Communication construction quality intelligent monitoring method and system and storage medium

Abstract

The invention relates to the field of monitoring, in particular to an intelligent monitoring method and system for communication construction quality and a storage medium, and specifically, an iterative recognition stage is entered by performing global preprocessing on an original image: constructing a vertex probability weighting model based on a gradient field, and screening a candidate feature point set; fitting a candidate hyperbolic curve set by adopting a random consistency sampling algorithm through non-uniform probability sampling and dielectric constant constraint, calculating a product of local energy and phase consistency for each candidate hyperbolic curve as a metric value, screening out a curve with the maximum metric value meeting a preset structure threshold value as a target hyperbolic curve, recording vertex and opening parameters, and obtaining a target hyperbolic curve; and constructing a multi-scale structure tensor by using the geometric morphology of the target hyperbolic curve, carrying out direction-guided anisotropic smoothing and target area removal on the image, using the updated image for the next round of iteration, and after the iteration is finished, summarizing the vertex and opening parameters of all the target hyperbolic curves as a monitoring result.

Description

Technical Field

[0001] This application belongs to the field of monitoring, and in particular relates to an intelligent monitoring method, system and storage medium for communication construction quality. Background Technology

[0002] Ground-penetrating radar (GPR), an important non-destructive testing technology, transmits high-frequency electromagnetic waves underground and receives the reflected signals. GPR is applied to monitoring projects such as underground pipeline laying, roadbed compaction, and tunnel lining quality. By analyzing the detection data, potential quality hazards such as voids, gaps, and abnormal pipeline locations can be detected in a timely manner. The raw data from GPR detection is typically presented as a two-dimensional profile, or B-scan image. In the B-scan image, the electromagnetic wave reflection signals from underground point or columnar targets form a hyperbolic shape with typical characteristics. Accurately, quickly, and automatically detecting and extracting the characteristic parameters of this hyperbola from the B-scan image is key to achieving intelligent monitoring of communication construction quality.

[0003] Image processing methods, such as algorithms based on Hough transform or template matching, while capable of identifying some clearly shaped hyperbolas, generally suffer from high computational costs, sensitivity to noise and clutter interference, and difficulty adapting to variations in hyperbola morphology. Particularly in actual construction sites, severe GPR signal attenuation leads to low image signal-to-noise ratios, and the complex underground environment often involves densely distributed pipelines, resulting in numerous overlapping and intersecting hyperbolas in B-scan images. Traditional algorithms struggle to reliably separate and identify each individual target, easily leading to missed or false detections. While some fitting algorithms based on random sampling consistency improve noise resistance to some extent, the random sampling process is inefficient, and fitting accuracy is difficult to guarantee when dealing with incomplete hyperbolas caused by target occlusion or weak signals. Furthermore, existing methods struggle to handle the challenge of multiple overlapping targets and lack reliable mechanisms to verify whether identified candidate curves are genuine targets, resulting in overall identification accuracy failing to meet the requirements of high-standard engineering quality monitoring. Summary of the Invention

[0004] To address the aforementioned problems, in the first aspect, this invention proposes an intelligent monitoring method for communication construction quality, comprising the following steps: Acquire the raw B-scan image from the ground penetrating radar and perform a global preprocessing on the image; Repeat the iterative identification steps until no new target hyperbola is identified, thus obtaining a set of candidate hyperbolas. For each candidate hyperbola, the pixel sequence on the hyperbola path is extracted, and the product of the local energy and phase consistency of the pixel sequence is calculated as the target metric. The candidate hyperbola with the largest metric value that is greater than the preset structure threshold is selected as the target hyperbola identified in this iteration, and the vertex and opening parameters are recorded. A multi-scale structure tensor is constructed using the geometry of the target hyperbola identified in this round, and the currently processed B-scan image is updated to obtain the B-scan image for the next iteration. After the iteration is completed, the vertex and opening parameters of all identified target hyperbolas are summarized as monitoring results.

[0005] Optionally, the iterative identification step is repeated until no new target hyperbola is identified, thus obtaining a candidate hyperbola set, specifically as follows: The gradient field is calculated based on the currently processed B-scan image, and a vertex probability weighted model is constructed according to the parallelism between the gradient direction and the vertical axis of the image. Each pixel is assigned a vertex weight, and a candidate feature point set is selected by combining the gradient magnitude. Based on the vertex weights, non-uniform probability sampling is performed to select initial fitting points. Within a dielectric constant constraint range determined by the identified target hyperbola opening parameters, the candidate feature point set is iteratively fitted using a random consistency sampling algorithm to obtain a candidate hyperbola set. If this is the first iteration, the constraint range is determined by a preset dielectric constant range.

[0006] Optionally, the update includes: performing directional anisotropic smoothing on the region adjacent to the target hyperbola based on the tensor, and removing the pixel region corresponding to the target hyperbola.

[0007] Optionally, the step of acquiring the raw B-scan image from the ground-penetrating radar and performing a global preprocessing on the image includes: For each time-domain signal in the original B-scan image, the DC component is eliminated by subtracting the mean of the signal from that channel; The energy ratio method is used to automatically pick up the arrival time of the direct wave of each signal, and all signals are shifted on the time axis to align the arrival times of all direct waves to a unified zero point. A Butterworth low-pass filter with a cutoff frequency twice that of the center frequency of the ground-penetrating radar antenna is used to filter the image to smooth high-frequency random noise.

[0008] Optionally, the step of constructing a vertex probability weighted model based on the parallelism between the gradient direction and the vertical axis of the image, assigning a vertex weight to each pixel, and filtering candidate feature point sets based on the gradient magnitude includes: The gradient magnitude G(x,y) and gradient direction angle of each pixel in the image are calculated using the 3x3 Sobel operator. ; According to the formula Calculate the vertex weight W(x,y) for each pixel; Pixels that simultaneously satisfy the conditions of gradient magnitude greater than the gradient magnitude threshold and vertex weight greater than the vertex weight threshold are selected as candidate feature points, forming a candidate feature point set.

[0009] Optionally, the step of iteratively fitting the candidate feature point set using a random consistency sampling algorithm to obtain a candidate hyperbola set includes: In each iteration, four points are randomly selected without replacement from the set of candidate feature points; Based on the four points, a hyperbola is fitted according to the hyperbola equation; Calculate the vertical distance from all other points in the candidate feature point set to the fitted hyperbola, and determine the points whose distance is less than the threshold as interior points; The number of interior points is counted. If the total number of interior points exceeds 15% of the total number of candidate feature points, the fitted hyperbola is recorded as a reliable candidate hyperbola and stored in the candidate hyperbola set.

[0010] Optionally, calculating the product of the local energy and phase consistency of the pixel sequence as the target metric includes: Construct a filter bank consisting of two-dimensional log-Gabor filters with four scales and six directions; The pixel sequence is convolved using the filter bank to obtain the even-symmetric response components of each pixel at each scale s and direction o. Sum of odd symmetric response components ; For the pixel sequence, the formula is used. Calculate phase consistency, where To prevent the division of tiny positive numbers with a denominator of zero; Through formula Calculate local energy; The calculated phase coherence is multiplied by the local energy to obtain the target metric.

[0011] Optionally, the step of constructing a multi-scale structure tensor using the geometry of the target hyperbola identified in this round includes: Calculate the gradient vector of each pixel on the hyperbolic path of the target identified in this round. ; Construct an initial structure tensor at each pixel location ; Define three standard deviations Two-dimensional Gaussian kernel functions for 1, 2, and 4 pixels respectively; The three Gaussian kernel functions are used to convolve and smooth the initial structure tensor field of the entire image. The smoothed structure tensor fields at the three scales are summed pixel by pixel and element by element to obtain the multi-scale structure tensor.

[0012] Optionally, the update includes: performing directional anisotropic smoothing on the region adjacent to the target hyperbola based on the tensor, and removing the pixel region corresponding to the target hyperbola, specifically: The adjacent region is defined as a strip-shaped region centered on the target hyperbola and extending 10 pixels to each side; Eigenvalue decomposition is performed on the multi-scale structure tensor of each pixel in the neighborhood region to obtain the main feature vector. ; Along the perpendicular to the principal eigenvector In the direction of the anisotropic diffusion filter with a step size of 0.1, the region is iteratively smoothed for 5 iterations. After smoothing, the pixel values ​​of the 3-pixel-wide pixel band on the target hyperbola path are set to 0, thus completing the removal.

[0013] In a second aspect, the present invention provides an intelligent monitoring system for communication construction quality, comprising the following modules: The acquisition module is used to acquire the raw B-scan image of the ground penetrating radar and perform a global preprocessing on the image; The execution module is used to repeatedly execute the iterative identification steps until no new target hyperbola is identified, thus obtaining a set of candidate hyperbolas. The calculation module is used to extract the pixel sequence on the hyperbola path for each candidate hyperbola, calculate the product of the local energy and phase consistency of the pixel sequence as the target metric value, select the candidate hyperbola with the largest metric value that is greater than the preset structure threshold as the target hyperbola identified in this iteration, and record the vertex and opening parameters. The monitoring module is used to construct a multi-scale structure tensor using the geometry of the target hyperbola identified in the current round, and to update the currently processed B-scan image to obtain the B-scan image for the next iteration; after the iteration is completed, the vertex and opening parameters of all identified target hyperbolas are summarized as monitoring results.

[0014] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect above.

[0015] This invention constructs a vertex probability weighted model and performs non-uniform probability sampling, which prioritizes high-probability pixels as initial fitting points, improving the search efficiency and accuracy of the random consistency sampling algorithm. During the iteration process, the hyperbolic parameters identified a priori are used to gradually narrow the search range of subsequent fitting parameters, improving the algorithm's convergence speed and fitting accuracy. Furthermore, using the product of local energy and phase consistency as a target metric filters out false targets caused by noise or clutter, ensuring the reliability of the monitoring results. After each successful target identification, the structure tensor is used to locally smooth the image and remove identified regions, reliably eliminating the occlusion effect of strongly reflective targets on surrounding weak or overlapping targets. This achieves precise extraction of complex and dense target groups, enhancing the detection integrity of the method under harsh conditions. Attached Figure Description

[0016] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of the fitting process; Figure 3 This is a diagram illustrating the comparison of target metrics. Detailed Implementation

[0017] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0018] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0019] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0020] In the first embodiment, the present invention proposes an intelligent monitoring method for communication construction quality, such as... Figure 1 This includes the following steps: S1, acquire the raw B-scan image of the ground penetrating radar, and perform a global preprocessing on the image.

[0021] Specifically, raw B-scan data is acquired from ground-penetrating radar equipment, forming a two-dimensional pixel matrix. Preprocessing includes removing the DC component, i.e., calculating the average value for each column of data in the image and subtracting the average value from each pixel value in that column; performing gain processing, for example, using a time-varying gain function to apply a gain coefficient that increases with depth to each row of the image from top to bottom to compensate for the energy attenuation of electromagnetic waves during underground propagation; and performing background noise suppression, calculating the average column vector of all columns in the entire image and subtracting the average column vector from each column of the image to eliminate horizontal stripe interference.

[0022] In an optional embodiment, acquiring the raw B-scan image from the ground-penetrating radar and performing a global preprocessing on the image includes: For each time-domain signal in the original B-scan image, the DC component is eliminated by subtracting the mean of the signal from that channel; The energy ratio method is used to automatically pick up the arrival time of the direct wave of each signal, and all signals are shifted on the time axis to align the arrival times of all direct waves to a unified zero point. A Butterworth low-pass filter with a cutoff frequency twice that of the center frequency of the ground-penetrating radar antenna is used to filter the image to smooth high-frequency random noise.

[0023] The global preprocessing process executes three steps sequentially to enhance the accuracy of subsequent target detection. Taking a raw B-scan image with a depth of 512 pixels and a width of 200 channels as an example, for each of the 200 channels of the A-scan signal, the average value of 512 pixels is calculated independently, and then the calculated average value is subtracted from all pixel values ​​of that channel. For example, if the average value of the signal in channel 10 is 55.3, then the grayscale value of all 512 pixels in that channel will be reduced by 55.3.

[0024] The time-axis offset of each channel signal caused by uneven ground or inconsistent acquisition trigger delays is corrected. For each channel signal after DC component elimination, the direct wave arrival time is calculated using the energy ratio method. This method calculates the ratio of the cumulative energy of the signal from the starting point to the total energy. When this ratio first exceeds a preset low threshold, such as 0.05, the corresponding time point is recorded as the direct wave arrival time. After obtaining the arrival times of all 200 channels, the minimum time value is found as the reference zero time point, and all other channel signals are shifted upwards on the time axis so that their direct wave arrival times are aligned with this reference zero time point.

[0025] The image processed in the first two steps may still contain random noise, especially in the high-frequency range. To smooth the noise, a Butterworth low-pass filter is applied to the entire image. The cutoff frequency of the filter is set according to the hardware parameters of the ground-penetrating radar antenna; for example, if the antenna center frequency is 400MHz, the cutoff frequency is set to 800MHz. Applying a second-order Butterworth low-pass filter to the entire 512x200 image matrix suppresses frequency components above 800MHz, thereby smoothing the image background and clarifying the hyperbolic features without damaging the reflected signal.

[0026] S2, repeat the iterative identification steps until no new target hyperbola is identified, and obtain a set of candidate hyperbolas.

[0027] In one embodiment, the gradient field is calculated based on the currently processed B-scan image, and a vertex probability weighted model is constructed according to the parallelism between the gradient direction and the vertical axis of the image. Each pixel is assigned a vertex weight, and a candidate feature point set is selected by combining the gradient magnitude. Based on the vertex weights, non-uniform probability sampling is performed to select initial fitting points, and within a dielectric constant constraint range determined by the identified target hyperbola opening parameters, the candidate feature point set is iteratively fitted using a random consistency sampling algorithm to obtain a candidate hyperbola set. If this is the first iteration, the constraint range is determined by a preset dielectric constant range.

[0028] The horizontal gradient of the current B-scan image is calculated using the Sobel operator. and vertical gradient Thus, the gradient magnitude G is obtained as and The square root of the sum of squares, with gradient direction θ as... and The arctangent value. The vertex probability-weighted model utilizes the prior knowledge that the gradient direction at the vertices of the hyperbola is close to the vertical direction. For each pixel, the vertex weight... This is the absolute value of the cosine of the angle between the gradient direction at that point and the vertical axis of the image. A gradient magnitude threshold is set when filtering candidate feature point sets. All gradient magnitudes G are greater than The pixels were identified as candidate feature points.

[0029] Three points are randomly selected from the candidate feature point set obtained above as initial fitting points, with the vertex weights of each candidate point as the basis. As for the probability of selection, the higher the vertex weight, the greater the probability of selection. Using the three initial fitting points, a fitting is performed according to the standard hyperbola equation, such as... Figure 2 As shown, the x-coordinates of the vertex of the hyperbola are solved. y-axis And the opening parameters that determine the shape. In the first iteration, the fitted dielectric constant is allowed to vary within a large preset range, such as 3 to 30. In subsequent iterations, this range is narrowed based on the dielectric constant of the target hyperbola successfully identified in the previous round. For example, if the dielectric constant identified in the previous round was 9, the constraint range for this round can be set to 8 to 10. The above sampling, fitting, and counting steps are repeated N times, for example 100 times. Each time, the number of points in the candidate feature point set whose distance to the currently fitted hyperbola is less than a distance threshold is calculated. These points are called inliers. The fitting result with the most inliers is retained as a candidate hyperbola.

[0030] To accurately extract pixels most likely belonging to the vertices of the hyperbola from the preprocessed image, in an optional embodiment, the step of constructing a vertex probability weighting model based on the parallelism of the gradient direction to the vertical axis of the image, assigning a vertex weight to each pixel, and filtering candidate feature point sets based on the gradient magnitude includes: The gradient magnitude G(x,y) and gradient direction angle of each pixel in the image are calculated using the 3x3 Sobel operator. ; According to the formula Calculate the vertex weight W(x,y) for each pixel; Pixels that simultaneously satisfy the conditions of gradient magnitude greater than the gradient magnitude threshold and vertex weight greater than the vertex weight threshold are selected as candidate feature points, forming a candidate feature point set.

[0031] The image is convolved using a 3x3 Sobel operator. The Sobel operator contains two convolution kernels in two directions, which are used to calculate the first-order partial derivatives of the image in the horizontal and vertical directions, respectively, thereby obtaining the gradient component of each pixel. and Based on the two components, using the formula... The gradient magnitude is calculated, reflecting the degree of grayscale change of a pixel; simultaneously, through... Calculate the gradient direction angle, which represents the direction of the fastest change in grayscale.

[0032] A vertex weight is constructed for each pixel. The vertex region of the hyperbola target appears as an approximately horizontal edge in the image, and the gradient direction is therefore close to the vertical direction. Approximately ±90°. The formula is used to calculate the vertex weight of each pixel. When the gradient direction angle is close to vertical, The absolute value of the gradient approaches 1, and the weight value also approaches 1; while when the gradient direction angle is close to horizontal, The absolute value of the value approaches 0, and the weight value also approaches 0. The exponent of 4 makes the weight more sensitive to changes in angle, and can more reliably highlight vertex features.

[0033] A dual thresholding strategy based on gradient magnitude and vertex weights is used. The arithmetic mean of the gradient magnitudes of all pixels in the entire image is calculated, and twice this mean is set as the gradient magnitude threshold. A vertex weight threshold of 0.9 is set. A pixel in the image must simultaneously meet two conditions to be selected into the candidate feature point set: its gradient magnitude must be greater than the gradient magnitude threshold, and its vertex weight must be greater than 0.9. This screening strategy ensures that the obtained candidate points are not only strong edge points, but also have a very high horizontal directional local structure, and are highly likely to be vertices of hyperbolas.

[0034] In an optional embodiment, the step of iteratively fitting the candidate feature point set using a random consistency sampling algorithm to obtain a candidate hyperbola set includes: In each iteration, four points are randomly selected without replacement from the set of candidate feature points; Based on the aforementioned four points, according to the hyperbola equation Fit a hyperbola; Calculate the vertical distance from all other points in the candidate feature point set to the fitted hyperbola, and determine the points whose distance is less than the threshold as interior points; The number of interior points is counted. If the total number of interior points exceeds 15% of the total number of candidate feature points, the fitted hyperbola is recorded as a reliable candidate hyperbola and stored in the candidate hyperbola set.

[0035] Using a random consistency sampling algorithm, a potential hyperbolic model is fitted from a set of candidate feature points containing numerous outliers. An iterative loop is used, with a maximum of 500 iterations. In each iteration, four points are randomly and non-repeatingly selected from the candidate feature point set as the minimum data samples required to fit the hyperbolic model for that round. After selecting the four points, their coordinates are substituted into the general equation of the hyperbola, and a system of linear equations is solved to determine the four parameters of the hyperbola: the coordinates of the center point. and The points, along with their semi-axis lengths a and b, uniquely determine a hyperbola passing through these four points. The process generates a new hyperbola model defined by different random sample points in each iteration.

[0036] After generating the hyperbola model, it needs to be validated. Further iterate through all points in the candidate feature point set except for the four points used for fitting, and calculate the nearest distance from each point to the fitted hyperbola. Set a distance threshold of 3 pixels; if the distance from a point to the hyperbola is less than 3 pixels, the point is classified as an inlier, indicating that it well supports the current hyperbola model. After traversing all points, count the total number of inliers obtained in this iteration. If the number of inliers exceeds 15% of the total number of points in the candidate feature point set (for example, if the candidate point set has 200 points, the number of inliers must exceed 30), then the hyperbola is considered a reliable model, and its parameters are saved to the candidate hyperbola set. Repeat this process 500 times to cover and find all possible target hyperbolas in the image. Figure 2 .

[0037] S3, For each candidate hyperbola, extract the pixel sequence on the hyperbola path, calculate the product of the local energy and phase consistency of the pixel sequence as the target metric value, select the candidate hyperbola with the largest metric value and greater than the preset structure threshold as the target hyperbola identified in this iteration, and record the vertex and opening parameters. For each candidate hyperbola output by the random consistency sampling algorithm, all pixels along the corresponding image path are extracted with a width of 3 pixels to form a pixel sequence. The local energy of this pixel sequence is calculated, which is the sum of the squares of the gray values ​​of all pixels in the sequence. A multi-scale, multi-directional logarithmic Gabor filter bank is used to filter this pixel sequence, and the phase consistency spectrum is calculated to obtain a phase consistency value reflecting the consistency of structural information along the path. The calculated local energy is multiplied by the phase consistency value to obtain the target metric. The metric values ​​of all candidate hyperbolas are compared, and the one with the largest value is selected. Figure 3 The metric value is then compared with a preset structural threshold. The comparison is performed, and if the result is greater than the threshold, the candidate hyperbola is determined as the target hyperbola successfully identified in this round, and the vertex coordinates and opening parameters are stored.

[0038] To represent the prominence of a pixel sequence as a target by analyzing local frequency information of an image, in an optional embodiment, calculating the product of the local energy and phase consistency of the pixel sequence as a target metric includes: Construct a filter bank consisting of two-dimensional log-Gabor filters with four scales and six directions; The pixel sequence is convolved using the filter bank to obtain the even-symmetric response components of each pixel at each scale s and direction o. Sum of odd symmetric response components ; For the pixel sequence, the formula is used. Calculate phase consistency, where To prevent the division of tiny positive numbers with a denominator of zero; Through formula Calculate local energy; The calculated phase coherence is multiplied by the local energy to obtain the target metric.

[0039] A two-dimensional logarithmic-Gabor filter bank is constructed, consisting of 24 independent filters with a Gaussian shape in the frequency domain, capable of reliably detecting local frequency and orientation information in an image. The 24 filters cover four different scales for detecting features of varying sizes, and six different orientations, typically 0°, 30°, 60°, 90°, 120°, and 150°, to achieve a response to edges and textures with diverse orientations.

[0040] The pixel sequence to be evaluated, such as a candidate hyperbola, is convolved with each filter in the filter bank. Since the log-Gabor filter is a complex filter, the convolution result is also a complex number. For each pixel, at each scale s and orientation o, the real part of the convolution result is denoted as the even-symmetric response component. The imaginary part is denoted as the odd-symmetric response component. Even-symmetric components respond most strongly to linear features, while odd-symmetric components respond most strongly to step-like edges.

[0041] Based on the response components of all 24 filters, two core metrics are calculated for each point in the pixel sequence. The first is phase consistency, calculated using a formula. A value close to 1 indicates high phase consistency across all frequency components, suggesting the presence of a clear structural feature; a value close to 0 indicates phase disorder, possibly noise. The second is local energy, calculated using a formula, which is the sum of the amplitudes of all filter responses, reflecting the signal strength. Multiplying these two values ​​yields a metric that considers both the structural clarity and energy intensity of the feature, reliably distinguishing real targets from artifacts.

[0042] S4. Construct a multi-scale structure tensor using the geometry of the target hyperbola identified in this round, and update the currently processed B-scan image to obtain the B-scan image for the next iteration. After the iteration ends, summarize the vertex and opening parameters of all identified target hyperbolas as monitoring results.

[0043] For the target hyperbola identified in this round, the tangent direction at each point on the path is calculated. A structure tensor is constructed based on this tangent direction, with its principal eigenvectors aligned with the tangent direction, representing the local geometry of the region. An anisotropic diffusion filtering algorithm guided by this structure tensor is used to smooth the surrounding area within a certain range around the target hyperbola. The smoothing process proceeds along the hyperbola's path, while directions perpendicular to the path remain sharp. After smoothing, the grayscale values ​​of all interior pixels belonging to the target hyperbola are set to 0, effectively erasing the target from the image. The updated B-scan image serves as the input for the next iteration.

[0044] The iteration terminates when, in a given round of identification, no candidate hyperbola's target metric value exceeds a preset structural threshold, indicating that no identifiable hyperbola target exists in the image. At this point, the vertex coordinates and opening parameters of all target hyperbolas recorded and stored in each previous iteration are summarized to form a list containing the location, depth, and size information of all detected underground targets. This list represents the communication construction quality monitoring results. More specifically, the electromagnetic wave propagation velocity in the local area of ​​each hyperbola is derived using its opening parameters; combined with this velocity and the ordinate of the hyperbola's vertex, the actual burial depth of each target, such as a communication pipeline, is calculated; and by integrating the target's burial depth and vertex abscissa, its two-dimensional spatial coordinates in the underground profile are determined. By summarizing the coordinate information of all targets, a spatial distribution map of the underground pipeline is generated and compared with the construction design drawings, thereby enabling the evaluation of key indicators such as pipeline burial depth, spacing, and location, and completing intelligent monitoring of communication construction quality.

[0045] In an optional embodiment, the step of constructing a multi-scale structure tensor using the geometry of the target hyperbola identified in this round includes: Calculate the gradient vector of each pixel on the hyperbolic path of the target identified in this round. ; Construct an initial structure tensor at each pixel location ; Define three standard deviations Two-dimensional Gaussian kernel functions for 1, 2, and 4 pixels respectively; The three Gaussian kernel functions are used to convolve and smooth the initial structure tensor field of the entire image. The smoothed structure tensor fields at the three scales are summed pixel by pixel and element by element to obtain the multi-scale structure tensor.

[0046] Construct a tensor field that represents the local geometry of the target hyperbola, and calculate the image gradient vector of each pixel constituting the path along the target hyperbola path identified in the previous round. The gradient vector is perpendicular to the tangent direction of the hyperbola and points in the direction of the fastest change in grayscale.

[0047] At each pixel location along the hyperbolic path, a 2x2 initial structure tensor is constructed using the gradient vector. This tensor is calculated by multiplying the gradient vector by itself. The tensor matrix contains information about the edge strength and direction within the pixel's neighborhood. For example, the direction of the principal feature vector is consistent with the gradient direction. The calculation can be extended to the entire image; in regions without hyperbolic features, the gradient is smaller, and the tensor components are correspondingly smaller.

[0048] To integrate structural information with features at different scales, a multi-scale analysis method is employed. Specifically, this involves defining three two-dimensional Gaussian kernel functions with different standard deviations. Set them to 1 pixel, 2 pixels, and 4 pixels respectively. Smaller A value of 1 focuses on very local, detailed structures, while larger values... A value of 4 will integrate structural information from a closer neighborhood. Using the three Gaussian kernels, the three independent component fields of the initial structural tensor field are respectively... field, field and The field is convolved and smoothed, generating three smoothed structure tensor fields at different scales. These three smoothed tensor fields are then summed element-wise at each pixel location to obtain a multi-scale structure tensor field. This tensor field encodes the local geometric features of the image dominated by the target hyperbola.

[0049] In an optional embodiment, the update includes: performing directional anisotropic smoothing on the region adjacent to the target hyperbola based on the tensor, and removing the pixel region corresponding to the target hyperbola, including: The adjacent region is defined as a strip-shaped region centered on the target hyperbola and extending 10 pixels to each side; Eigenvalue decomposition is performed on the multi-scale structure tensor of each pixel in the neighborhood region to obtain the main feature vector. ; Along the perpendicular to the principal eigenvector In the direction of the anisotropic diffusion filter with a step size of 0.1, the region is iteratively smoothed for 5 iterations. After smoothing, the pixel values ​​of the 3-pixel-wide pixel band on the target hyperbola path are set to 0, thus completing the removal.

[0050] Intelligent smoothing of the area surrounding the target and erasure of the identified target from the image are performed. An operation area is defined, which is a strip-shaped region extending 10 pixels to each side of the hyperbola of the newly identified target as its center line. This region is considered to be the neighboring area affected by the target.

[0051] Direction-guided anisotropic smoothing is performed within this strip region. For each pixel in this region, eigenvalue decomposition is performed on the multi-scale structure tensor calculated in the previous step. The decomposition yields two eigenvectors, where the principal eigenvector corresponds to the larger eigenvalue. The direction of the local edges is specified, i.e., along the hyperbola. The smoothing process uses the Perona-Malik diffusion equation, a nonlinear diffusion model where the diffusion coefficient changes with the image gradient. This allows for smoothing of uniform areas while preserving edges. The diffusion is constrained to be perpendicular to the principal feature vector. The smoothing operation is performed along the normal direction of the hyperbola, suppressing noise and artifacts around the hyperbola while maintaining the hyperbola's own sharpness. This process iterates 5 times with a step size of 0.1 to achieve local cleanup.

[0052] After smoothing, to prevent the same target from being repeatedly identified in the next round of detection, the currently identified hyperbola needs to be removed from the image. A pixel band with a width of 3 pixels is defined along the path of the target hyperbola. The grayscale value of all pixels within this pixel band is forced to 0, making the target invisible in subsequent gradient calculations and feature extraction. This allows the algorithm to focus on processing other remaining, unidentified potential targets in the image.

[0053] In a second embodiment, the present invention also provides an intelligent monitoring system for communication construction quality, comprising the following modules: The acquisition module is used to acquire the raw B-scan image of the ground penetrating radar and perform a global preprocessing on the image; The execution module is used to repeatedly execute the iterative identification steps until no new target hyperbola is identified, thus obtaining a set of candidate hyperbolas. The calculation module is used to extract the pixel sequence on the hyperbola path for each candidate hyperbola, calculate the product of the local energy and phase consistency of the pixel sequence as the target metric value, select the candidate hyperbola with the largest metric value that is greater than the preset structure threshold as the target hyperbola identified in this iteration, and record the vertex and opening parameters. The monitoring module is used to construct a multi-scale structure tensor using the geometry of the target hyperbola identified in the current round, and to update the currently processed B-scan image to obtain the B-scan image for the next iteration; after the iteration is completed, the vertex and opening parameters of all identified target hyperbolas are summarized as monitoring results.

[0054] In a third embodiment, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in the first embodiment.

[0055] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, compact disc read-only memory (CD-ROM), optical storage, etc.) containing computer-usable program code.

[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0057] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for intelligent monitoring of communication construction quality, characterized in that, Includes the following steps: Acquire the raw B-scan image from the ground penetrating radar and perform a global preprocessing on the image; Repeat the iterative identification steps until no new target hyperbola is identified, thus obtaining a set of candidate hyperbolas. For each candidate hyperbola, the pixel sequence on the hyperbola path is extracted, and the product of the local energy and phase consistency of the pixel sequence is calculated as the target metric. The candidate hyperbola with the largest metric value that is greater than the preset structure threshold is selected as the target hyperbola identified in this iteration, and the vertex and opening parameters are recorded. A multi-scale structure tensor is constructed using the geometry of the target hyperbola identified in this round, and the currently processed B-scan image is updated to obtain the B-scan image for the next iteration. After the iteration is completed, the vertex and opening parameters of all identified target hyperbolas are summarized as monitoring results.

2. The method according to claim 1, characterized in that, The iterative identification step is repeated until no new target hyperbola is identified, resulting in a candidate hyperbola set. Specifically: The gradient field is calculated based on the currently processed B-scan image, and a vertex probability weighted model is constructed according to the parallelism between the gradient direction and the vertical axis of the image. Each pixel is assigned a vertex weight, and the candidate feature point set is selected by combining the gradient magnitude. Based on the vertex weights, non-uniform probability sampling is used to select initial fitting points. Within a dielectric constant constraint range determined by the identified target hyperbola opening parameters, a random consistency sampling algorithm is used to iteratively fit the candidate feature point set to obtain a candidate hyperbola set. If the current iteration is the first iteration, the constraint range is determined by a preset dielectric constant range.

3. The method according to claim 1, wherein the update comprises: Based on the tensor, directional anisotropic smoothing is performed on the region adjacent to the target hyperbola, and the pixel region corresponding to the target hyperbola is removed.

4. The method according to any one of claims 1-3, characterized in that, The acquisition of the raw B-scan image from the ground-penetrating radar and the performance of a global preprocessing step on the image include: For each time-domain signal in the original B-scan image, the DC component is eliminated by subtracting the mean of the signal from that channel; The energy ratio method is used to automatically pick up the arrival time of the direct wave of each signal, and all signals are shifted on the time axis to align the arrival times of all direct waves to a unified zero point. A Butterworth low-pass filter with a cutoff frequency twice that of the center frequency of the ground-penetrating radar antenna is used to filter the image to smooth high-frequency random noise.

5. The method according to claim 2, characterized in that, The step involves constructing a vertex probability weighting model based on the parallelism between the gradient direction and the vertical axis of the image, assigning a vertex weight to each pixel, and filtering candidate feature point sets based on the gradient magnitude, including: The gradient magnitude G(x,y) and gradient direction angle of each pixel in the image are calculated using the 3x3 Sobel operator. ; According to the formula Calculate the vertex weight W(x,y) for each pixel; Pixels that simultaneously satisfy the conditions of gradient magnitude greater than the gradient magnitude threshold and vertex weight greater than the vertex weight threshold are selected as candidate feature points, forming a candidate feature point set.

6. The method according to claim 2, characterized in that, The step of iteratively fitting the candidate feature point set using a random consistency sampling algorithm to obtain a candidate hyperbola set includes: In each iteration, four points are randomly selected without replacement from the set of candidate feature points; Based on the four points, a hyperbola is fitted according to the hyperbola equation; Calculate the vertical distance from all other points in the candidate feature point set to the fitted hyperbola, and determine the points whose distance is less than the threshold as interior points; The number of interior points is counted. If the total number of interior points exceeds 15% of the total number of candidate feature points, the fitted hyperbola is recorded as a reliable candidate hyperbola and stored in the candidate hyperbola set.

7. The method according to any one of claims 1-3, characterized in that, The calculation of the product of the local energy and phase consistency of the pixel sequence as the target metric includes: Construct a filter bank consisting of two-dimensional log-Gabor filters with four scales and six directions; The pixel sequence is convolved using the filter bank to obtain the even-symmetric response components of each pixel at each scale s and direction o. Sum of odd symmetric response components ; For the pixel sequence, the formula is used. Calculate phase consistency, where To prevent the division of tiny positive numbers with a denominator of zero; Through formula Calculate local energy; The calculated phase coherence is multiplied by the local energy to obtain the target metric.

8. The method according to any one of claims 1-3, characterized in that, The construction of a multi-scale structure tensor using the geometric shape of the target hyperbola identified in this round includes: Calculate the gradient vector of each pixel on the hyperbolic path of the target identified in this round. ; Construct an initial structure tensor at each pixel location ; Define three standard deviations Two-dimensional Gaussian kernel functions for 1, 2, and 4 pixels respectively; The three Gaussian kernel functions are used to convolve and smooth the initial structure tensor field of the entire image. The smoothed structure tensor fields at the three scales are summed pixel by pixel and element by element to obtain the multi-scale structure tensor.

9. The method according to claim 3, characterized in that, The update includes: performing directional anisotropic smoothing on the region adjacent to the target hyperbola based on the tensor, and removing the pixel region corresponding to the target hyperbola, specifically: The adjacent region is defined as a strip-shaped region centered on the target hyperbola and extending 10 pixels to each side; Eigenvalue decomposition is performed on the multi-scale structure tensor of each pixel in the neighborhood region to obtain the main feature vector. ; Along the perpendicular to the principal eigenvector In the direction of the anisotropic diffusion filter with a step size of 0.1, the region is iteratively smoothed for 5 iterations. After smoothing, the pixel values ​​of the 3-pixel-wide pixel band on the target hyperbola path are set to 0, thus completing the removal.

10. A smart monitoring system for communication construction quality, characterized in that, Includes the following modules: The acquisition module is used to acquire the raw B-scan image of the ground penetrating radar and perform a global preprocessing on the image; The execution module is used to repeatedly execute the iterative identification steps until no new target hyperbola is identified, thus obtaining a set of candidate hyperbolas. The calculation module is used to extract the pixel sequence on the hyperbola path for each candidate hyperbola, calculate the product of the local energy and phase consistency of the pixel sequence as the target metric value, select the candidate hyperbola with the largest metric value that is greater than the preset structure threshold as the target hyperbola identified in this iteration, and record the vertex and opening parameters. The monitoring module is used to construct a multi-scale structure tensor using the geometry of the target hyperbola identified in the current round, and to update the currently processed B-scan image to obtain the B-scan image for the next iteration; after the iteration is completed, the vertex and opening parameters of all identified target hyperbolas are summarized as monitoring results.

Images