medium. Because pipes are often buried horizontally within the effective depth range of the ground-penetrating radar (GPR) antenna, and the GPR detection direction should be perpendicular to the pipe's direction (in practical applications, the direction of the buried pipe can be initially determined using existing pipe maps, and then the specific direction can be determined through multiple detections in different directions), a hyperbola will appear on the GPR B-scan image when detecting the pipe. The location and size of the relevant pipe within the medium can be obtained by fitting the generated hyperbola and examining its parameters.

[0004] In practical applications, ground-penetrating radar (GPR) images can be noisy due to system noise, the heterogeneity of the medium, and the interaction of mutual waves. Therefore, automatically processing GPR images using computers to extract hyperbolas has become a research topic. Many strategies have emerged in this field to address this problem, including but not limited to methods based on Hough transform, least squares methods, machine learning-based methods, and combinations thereof.

[0005] Currently, machine learning-based methods are the mainstream approach; however, their recognition performance is highly dependent on the quality of the training set, and finding a large number of suitable, high-quality hyperbolic image training sets is a challenge. Without machine learning, existing methods such as least squares still have room for improvement in recognition performance under conditions of low signal-to-noise ratio and high clutter. Summary of the Invention

[0006] In view of this, the present invention provides a hyperbola linearization recognition method based on lower opening scanning, which can solve the problem that the existing methods have large errors when the hyperbola features are not obvious or the hyperbolas overlap, thus affecting the recognition accuracy, and improves the accuracy and robustness of the recognition method.

[0007] The hyperbola linearization recognition method based on lower opening scanning of the present invention includes:

[0008] Step 1: After removing the direct wave from the ground penetrating radar B scan image, perform binary segmentation to obtain a binary image;

[0009] Step 2: Perform a lower opening scan on the binary image to obtain point clusters for each lower opening region;

[0010] Step 3: Crop the image containing the point clusters of each lower opening region, and scan column by column to extract the point segment information of the point clusters of each lower opening region. The point segment information includes the number of point segments in each column, the ordinates of the start and end points of the point segments, and the length of the point segments. Take the midpoint of each point segment and perform HPL processing to obtain a line graph H. The HPL processing specifically involves:

[0011] S341, Construct a new orbital coordinate system O ′ x ′ y The new orbital coordinate system is a rectangular coordinate system, with the origin being a set reference point;

[0012] S342, obtain the midpoint of the point segment extracted by column in S33 and use... Transform the midpoint of the point segment to the new orbital coordinate system. O ′ x ′ y Below, a broken line curve is obtained; where, ( x , y () represents the coordinates of the midpoint of the point segment in the original image; x ′, y ′) represents the coordinates of the midpoint of the point segment in the new orbital coordinate system; where, , y D The maximum value of the midpoint of the point segment extracted from S33; r This refers to the scanning step distance of the ground-penetrating radar antenna. v The speed at which electromagnetic waves propagate in the background medium. t For time;

[0013] S343, based on the point segment length obtained in S33, expands the polyline curve of S342 into a polyline graph H;

[0014] Step 4, construct a series of correlation matrices Si The parameters of the correlation matrix S include: slider width, left step count, right step count, and single step distance; where the slider width is 1.5 to 3 times the width of the top of the line chart in step 3; the left and right step counts are respectively taken as the distance column numbers from the top midpoint to the left and right image boundaries in the line chart H in step 3; the single step distance iterates between 1 and the maximum single step distance, i.e. i =1, 2, …, maximum single step distance; the maximum single step distance is the distance in rows from the top midpoint of the line graph in step 3 to the bottom of the image;

[0015] Calculate the correlation values ​​between the series correlation matrices and the line graph from step 3, and obtain the maximum correlation value;

[0016] Step 5: If the maximum correlation value obtained in step 4 is greater than or equal to the judgment threshold, then the point cluster corresponding to the line chart is determined to be a hyperbola point cluster.

[0017] A better approach is to use an adaptive threshold to achieve binary segmentation, specifically including:

[0018] S11, perform median filtering on the image after removing the direct wave to obtain image C1;

[0019] S12, calculate the global average value of image C1; replace the values ​​of all pixels in the image that are less than the average value with the average value to obtain image C2;

[0020] S13, using an adaptive thresholding algorithm based on longitudinal gradient information, threshold segmentation is performed on image C2 to obtain binary image C3.

[0021] In a more optimal form, S13, the gradient region set is defined as:

[0022]

[0023] in,( i , j ) represents the pixel coordinates; g is the set minimum gradient threshold;

[0024] Define adaptive threshold T for:

[0025]

[0026] Using adaptive thresholds T Image C2 is segmented to obtain a binarized image C3:

[0027] .

[0028] A better approach is to remove direct waves by means cancellation, that is, for each pixel, subtract the average value of the row from the value of that pixel.

[0029] Preferably, in step 1, a closing-then-opening operation is performed on the binary image.

[0030] Preferred methods for obtaining the point clusters in the lower opening region include:

[0031] S21, For a binary image, scan the point segments with downward opening features row by row to obtain the downward opening region and establish point clusters;

[0032] S22, perform a top-up search and a left and right-tail-down search on the lower opening region. During the upward search, if the point segments in the previous row have vertical overlap with the point segments in the current row, cut off the pixels in the previous row that extend beyond the horizontal area of ​​the current row point segments, and assign the cluster class number of the current row point segments to the pixels in the previous row point segments that do not extend beyond the horizontal area of ​​the current row point segments. Continue searching upward until there are no vertically overlapping point segments above. During the downward search, if the point segments in the next row have vertical overlap with the point segments in the current row, cut off the pixels in the next row that extend beyond the horizontal area of ​​the current row point segments, and assign the cluster class number of the current row point segments to the pixels in the next row point segments that do not extend beyond the horizontal area of ​​the current row point segments. Continue searching downward until there are no vertically overlapping point segments below, and obtain the final point cluster of the current lower opening region.

[0033] Preferably, in step 3, the image containing the point clusters of each lower opening region is cropped as follows:

[0034] S31, the binary image is divided into multiple sub-images, each sub-image containing only one lower opening region point cluster; the lower opening region point clusters in each sub-image are cropped to obtain the cropped lower opening region image;

[0035] S32, Perform vertical bilinear interpolation on the cropped lower opening region image;

[0036] S33, perform column scanning on the interpolated image and extract point segment information by column.

[0037] Preferably, in step S31, the image cropping specifically involves: for each sub-image, marking four midpoints of the point cluster: the top, left tail, right tail, and the midpoint of the lower opening; and cropping the point cluster region based on the four midpoints to obtain the cropped lower opening region image.

[0038] Preferably, in step 4, when calculating the correlation value, the correlation between the correlation matrix S and the line graph H is calculated first, and then the sum of the redundancy values ​​of all elements of the redundancy matrix is ​​subtracted to obtain the correlation value for that single step distance. The redundancy matrix is: the region where the difference between the line graph H and the correlation matrix S is 1, and the redundancy value of each point in the region is assigned according to the distance between each point in the region and the slider, which is a matrix with the region being zero.

[0039] Preferably, in step 5, the determination threshold is: C 3 × slider width × (number of left steps + number of right steps + 1), where C3 is a constant.

[0040] Beneficial effects:

[0041] (1) This invention transforms the hyperbola into an equally spaced model, changing the question of whether the target point cluster is a hyperbola into the question of whether the point cluster exhibits equally spaced characteristics, thus making the parameters to be estimated of the hyperbola more reasonable. a , b Reduced to one d t / r At the same time, it retains all the information about the original hyperbola shape, resulting in better accuracy compared to ordinary fitting methods. Ordinary fitting methods (such as least squares) may produce incorrect results due to the special characteristics of the hyperbola shape, but this method can effectively handle such situations.

[0042] (2) The present invention achieves adaptive threshold binary segmentation based on improved longitudinal gradient information, which is suitable for situations where the background noise area is large and the background noise cannot be completely filtered, and can achieve better segmentation effect and adaptive threshold segmentation.

[0043] (3) In order to achieve better recognition effect during the lower opening scanning process, the present invention makes appropriate modifications to the lower opening scanning process, no longer recognizes the upper opening and the "x" point segment, and sets the point segment range of each position of the hyperbola marked as the lower opening, and reasonably deletes the pixel area that exceeds the range, which is closer to the characteristics of the actual hyperbola and enhances the robustness of the algorithm.

[0044] (4) By introducing a redundancy matrix, this invention takes into account the correlation calculation error caused by the pixel of the line graph exceeding the slider range in the line graph, thereby reducing the recognition error. Attached Figure Description

[0045] Figure 1 The results are the mean cancellation results; (a) before processing; (b) after processing.

[0046] Figure 2(a) Schematic diagram of threshold segmentation method based on traditional longitudinal gradient information; (b) Actual effect of traditional method; (c) Adaptive threshold segmentation effect of the present invention.

[0047] Figure 3 (a) shows the result of the opening and closing operations; (b) shows the result of the closing operation.

[0048] Figure 4 The diagram shows the lower opening segment of OSCA; (a) is the lower opening; (b) is the upper opening.

[0049] Figure 5 This refers to the case of overlapping points and segments.

[0050] Figure 6 The impact of the original discrimination condition is shown in (a), which is the case where the tip of the point cluster has burrs; (b) is the result of the traditional discrimination on (a).

[0051] Figure 7 The following diagram illustrates the cutting effect of the scanning process under the present invention: (a) shows the processing of the point cluster with burrs at the top; (b) shows the processing of the point cluster with burrs at the tail.

[0052] Figure 8 This is a diagram illustrating pixel clustering.

[0053] Figure 9 This is the result after OSCA processing.

[0054] Figure 10 This is a schematic diagram of a ground-penetrating radar scenario.

[0055] Figure 11 This is a schematic diagram of HPL transformation.

[0056] Figure 12 The effect of variable substitution on the width of the point segment.

[0057] Figure 13 This is a schematic diagram of point segment deformation.

[0058] Figure 14 This is the reason why interpolation is needed.

[0059] Figure 15 Four points are used for calibration.

[0060] Figure 16 This is the converted image.

[0061] Figure 17 These are the relevant values ​​under different timelines.

[0062] Figure 18The results are as follows: (a) to (g) are the recognition results of 7 LTE radar images identified using the method of the present invention; (h) to (n) are the recognition results of 7 LTE radar images identified using the PFJ method.

[0063] Figure 19 This is a flowchart of the present invention. Detailed Implementation

[0064] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0065] This invention provides a hyperbolic linearization recognition method based on bottom-opening scanning. First, the B-scan image is converted into a binary image through preprocessing methods (mean cancellation, adaptive thresholding, opening and closing operations, etc.). Then, the bottom-opening scanning algorithm (OSCA) is used to identify point clusters with bottom-opening features, and each pixel is clustered. During the bottom-opening scanning process, cases where two point segments in the same column of the same cluster should be eliminated as much as possible for subsequent processing. Then, the hyperbolic point segment linearization (HPL) method is used to convert the point clusters into equally spaced linear models. The ideal equally spaced image is represented by a computer. By comparing and evaluating the equally spaced features of the converted image, the similarity between the point cluster and the ideal hyperbolic point cluster is determined.

[0066] The specific steps are as follows:

[0067] Step 1: Preprocess the ground-penetrating radar B-scan images

[0068] The preprocessing process consists of three steps: First, mean cancellation is performed on the image to remove direct waves; second, adaptive thresholding is performed on the image; and third, opening and closing operations are performed on the image to remove freckles and connect breakpoints.

[0069] S1, Remove direct waves:

[0070] The direct wave is removed using the mean cancellation method, specifically by: for each pixel ( i , j The expression is as follows: Subtract the average value of the row from the value of the given pixel.

[0071]

[0072] Among them, B M×N Represents the original B-scan image, C M×N This represents the image after the mean is canceled out.

[0073] The comparison effect of mean cancellation treatment is as follows Figure 1 As shown, the direct wave portion is effectively removed.

[0074] S2, Adaptive Threshold Binary Segmentation Algorithm:

[0075] An adaptive thresholding algorithm based on vertical gradient information can be used. This involves extracting regions with large vertical gradients from the image, calculating the average value of these regions as the threshold T for binary segmentation. Figure 2 As shown in (a). However, when the background noise area is large, since the background noise value is often between the positive and negative peaks, if only the area with the larger gradient in the image is averaged, then the background noise cannot be completely filtered out, such as... Figure 2 As shown in (b), to achieve better segmentation results, the negative peaks of the image need to be removed. Therefore, this embodiment makes the following improvements:

[0076] The adaptive threshold algorithm in this embodiment consists of three parts:

[0077] ① Use median filtering to remove image noise and reduce gradient values ​​in flat areas. This step is to avoid image noise affecting the overall threshold. Image C1 is obtained after this step.

[0078] ② Calculate the global average value *m* of image C1. For each pixel, if its value is less than the average, set its value to *m*. This step is to prevent negative peak regions from lowering the overall threshold. *Mean{}* represents taking the matrix average, which can be expressed mathematically as:

[0079]

[0080]

[0081] ③ Binary segmentation is performed using an adaptive threshold algorithm based on vertical gradient information; specifically, the gradient region set is defined as:

[0082]

[0083] Where g is the minimum gradient threshold.

[0084] Define the adaptive threshold T:

[0085]

[0086] After segmentation with a threshold T, a binarized image C3 is obtained:

[0087]

[0088] The effect after processing with the adaptive threshold method is as follows: Figure 2 As shown in (c), after thresholding, the original grayscale image is transformed into a binary image.

[0089] S3, Open-Close Operation:

[0090] The opening-closing operation consists of two steps—the opening operation and the closing operation—both of which are morphological operations representing expansion and erosion. Assuming a structuring element dk, the opening operation is defined as:

[0091]

[0092] Similarly, the closing operation is defined as:

[0093]

[0094] This represents the erosion of image edges by the structuring element; the magnitude of the erosion depends on the choice of the structuring element radius. Similarly, This represents the expansion of the image edges by the structuring element. This scheme uses a closing-then-opening operation. The radius of the structuring element for the closing operation is set to 3, and the radius of the structuring element for the opening operation is set to 2. The effect is as follows... Figure 3 As shown. Figure 3 Figure (a) shows the processing effect of the closing operation. Figure 3 (b) The figure shows the processing effect of the opening operation.

[0095] Step 2: Perform improved bottom-aperture scan processing (OSCA) on the preprocessed image.

[0096] Traditional bottom-opening scan processing (OSCA) identifies bottom-opening point clusters by scanning point segments with bottom-opening characteristics row by row. The concept of a bottom opening is defined as follows: below a point segment (segment 1) in the i-th row, there are two point segments, segment 2 and segment 3, that overlap vertically with it. The starting point of segment 1 lies between the starting and ending points of segment 2, and the ending point of segment 1 lies between the starting and ending points of segment 3. The three point segments are interconnected, as shown below. Figure 4 As shown in (a). In Figure 4 In (a), segment 2 and segment 3 have a gap, and the length of each segment is greater than 2. Thus, segment 1, segment 2, and segment 3 form a lower opening region. Similarly, if segment 2 and segment 3 are located above segment 1 in an adjacent row, and the three segments overlap vertically with a gap between segment 2 and segment 3, then these three segments form an upper opening region, as shown below. Figure 4As shown in (b). After the computer identifies the lower opening region, it assigns a cluster class number to this region, establishing a new point cluster. The portion above the opening is called the top of the hyperbola, and the portion below is called the tail of the hyperbola. After the program finds the lower opening region, it performs one upward search and two downward searches. During the upward search, if the point segment in the previous row does not exceed the start and end range of the adjacent point segment in the next row, all pixels of that point segment are assigned the cluster class number of the lower point segment, and the upward scan continues until there are no overlapping point segments above or the overlapping point segments exceed the horizontal region corresponding to the lower point segment. The two downward scans search along the left and right tail boundaries of point clusters for point segments 2 and 3, respectively. That is, if a point segment in the next row overlaps with an adjacent point segment in the previous row, all pixels of that next row point segment are assigned the cluster class number of the upper point segment, and the downward scan continues until there are no overlapping point segments below. Furthermore, when adding a point segment to an existing cluster, sometimes two point clusters overlap, such as... Figure 5 As shown. When encountering situations where two point segments overlap, the traditional method employs a strategy of deleting the overlapping portion.

[0097] To ensure better accuracy and seamless integration with subsequent steps, this invention improves upon the traditional lower-aperture scanning process as follows:

[0098] Firstly, areas where the points exceed a reasonable range should be cut off.

[0099] In actual images, the top region of a hyperbola point cluster does not entirely follow the condition that the upper point segment is within the start and end interval of the adjacent lower point segment, such as... Figure 6 As shown in (a), tiny burrs may appear at the top of the hyperbola point cluster in the actual image. Failure to meet this condition will cause all the areas above to be cut off from the origin point cluster, thus severely affecting the shape of the hyperbola. Figure 6 As shown in (b), the present invention makes the following modifications:

[0100] Assuming that point segment X is identified as part of a lower-opening cluster, the interval of the point segment is... And there exists a point segment Y above that overlaps with X in the vertical direction, the interval of which is... Then take the interval The pixels within are placed into the point cluster containing point segment X, where:

[0101]

[0102]

[0103] `max()` and `min()` represent taking the maximum and minimum values, respectively. The same process is performed during the search of the left and right tails. During the downward scan, the endpoints of the adjacent segments below the left tail and the starting points of the adjacent segments below the right tail are cut. Assume the segments of the left and right tails are L and R, respectively, and they overlap with the segments below them LL and RR. The horizontal intervals of L and R are respectively... , ;LL and RR are , Then the interval , The pixels within are placed into the point cluster containing L and R, where

[0104]

[0105]

[0106] Cutting effect as Figure 7 As shown. The impact of cutting is far from enough to affect the program's judgment, but it can greatly improve the clustering effect when there are too many burrs in the hyperbolic point cluster.

[0107] Secondly, under the one-based clustering condition, pixels in overlapping regions are allowed to be added to multiple point clusters, such as... Figure 8 As shown in (a).

[0108] The intersecting regions in the diagram are considered common areas of multiple point clusters. During clustering, these common areas become components of all the point clusters that make up that region. The pixels in these common areas contain the class numbers of all the point clusters that make up that region, and these class numbers appear in different point clusters. The clustering results are as follows: Figure 8 As shown in (b) and (c).

[0109] The image after the lower aperture scanning process is as follows: Figure 9 As shown, non-downward-opening point clusters have been filtered out.

[0110] Step 3: Convert the hyperbolic point cluster into an equally spaced lattice.

[0111] In ground-penetrating radar, assuming the target's x-coordinate is 0, such as... Figure 10 As shown, the relationship between the distance from the antenna to the target and the antenna's horizontal coordinate can be expressed as:

[0112]

[0113] Let r represent the step distance of the antenna in each step, x represent the number of orbits between the antenna and the target, d represent the target depth, and s represent the straight-line distance from the antenna to the target. Then the formula can be simplified to:

[0114]

[0115] Sampling is performed on the time axis, and it is known that... x The target appears in the image at a certain time. y ,Depend on:

[0116]

[0117] Substituting, we get:

[0118]

[0119] parameter y D This indicates the time when the echo appears in the image when the antenna is directly above the target, i.e., the ordinate of the hyperbola's vertex, such as... Figure 11 As shown in the figure. v represents the propagation speed of electromagnetic waves in the background medium, parameter y D As can be seen from the formula, it is only necessary to estimate the ratio of the product of sampling time and velocity to the antenna step, v*dt / r This allows for the unique determination of the hyperbola equation corresponding to a specific hyperbola region in the image. By directly performing correlation operations between the hyperbola point clusters processed by OSCA in the image and the ideal hyperbola, the similarity between the hyperbola point clusters and the ideal hyperbola can be obtained, thus enabling the identification of the hyperbola. This method closely resembles human visual perception.

[0120] However, in practical image processing, since the radar step distance during sampling can only be obtained from specific radar sampling files, we need a widely applicable method to determine hyperbolas. This invention will directly determine hyperbola point clusters from the perspective of image processing, based on the characteristics of hyperbolas.

[0121] like Figure 11 As shown, the image contains a reference point and a track for sliding point segments. If the point segments of each column of the hyperbola are placed on the track, and the ordinate of each point segment is converted into the straight-line distance from the reference point, a set of point segments is obtained in this way. Since the hyperbola is an image obtained by a radar antenna under equally spaced sampling, rearranging the point segments of the hyperbola in this way should result in all point segments having equally spaced distances on the track.

[0122] This can also be explained mathematically:

[0123] If let Then we have:

[0124]

[0125] Simplifying, we get:

[0126]

[0127] This shows that y 'and x There is a linear relationship. y This is equivalent to the coordinates of a point segment on the "slide rail" relative to the reference point, such as... Figure 11 As shown, y '- x The absolute value of the slope of the image is parameter 2. r / vdt, therefore the original problem is transformed into y '- x Whether a linear relationship is satisfied. This algorithm avoids the problem of parameter estimation in hyperbolas. Therefore, it will... y - x Image converted y '- x Images are judged using design evaluation criteria. y 'and x By determining whether a linear relationship is satisfied, we can determine whether a point cluster satisfies the characteristics of a hyperbolic point cluster, thus avoiding errors caused by parameter estimation in the hyperbolic equation.

[0128] In actual images, point segments have a certain width along their vertical axis. When determining the coordinate range of a point segment on the "sliding rail," the impact of this transformation on the segment's width needs to be considered. Figure 12 As shown, according to the properties of triangles, the length of the transformed point segment will always be greater than the length of the original point segment. y 2- y 1) Therefore, it can be concluded that the length of each point segment will be distorted to different degrees during the transformation process. That is, point segments that were originally of the same length may become different in length after transformation due to different positions.

[0129] Therefore, in the actual process of constructing equally spaced point segments, it is necessary to first extract a point in the point segment and perform coordinate transformation, and then reconstruct the point segment based on the position of that point in the point segment and the length of the point segment, such as... Figure 13 As shown in (b), instead of transforming all points in the point segment to the new coordinate axis in the same way as above, which would lead to image distortion, this invention extracts the midpoint of the hyperbola and processes it, arranging the point segment positions according to the rearranged midpoints, thus transforming the hyperbola into an equally spaced polyline.

[0130] Suppose the coordinates of the midpoint of a certain point segment are y m If the length of the point segment is L, then the coordinate range of the point segment on the "slide rail" is: ,in `cell()` represents rounding up. The process of equalizing the spacing of point segments in each column of the hyperbola using the above method is called the Hyperbola Point Segment Linearization (HPL) method in this invention.

[0131] Because in digital images, the displacement of a point segment can only take integer values, while the point segment interval is 2... r / dt can be a decimal, in some cases (such as...) Figure 14 As shown, this will result in a low correlation between the image and the equally spaced sliders, so the original image needs to be interpolated vertically.

[0132] Figure 14 (a) shows the position of the slider relative to the actual point segment when the slider step is 1. Figure 14 (b) shows the position of the slider relative to the actual point segment when the slider step is 2. The white area represents the position of the actual point segment, and the black area represents the position of the slider. It can be seen that the interval of the actual point segment should be between 1 and 2, while the slider interval can only be an integer. Performing related operations under this condition will result in a lower correlation value and may lead to misjudgment.

[0133] Therefore, the specific process for step 3 is as follows:

[0134] The image is segmented according to the cluster number, ensuring that each image belongs to only one category. A scan is then performed to identify four points: the top, left tail, right tail, and the midpoint of the bottom opening of each cluster. These four points are used to determine the position of the cropped image within the original image. Figure 15 As shown;

[0135] Assume the coordinates of the four points are as follows: ; ; ; .

[0136] The image is cut based on these four points to obtain the cropped hyperbolic region; the dimensions of the cropped image are: ;

[0137] Perform vertical bilinear interpolation on the cropped image;

[0138] Perform a column scan to obtain column point segment information distributed by column, including the number of points in each column, the length of each point segment, and the ordinates of the start and end points of each point segment.

[0139] Take the midpoint of each segment and perform HPL processing;

[0140] By following the steps above, a line graph derived from a hyperbola can be obtained, as shown in the image below. Figure 16 As shown, assuming the image is H Mh×Nh .

[0141] Step four: Evaluate the equidistant spacing of the converted image to determine if the original image is a hyperbola.

[0142] By stitching together the positions of the sliders in each column, we obtain an inverted "V" shaped image, which we call the correlation matrix. The correlation matrix has four parameters: slider width, number of left steps, number of right steps, and distance per step. The slider width depends on the width of the segment containing the vertex. In real-world images, many hyperbolas that are visually recognizable are not perfectly symmetrical. Therefore, when designing the number of slider steps, the number of left and right steps should be designed separately. The number of left steps depends on the lateral distance between the vertex and the left tail; the number of right steps depends on the lateral distance between the vertex and the right tail. Assume the vertex is located at the [number missing]th [position missing] in the column. k On a sequence of point segments, the slider width is z The number of steps to the left and right are respectively w l , w r :

[0143]

[0144]

[0145]

[0146] in Indicates the image number 1 i The first in the column from top to bottom j A sequence of point segments, C 1 is a constant

[0147] The single step distance traverses within the range after clipping interpolation, and the slider step corresponding to the maximum relevance value is taken as the selected step, expressed as:

[0148]

[0149]

[0150] argmax(Cor( r )) indicates Cor( r When taking the maximum value r The value of Cor( r ) indicates that the distance of a single step is r When the number of pixels is 1, the correlation matrix and H The correlation values ​​of the matrix will be explained in detail in the next step.

[0151] For example, the following is a matrix under different timing conditions. H The degree of correlation with the correlation matrix corresponds to Figure 17 :

[0152] Relevant values ​​under different paces:

[0153]

[0154] When the slider's single step distance is 6 pixels, the correlation matrix and the matrix H The correlation is the highest, therefore the slider step in this example is... r Take 6.

[0155] when H When pixels in the matrix exceed the slider's range, their correlation with the correlation matrix should decrease. Therefore, we introduce a redundancy matrix. The redundancy matrix represents the original matrix. H The distribution of pixels outside the relevant matrix region, i.e., the matrix H The pixel region whose value is one is obtained by subtracting the correlation matrix. For all pixels with a value of 1 in the redundancy matrix, the redundancy level of that point needs to be evaluated, which is referred to as the redundancy value in this invention. The calculation of the redundancy value should follow the following principles:

[0156] H

[0157] Based on the above conditions, the method for calculating redundancy is defined as follows:

[0158]

[0159]

[0160]

[0161] λ For buffer, for slider width C 2 times, C 2 is a constant; l This represents the shortest distance from the pixel to the slider (top or bottom boundary) of its row, where z is the slider width. l The impact on the overall correlation value grows geometrically.

[0162] Relevance matrix S With matrix H Perform correlation calculations, then subtract the sum of redundancy values ​​for all elements of the redundancy matrix. This yields the correlation value for that step, expressed as:

[0163]

[0164]

[0165] Set the judgment threshold as follows: C 3 × slider width × (number of left steps + number of right steps + 1), where C3 is a constant, expressed as:

[0166]

[0167] if If the condition is met, then the point cluster is determined to be a hyperbolic point cluster.

[0168] Example

[0169] To verify the accuracy of this invention in identifying hyperbolic underground targets, this experiment used seven LTE radar images as input, with the step size r and dt of the radar data considered unknown. The experiment used recognition accuracy and recall as performance indicators and compared them with the PFJ method. All images used were radar detection results under poor image quality conditions. The experimental results are as follows: Figure 10 As shown in Figure 8

[0170] The identification result of this scheme is as follows:

[0171]

[0172] PFJ method:

[0173] PFJ (Parabolic Fitting-Based Judgment Method) is a method that fits point clusters to parabolas and then discriminates them. The discrimination criteria are as follows:

[0174] A) The fitted hyperbola needs to open downwards.

[0175] B) The fitted hyperbola needs to satisfy the condition that its axis of symmetry lies within the horizontal coordinate range of the segment with the lower opening.

[0176] The following are the identification results using the PFJ method:

[0177]

[0178] (Note: If the connected volume of multiple hyperbolas is identified as a single hyperbola, it is considered an incorrect identification.)

[0179] The comparison shows that the method in this invention has significant advantages in identifying hyperbolas with low signal-to-noise ratios.

[0180] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A hyperbola linearization recognition method based on lower opening scanning, characterized in that, include: Step 1: After removing the direct wave from the ground penetrating radar B scan image, perform binary segmentation to obtain a binary image; Step 2: Perform a lower opening scan on the binary image to obtain point clusters for each lower opening region; Step 3: Crop the image containing the point clusters of each lower opening region, and scan column by column to extract the point segment information of the point clusters of each lower opening region. The point segment information includes the number of point segments in each column, the ordinates of the start and end points of the point segments, and the length of the point segments. Take the midpoint of each point segment and perform HPL processing to obtain a line graph H. The HPL processing specifically involves: S341, Construct a new orbital coordinate system O ′ x ′ y The new orbital coordinate system is a rectangular coordinate system, with the origin being a set reference point; S342, obtain the midpoint of the point segment extracted by column in S33 and use... Transform the midpoint of the point segment to the new orbital coordinate system. O ′ x ′ y The result is a broken line curve. in,( x , y () represents the coordinates of the midpoint of the point segment in the original image; x ′, y ′) represents the coordinates of the midpoint of the point segment in the new orbital coordinate system; where, , y D The maximum value of the midpoint of the point segment extracted from S33; r This refers to the scanning step distance of the ground-penetrating radar antenna. v The speed at which electromagnetic waves propagate in the background medium. t For time; S343, based on the point segment length obtained in S33, expands the polyline curve of S342 into a polyline graph H; Step 4, construct a series of correlation matrices S i The parameters of the correlation matrix S include: slider width, left step count, right step count, and single step distance; where the slider width is 1.5 to 3 times the width of the top of the line chart in step 3; the left and right step counts are respectively taken as the distance column numbers from the top midpoint to the left and right image boundaries in the line chart H in step 3; the single step distance iterates between 1 and the maximum single step distance, i.e. i =1, 2, …, maximum single step distance; the maximum single step distance is the distance in rows from the top midpoint of the line graph in step 3 to the bottom of the image; Calculate the correlation values ​​between the series correlation matrices and the line graph from step 3, and obtain the maximum correlation value; Step 5: If the maximum correlation value obtained in step 4 is greater than or equal to the judgment threshold, then the point cluster corresponding to the line chart is determined to be a hyperbola point cluster.

2. The method as described in claim 1, characterized in that, Binary segmentation is achieved using an adaptive threshold, specifically including: S11, perform median filtering on the image after removing the direct wave to obtain image C1; S12, calculate the global average value of image C1; replace the values ​​of all pixels in the image that are less than the average value with the average value to obtain image C2; S13, using an adaptive thresholding algorithm based on longitudinal gradient information, threshold segmentation is performed on image C2 to obtain binary image C3.

3. The method as described in claim 2, characterized in that, In S13, the gradient region set is defined as: in,( i , j ) represents the pixel coordinates; g is the set minimum gradient threshold; Define adaptive threshold T for: Using adaptive thresholds T Image C2 is segmented to obtain binary image C3: 。 4. The method as described in claim 1, characterized in that, The direct wave is removed by means cancellation, that is, for each pixel, the average value of the row is subtracted from the value of that pixel.

5. The method according to any one of claims 1 to 4, characterized in that, In step 1, a closing-then-opening operation is performed on the binary image.

6. The method as described in claim 1, characterized in that, Methods for obtaining point clusters in the lower opening region include: S21, For a binary image, scan the point segments with downward opening features row by row to obtain the downward opening region and establish point clusters; S22, perform a top-up search and a left and right-tail-down search on the lower opening region. During the upward search, if the point segments in the previous row have vertical overlap with the point segments in the current row, cut off the pixels in the previous row that extend beyond the horizontal area of ​​the current row point segments, and assign the cluster class number of the current row point segments to the pixels in the previous row point segments that do not extend beyond the horizontal area of ​​the current row point segments. Continue searching upward until there are no vertically overlapping point segments above. During the downward search, if the point segments in the next row have vertical overlap with the point segments in the current row, cut off the pixels in the next row that extend beyond the horizontal area of ​​the current row point segments, and assign the cluster class number of the current row point segments to the pixels in the next row point segments that do not extend beyond the horizontal area of ​​the current row point segments. Continue searching downward until there are no vertically overlapping point segments below, and obtain the final point cluster of the current lower opening region.

7. The method as described in claim 1, characterized in that, In step 3, the image containing the point clusters of each lower opening region is cropped as follows: S31, the binary image is divided into multiple sub-images, each sub-image containing only one lower opening region point cluster; the lower opening region point clusters in each sub-image are cropped to obtain the cropped lower opening region image; S32, Perform vertical bilinear interpolation on the cropped lower opening region image; S33, perform column scanning on the interpolated image and extract point segment information by column.

8. The method as described in claim 7, characterized in that, In step S31, image cropping specifically involves: for each sub-image, marking four midpoints of the point cluster: the top, left tail, right tail, and the midpoint of the lower opening; and cropping the point cluster region based on the four midpoints to obtain the cropped lower opening region image.

9. The method as described in claim 1, characterized in that, When calculating the correlation value in step 4, first calculate the correlation between the correlation matrix S and the line graph H, then subtract the sum of the redundancy values ​​of all elements of the redundancy matrix, and use it as the correlation value for that single step distance; the redundancy matrix is: the area where the difference between the line graph H and the correlation matrix S is 1, and the redundancy value is assigned to each point in the area based on the distance between each point in the area and the slider, and the matrix is ​​zero between the area and the redundancy matrix.

10. The method as described in claim 1 or 9, characterized in that, In step 5, the threshold is determined as follows: c 3 × slider width × (number of left steps + number of right steps + 1), where, c 3 is a constant.

Images