Rebar concrete defect positioning method and system based on ground penetrating radar feature fusion
By employing Radon transform, Curvelet transform, and pseudo-color three-channel fusion image technology, combined with the Faster R-CNN network, the problem of blurred defect boundaries caused by steel reinforcement clutter interference was solved, achieving stable localization and accurate detection of internal defects in reinforced concrete.
Patent Information
- Application Number
- CN202611122233.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-08-25
AI Technical Summary
In existing technologies, the interference from steel reinforcement noise is severe, which leads to blurred boundaries and weakened features of internal defects in reinforced concrete, resulting in decreased positioning accuracy and making it difficult to achieve stable positioning in complex steel reinforcement environments.
A clutter suppression method combining Radon transform and Curvelet transform is adopted. By combining region of interest extraction and Gaussian smoothing, a pseudo-color three-channel fused image is constructed. Then, a fast region convolutional neural network (Faster R-CNN) is used for feature fusion and localization to achieve clutter suppression and preservation of defect boundary information.
It improves the accuracy of defect location in complex reinforced concrete environments, reduces the false detection and false detection rates, and achieves stable identification and reliable location of defect boundaries.
Smart Images
Figure CN122636736A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of structural non-destructive testing and ground-penetrating radar signal processing, and in particular relates to a method and system for locating defects in reinforced concrete based on ground-penetrating radar feature fusion. Background Technology
[0002] Reinforced concrete structures are widely used in bridges, tunnels, underground engineering, and building structures due to their excellent mechanical properties and durability. However, during long-term service, defects such as voids, honeycombing, and delamination may develop within the concrete, reducing structural safety and service performance. Therefore, conducting non-destructive testing for internal defects in reinforced concrete is of significant engineering importance.
[0003] Ground penetrating radar (GPR) has been widely used for detecting internal defects in reinforced concrete due to its advantages such as non-contact operation, high detection speed, and sensitivity to internal structures. However, in practical engineering, the strong reflection signals generated by the reinforcing bars can create significant clutter, obscuring and interfering with the defect echoes, leading to blurred defect boundaries, weakened features, and decreased positioning accuracy. Current technologies typically employ filtering, background removal, or time-frequency analysis to suppress this clutter, but these traditional methods, while suppressing rebar reflections, easily result in the loss of defect boundary information, making it difficult to balance preserving the original structural information with enhancing defect features.
[0004] With the development of deep learning technology, target detection methods based on convolutional neural networks are gradually being applied to ground-penetrating radar image analysis. However, most existing methods are trained directly on single-source images, failing to fully utilize the complementary information between the original image and the clutter-suppressed image. This results in room for further improvement in defect localization performance in complex reinforced concrete environments. Therefore, it is urgent to propose a method for locating internal defects in reinforced concrete that can simultaneously address the issues of clutter suppression, preservation of original structural information, and robust localization capabilities in complex environments. Summary of the Invention
[0005] The technical problem to be solved by this invention is to provide a method and system for locating defects in reinforced concrete based on ground-penetrating radar feature fusion, which solves the technical problems of severe reinforcement clutter interference, easy loss of defect boundary information, and insufficient positioning robustness in complex reinforcement environments in the prior art, making it difficult to achieve stable positioning of internal defects in reinforced concrete in complex reinforcement environments.
[0006] The present invention adopts the following technical solution: The method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion includes the following steps, which are performed sequentially: Step 1: Use ground-penetrating radar equipment to scan the reinforced concrete structure and obtain the original ground-penetrating radar B scan data. The original ground-penetrating radar B scan data includes a radar echo data matrix and an original grayscale image generated from the radar echo data matrix. Step 2: Based on Radon transform and Curvelet transform, clutter suppression processing is performed on the radar echo data matrix. Then, combined with region of interest extraction and spatial domain Gaussian smoothing fusion reconstruction, ground penetrating radar data with steel bar clutter suppression and corresponding clutter suppression image are obtained. Step 3: Execute a multi-feature channel-level fusion strategy. Perform pixel-by-pixel difference and absolute value operation between the original grayscale image and the clutter suppression image to obtain the clutter image. Map the original grayscale image, the clutter suppression image, and the clutter image to different color channels of the target three-channel image, and cascade them along the channel dimension to construct a pseudo-color three-channel fused image and obtain pseudo-color three-channel feature fused ground-penetrating radar data. Step 4: Scale the pseudo-color three-channel fused image to the input size required by the target detection network, and input it into the pre-trained target detection network to output the defect category and the corresponding bounding box coordinates; Step 5: Inversely scale and map the bounding box coordinates back to the coordinate system of the original grayscale image, then convert them into the actual physical coordinates of the defect, and overlay the corresponding target detection box onto the original grayscale image to realize the localization of internal defects in reinforced concrete.
[0007] The Radon transform is used for direct wave removal, specifically as follows: a. Mapping radar echo data to the slow-time domain using a linear Radon transform: ; in: Indicates slowness and intercept time The corresponding linear Radon transform result; Indicates along a linear trajectory The intercepted ground-penetrating radar echo signal; This represents the slowness parameter, which is the reciprocal of the speed. This represents the intercept time, which is the time when the trajectory is at zero in the horizontal position; This indicates the horizontal scanning position of the ground-penetrating radar antenna along the survey line, i.e., the spatial position of each radar channel; b. By setting a slowness threshold Extract the direct wave component; c. Subsequently, the direct wave is reconstructed through inverse transformation and removed from the original data to obtain the de-direct wave signal. : ; in: This is the radar echo data matrix; The direct wave signal component reconstructed by the inverse Radon transform; The Curvelet transform is used for multi-scale decomposition and directional constraints, specifically as follows: a. to Perform a fast discrete Curvelet transform based on the Wrapping algorithm to obtain multi-scale, multi-directional coefficients. : ; In the formula, For scale ,direction Spatial location The corresponding Curvelet base functions; For scale parameters, For direction parameters, Spatial location parameters; b. Based on the correspondence between the Curvelet directional sub-bands and the rebar reflection directions, construct a set of rebar clutter directions. ; c. For Curvelet coefficients belonging to the direction set of reinforcement clutter, directional suppression is achieved by proportional attenuation in the non-interest region: ; in: The attenuation coefficient is... ; These are the Curvelet coefficients after direction selection and proportional attenuation suppression; d. Construct statistically adaptive thresholds for each scale: ; in: For the first The mean of the magnitudes of the Curvelet coefficients; For the first Standard deviation of Curvelet coefficient amplitude; To adjust the parameters; For the first Statistical adaptive threshold for scale; e. For coefficients that satisfy the following formula: If so, the corresponding coefficient is set to zero to remove residual weak clutter and random noise; For a strong reflection coefficient that satisfies the following formula: Then further proportional compression is applied: ; in: The strong reflection determination coefficient; For the remaining coefficients, their processed amplitudes remain unchanged.
[0008] The extraction of the region of interest and the spatial domain Gaussian smoothing fusion reconstruction include: a. Region of Interest (ROI) Extraction and Constraints: To avoid excessive attenuation of the defect signal, Hilbert transform and two-dimensional spatial smoothing are used to calculate the local energy of the B-scan data after removing the direct wave. Statistical thresholding is applied to separate high-energy scattering regions above the statistical threshold as the Region of Interest (ROI), forming an ROI mask. Threshold constraints are applied within the ROI, and the initial filtered signal is obtained through inverse transform reconstruction. ; b. Spatial Domain Smoothing Fusion: Introducing a Gaussian Smoothing-based Weighting Function By mitigating abrupt changes at the ROI boundary and achieving a smooth transition, ground-penetrating radar data with reinforced concrete clutter suppressed is obtained. : .
[0009] The pre-trained target detection network employs a Faster R-CNN (Fast Region Convolutional Neural Network) and uses a 50-layer ResNet-50 residual network as the backbone feature extraction network. Its training process is as follows: Acquire historical sample ground-penetrating radar data and corresponding real target bounding box annotation information; The historical sample ground-penetrating radar data includes measured data collected from actual reinforced concrete specimens and simulation data generated based on electromagnetic simulation software; among them, the measured data includes B-scan data collected from the top and side scans through different scanning directions; After processing the historical sample ground-penetrating radar data according to steps one through three, a three-channel feature fusion training dataset is constructed; and manual defect bounding box annotation is performed on the training data. The constructed training samples are uniformly scaled to the input size required by the object detection network, and the coordinates of the corresponding real object bounding boxes are adjusted synchronously according to the image scaling ratio; In the data preprocessing stage, after data augmentation, random affine transformation is introduced, including horizontal mirror flipping with a set proportional probability and random pixel translation with a set pixel value. Invalid bounding boxes that have moved out of the image boundary after translation are simultaneously screened out using the intersection-union threshold. The initial target detection network is then fed into the network, and the network is iteratively trained using a stochastic gradient descent optimizer (SGDM) with momentum. The selected learning rate adjustment strategy is then used to optimize the training, enabling the target detection network to learn the distinction between clutter and defects in multiple feature channels until the target detection network converges, and finally, an internal defect localization model is established.
[0010] The inverse scaling mapping formula is: ; ; In the formula, , These are the required input width and height for the object detection network, respectively. , These are the width and height of the original grayscale image, respectively; These represent the pixel coordinates of the top-left or bottom-right corner of the bounding box, respectively. ,in, and These are the pixel coordinates of the top left and bottom right corners of the bounding box, respectively. and These are the inverse scaling factors for the horizontal and vertical directions, respectively; This represents the original bounding box pixel coordinates after inverse scaling. Similarly, ,in, and These are the pixel coordinates of the top left and bottom right corners of the original bounding box, respectively.
[0011] The conversion of the actual physical coordinates of the defect is as follows: based on the channel spacing of the ground penetrating radar scan, the original horizontal pixel coordinates in the pixel coordinates of the upper left and lower right corners of the original boundary box are converted into horizontal detection distance; based on the relative permittivity of concrete, the propagation speed of electromagnetic waves in the medium is determined, and combined with the propagation speed, the single pixel sampling interval of the ground penetrating radar time window, and the time zero point correction amount, the vertical coordinates representing the original longitudinal time pixels in the original boundary box are converted into the actual burial depth of the defect. Wherein, the horizontal detection distance The calculation formula is: ; In the formula, The spacing between tracks in a ground-penetrating radar scan; The propagation speed of the electromagnetic waves in concrete Based on the relative permittivity of concrete It is determined that, under the condition that the relative permeability approaches 1, the propagation speed is: ; In the formula, The speed of light in a vacuum; The actual burial depth of the defect : ; In the formula, The time sampling interval corresponding to a single pixel. This is the zero-point correction value for ground-penetrating radar. Coefficient 2 is used to convert the two-way propagation time of electromagnetic waves into one-way propagation distance.
[0012] The reinforced concrete defect localization system based on ground-penetrating radar feature fusion adopts the aforementioned ground-penetrating radar feature fusion-based reinforced concrete defect localization method, which includes a ground-penetrating radar acquisition unit, a clutter suppression unit, a multi-feature channel-level fusion unit, a defect detection unit, and a coordinate mapping unit. The ground-penetrating radar acquisition unit is used to acquire the original ground-penetrating radar B scan data; The clutter suppression unit is used to obtain a clutter suppression image; The multi-feature channel-level fusion unit is used to construct a pseudo-color three-channel fused image and obtain pseudo-color three-channel feature fused ground-penetrating radar data; The defect localization unit is used to output the defect category and the corresponding bounding box coordinates; The result mapping unit is used to locate internal defects in reinforced concrete.
[0013] Through the above design scheme, the present invention can bring the following beneficial effects: 1. This invention combines Radon transform with Curvelet multi-scale directional constraints to directionally suppress hyperbolic strong reflection clutter from reinforcing bars, reducing interference from reinforcing bar occlusion on weak defect signals. Simultaneously, it combines Hilbert transform for region of interest extraction and introduces a Gaussian smoothing strategy during reconstruction to retain weak defect boundary information while filtering out clutter, solving the defect edge degradation problem easily caused by traditional global filtering and improving the identifiability of defects in complex reinforcing bar environments.
[0014] 2. This invention obtains a clutter image by calculating the double-precision absolute difference between the original grayscale image and the clutter-suppressed image, and then concatenates it with the original grayscale image and the clutter-suppressed image to construct a pseudo-color three-channel fused image. In the channel color space, the aliased clutter and defects are physically separated, increasing the feature gradient between the target signal and the interfering background. This enables the target detection network to simultaneously extract complementary features from the three channels for feature representation, enhancing the network's ability to describe defect boundaries, improving the stability of defect localization, and reducing the false detection and false negative rates in complex backgrounds.
[0015] 3. This invention combines the scanning channel spacing of ground penetrating radar, the time sampling interval, and the dielectric constant of concrete to physically map the pixel coordinates of the bounding box output by the target detection network, directly outputting the actual horizontal detection distance and defect burial depth parameters, thus realizing a data closed loop from image processing algorithm to engineering measurement and positioning. This solution is applicable to various reinforced concrete defect detection scenarios such as voids, honeycombing, and delamination, improving the robustness and reliability of defect positioning in complex and variable environments. Attached Figure Description
[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: Figure 1 This is a flowchart of the method in the ground-penetrating radar feature fusion-based method and system for locating defects in reinforced concrete according to the present invention. Figure 2 This is a schematic diagram comparing ground-penetrating radar images before and after reinforcement clutter suppression; where: Figure 2 (a) is the original grayscale image containing reinforcement clutter; Figure 2 (b) is a ground-penetrating radar image after clutter suppression; Figure 3 This is a flowchart of the input process for a three-channel feature fusion and target detection network. Figure 4 This is a schematic diagram showing the location of internal defects in the concrete; where: Figure 4 (a) is an example of the top scan detection results; Figure 4 (b) is an example of the side scan detection results. Detailed Implementation
[0017] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.
[0018] The method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion includes the following steps: S1: The ground-penetrating radar equipment is used to scan the reinforced concrete structure to obtain the original ground-penetrating radar B scan data. The original ground-penetrating radar B scan data includes a radar echo data matrix and an original grayscale image generated from the radar echo data matrix. S2: Perform clutter suppression processing on the radar echo data matrix by combining Radon transform and Curvelet transform. Specifically, this includes: extracting and removing the direct wave component in the slow-time domain using linear Radon transform; performing Curvelet multi-scale decomposition on the signal after removing the direct wave; suppressing the rebar clutter in a directional manner by selecting the direction sub-band and using adaptive threshold constraints based on statistical characteristics; and obtaining the clutter-suppressed image by combining region of interest extraction and spatial domain Gaussian smoothing fusion reconstruction. S3: Execute a multi-feature channel-level fusion strategy, perform pixel-by-pixel difference and absolute value operation between the original grayscale image and the clutter suppression image to obtain a clutter image reflecting clutter features; map the original grayscale image, clutter suppression image and clutter image to different color channels of the target three-channel image respectively, and cascade the channels in the depth dimension to construct a pseudo-color three-channel fused image containing original structural features, defect enhancement features and rebar clutter features, and obtain pseudo-color three-channel feature fused ground penetrating radar data; S4: Scale the pseudo-color three-channel fused image to the input size required by the target detection network, and input it into the pre-trained target detection network to output the defect category and the corresponding bounding box coordinates; S5: Use the target detection network to extract multi-feature information from the pseudo-color three-channel fused image, and output the defect category and bounding box coordinates; first, inversely scale and map the bounding box coordinates back to the original grayscale image coordinates, and then convert them into the actual physical coordinates of the defect according to the ground penetrating radar scanning parameters and the dielectric constant of concrete, and then overlay the corresponding target detection box on the original grayscale image to realize the location of reinforced concrete defects.
[0019] In step S2, discrete Curvelet transform is used for multi-scale decomposition to obtain Curvelet coefficients at the corresponding scale, direction, and spatial location. At the scale, low-frequency background components in the preset low-frequency scale set are suppressed as a whole. At the direction and amplitude, Curvelet coefficients belonging to the direction set of rebar clutter in the non-interest region are subjected to the first round of proportional attenuation using an attenuation coefficient. At the same time, an adaptive threshold is constructed based on the statistical characteristics of the Curvelet coefficients at each scale. Coefficients below the adaptive threshold are set to zero to remove noise, and coefficients above the strong reflection judgment threshold are further subjected to the second round of proportional compression, thereby achieving deep suppression of rebar clutter.
[0020] The combination of region of interest extraction and spatial domain smoothing reconstruction in step S2 specifically includes: calculating the local energy of the B-scan data after removing the direct wave using Hilbert transform and two-dimensional spatial smoothing; applying statistical thresholding to separate high-energy scattering regions to form a region of interest mask; within the region of interest, applying a weaker threshold constraint to the Curvelet coefficients to avoid loss of target information; reconstructing the preliminary filtered signal through inverse Curvelet transform; and introducing a weighting function based on Gaussian smoothing to achieve a smooth transition between the target region and the background clutter suppression results, thus obtaining the final clutter-suppressed image.
[0021] In step S3, the process of constructing the pseudo-color three-channel fused image is to convert the original grayscale image and the clutter suppression image into double-precision floating-point data and then perform subtraction and absolute value operations to obtain the clutter image. The original grayscale image is mapped to the red channel to represent the original structure reflection characteristics, the clutter suppression image is mapped to the green channel to represent the target enhancement characteristics after clutter suppression, and the clutter image is mapped to the blue channel to represent clutter characteristics.
[0022] The pre-trained object detection network described in step S4 is obtained through the following training steps: Obtain the original grayscale images of historical samples and the corresponding coordinates of the ground truth target bounding boxes; According to the method described in steps S2 and S3, the original grayscale image of the historical sample is processed and constructed into a pseudo-color three-channel fused image of the historical sample; wherein, the original grayscale image, clutter suppression image and clutter image are all derived from the same original ground penetrating radar B scan data; the original grayscale image, clutter suppression image and clutter image in the pseudo-color three-channel fused image maintain a pixel-by-pixel spatial correspondence, and the position of the corresponding real target bounding box in the three channels is consistent with the defect annotation position in the original grayscale image; The historical sample pseudo-color three-channel fused image is scaled to the input size required by the network, and the corresponding real target bounding box coordinates are adjusted synchronously according to the image scaling ratio; After data augmentation, the data is input into the initial target detection network for iterative training until the network converges, thus obtaining the pre-trained target detection network. The data augmentation process includes introducing random affine transformations, specifically including: horizontal mirror flipping with a set probability and random translation within a specified pixel range in the X and Y axes, and synchronously filtering out invalid bounding boxes that have moved out of the image boundary after translation using a set intersection-union ratio threshold.
[0023] The object detection network mentioned in step S4 is a two-stage object detection network based on deep learning.
[0024] The target detection network adopts the Faster Region-based Convolutional Neural Network (FasterR-CNN) network architecture; the backbone feature extraction network of the target detection network adopts the 50-layer Residual Network (ResNet-50) series network.
[0025] The process of inverse scaling and mapping back to the original physical units to locate defects in reinforced concrete in step S5 includes: inversely mapping the bounding box pixel coordinates output by the network back to the original grayscale image based on the input size of the target detection network and the size of the original grayscale image to obtain the original horizontal and vertical pixel coordinates; converting the original horizontal pixel coordinates into a horizontal detection distance based on the channel spacing of the ground penetrating radar scan; and based on the relative permittivity of the concrete... Determine the speed of electromagnetic waves in a medium The original longitudinal time pixel coordinates are converted into the actual burial depth of the defect by using the propagation speed, the sampling interval of the ground penetrating radar time window, and the time zero point correction amount.
[0026] The reinforced concrete defect localization system based on ground-penetrating radar feature fusion, employing the aforementioned ground-penetrating radar feature fusion-based reinforced concrete defect localization method, includes: Ground penetrating radar acquisition unit is used to acquire raw ground penetrating radar B scan data, which includes radar echo data matrix and raw grayscale image generated from the radar echo data matrix. The clutter suppression unit is used to perform clutter suppression processing combining Radon transform and Curvelet transform on the radar echo data matrix to obtain a clutter suppression image. A multi-feature channel-level fusion unit is used to obtain a clutter image by subtracting the original grayscale image from the clutter suppression image and taking the absolute value, and to map the original grayscale image, the clutter suppression image and the clutter image into red, green and blue color channels and cascade them in the depth dimension to construct a pseudo-color three-channel fusion image; The defect localization unit is used to input the scaled pseudo-color three-channel fused image into a pre-trained target detection network, extract multiple feature information, and output the defect category and bounding box coordinates. The result mapping unit is used to map the bounding box coordinates back to the original physical units through inverse scaling, and to overlay the corresponding target detection box on the original grayscale image to realize the location of reinforced concrete defects.
[0027] An electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion.
[0028] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion. Example
[0029] This embodiment provides a method and system for locating defects in reinforced concrete based on ground-penetrating radar feature fusion. Figure 1 This is a schematic diagram of the overall process of the method in this invention. Reinforced concrete structures typically contain hidden defects such as voids, honeycombing, or delamination. The defect location method specifically includes the following steps: S1: Ground Penetrating Radar Data Acquisition Ground-penetrating radar (GPR) equipment is used to scan the reinforced concrete structure to obtain raw GPR B scan data. The raw GPR B scan data includes a radar echo data matrix, denoted as... And the original grayscale image generated from the radar echo data matrix, denoted as ;therefore, and These data originate from the same set of raw ground-penetrating radar data, but are presented in different processing stages.
[0030] S2: Clutter suppression processing combining Radon and Curvelet transforms Reinforcement clutter suppression processing is applied to the radar echo data matrix, such as... Figure 2 As shown, Figure 2 (a) is the original grayscale image containing reinforcement clutter. Figure 2 (b) is the ground-penetrating radar image after clutter suppression. Specific operations include: S2.1 Direct Wave Removal Based on Radon Transform: A linear Radon transform is used to map radar echo data to the slow-time domain. ; in: Indicates slowness and intercept time The corresponding linear Radon transform result; Indicates along a linear trajectory The intercepted ground-penetrating radar echo signal; This represents the slowness parameter, which is the reciprocal of the speed. This represents the intercept time, which is the time when the trajectory is at zero horizontal position; This indicates the horizontal scanning position of the ground-penetrating radar antenna along the survey line, i.e., the spatial position of each radar channel; By setting a slowness threshold The direct wave component is extracted; then, the direct wave is reconstructed through inverse transform and removed from the original data to obtain the de-direct wave signal. ; in: This is the radar echo data matrix; This represents the direct wave signal component reconstructed through the inverse Radon transform.
[0031] S2.2 Curvelet Multiscale Decomposition and Oriented Constraints: For Perform a fast discrete Curvelet transform based on the Wrapping algorithm to obtain multi-scale, multi-directional coefficients. : ; In the formula, For scale ,direction Spatial location The corresponding Curvelet base functions; For scale parameters, For direction parameters, These are spatial location parameters.
[0032] Hyperbolic reflections of reinforcing bars typically exhibit energy concentration in specific directional subbands within the Curvelet domain. Based on this characteristic, this embodiment first performs overall suppression of low-frequency scale components to reduce the impact of background components and direct wave remnants on subsequent processing.
[0033] Furthermore, based on the correspondence between the Curvelet directional sub-bands and the rebar reflection directions, a set of rebar clutter directions is constructed. For Curvelet coefficients belonging to the reinforcement clutter direction set, directional suppression is achieved using proportional attenuation in the non-interest region: ; in: The attenuation coefficient is... In this embodiment, Take 0.1; These are the Curvelet coefficients after direction selection and proportional attenuation suppression.
[0034] Simultaneously, a statistically adaptive threshold is constructed for each scale: ; in: For the first The mean of the magnitudes of the Curvelet coefficients; For the first Standard deviation of Curvelet coefficient amplitude; To adjust the parameters, For the first Statistically adaptive threshold for scale.
[0035] For coefficients that satisfy the following formula: ; Then set the corresponding coefficient to zero to remove residual weak clutter and random noise.
[0036] For a strong reflection coefficient that satisfies the following formula: ; Further proportional compression is applied: ; in: This is the strong reflection determination coefficient. In this embodiment, Take 2.5.
[0037] For the remaining coefficients, their processed amplitudes remain unchanged.
[0038] The directional suppression of steel reinforcement clutter is achieved by combining low-frequency scale suppression, direction selection, proportional attenuation, adaptive threshold constraint, and strong reflection secondary suppression. S2.3 Region of Interest (ROI) Extraction and Constraints: To avoid excessive attenuation of the defect signal, Hilbert transform and two-dimensional spatial smoothing are used to calculate the local energy of the B-scan data after removing the direct wave. Statistical thresholding is applied to separate high-energy scattering regions to form ROI masks. Weak threshold constraints are applied within the ROIs, and the initial filtered signal is obtained through inverse transform reconstruction. ; S2.4 Spatial Domain Smoothing Fusion: Introducing a Gaussian Smoothing-Based Weighting Function Reduce abrupt changes at ROI boundaries and achieve a smooth transition: .
[0039] Finally, ground-penetrating radar data after suppressing steel reinforcement clutter was obtained. The corresponding images are collectively referred to as clutter suppression images, denoted as... The clutter suppression image, because it highlights the target features, can also be called a defect enhancement image.
[0040] Among them, the spatial weight function The definition is as follows: ; in, This represents a binary mask for the ROI region. Represents a two-dimensional Gaussian kernel. and These represent the smoothed scales in the time and space directions, respectively. This represents a two-dimensional convolution operation. Represents the nonlinear enhancement parameter. This indicates a truncation normalization operation, used to limit the weights to a certain value. Within the range.
[0041] S3: Multi-feature channel-level fusion Figure 3 This is a flowchart of the input process for a three-channel feature fusion and target detection network.
[0042] S3.1 Clutter Image Construction: The original grayscale image is denoted as... The clutter suppression image is denoted as The pixel-level absolute difference between the original grayscale image and the clutter-suppressed image is calculated to effectively capture the amplitude intensity distribution of positive and negative clutter, thus obtaining a clutter image reflecting the energy of the suppressed reinforcing steel. The clutter image is defined as follows: ,in This represents pixel-by-pixel absolute value operation; S3.2 Three-Channel Mapping: The original grayscale image, clutter-suppressed image, and clutter image are paired according to their correspondence and mapped to the red R channel, green G channel, and blue B channel of the target three-channel image, respectively. Channel-level stitching is then performed in the depth dimension to construct pseudo-color three-channel feature fusion ground-penetrating radar data. .
[0043] Specifically: the R channel preserves the original reflection signal of the structure; the G channel enhances the response of internal defects; and the B channel separates the interference energy of the reinforcing steel bars. The three-channel feature fusion data, through multi-dimensional representation mapping at the physical level, separates the aliased clutter and defects in the grayscale space to different color frequency bands, increasing the feature gradient between the target signal and the interference background.
[0044] S4: Image Preprocessing, Data Augmentation, and Network Input The constructed three-channel feature fusion ground-penetrating radar data is input into a pre-trained deep learning target detection network. In this embodiment, the target detection network uses a Faster R-CNN (Fast Region Convolutional Neural Network) and a 50-layer ResNet-50 residual network as the backbone feature extraction network. To adapt to the network input, bicubic interpolation is used. The image size is scaled up; in this embodiment, the scaled image size is 224×224 pixels.
[0045] Furthermore, the training and optimization process of the object detection network is as follows: Historical ground-penetrating radar (GPR) data and corresponding real target bounding box annotations were acquired. The historical GPR data included data from actual reinforced concrete specimens and numerical simulation data generated using electromagnetic simulation software (such as gprMax). The measured data included B-scan data from top and side scans collected from different scanning directions. A three-channel feature fusion training dataset was constructed using methods S1-S3 described above; manual defect bounding boxes were then annotated on the training data. The constructed training samples were uniformly scaled to the input size required by the network, and the coordinates of the corresponding real target bounding boxes were synchronously adjusted according to the image scaling ratio. In the data preprocessing stage, after data augmentation, random affine transformations were introduced, including a 50% probability horizontal mirror flip and a random translation of ±3 pixels. Invalid annotation boxes that moved out of the image boundary after translation were synchronously filtered out using an intersection-over-union (IoU) threshold (IoU=0.5). The initial target detection network is then fed into the network and iteratively trained using a Stochastic Gradient Descent with Momentum (SGDM) optimizer. This training is further optimized using a learning rate adjustment strategy (such as piecewise decay of the learning rate) to enable the network to learn the distinction between clutter and defects across multiple feature channels until the network converges, ultimately establishing an internal defect localization model. In this embodiment, a Faster R-CNN target detection network is used for training. The main hyperparameter settings used for network training are shown in Table 1. It should be noted that the parameters in Table 1 are only one parameter configuration used in this embodiment and do not constitute a limitation on the scope of protection of this invention.
[0046]
[0047] After completing network training according to the parameters shown in Table 1, a pre-trained target detection network is obtained and used for subsequent defect identification and localization experiments. Here, `activation_40_relu` represents the ReLU activation feature layer output by the fourth-stage convolutional block of the ResNet-50 network. When the input image size is 224×224 pixels, the output feature map space size of this layer is 14×14.
[0048] S5: Object Detection and Coordinate Inverse Scaling The trained object detection network is used to detect the test data. The network automatically outputs the category of internal defects in reinforced concrete (void, honeycomb, delamination) and the pixel coordinates of the bounding box in the image feature space. Figure 4 This is a schematic diagram showing the location of internal defects in the concrete. Figure 4 (a) is an example image of the top scan detection results. Figure 4(b) is an example of the side scan detection results. Void represents void defects, Honeycombing represents honeycomb defects, and Delamination represents delamination defects. The numerical value following each category name represents the prediction confidence of the target detection network for the corresponding defect category.
[0049] Since the image was scaled before being input into the network, to ensure the accuracy of engineering measurements, the pixel coordinates of the bounding box output by the network need to be inversely scaled to the original grayscale image size. Let the required input size of the target detection network be... (In this embodiment, the size is set to 224×224 pixels according to step S4), the original grayscale image size is The coordinates of a defect bounding box output by the target detection network are: ,in and These are the pixel coordinates of the top left and bottom right corners of the bounding box, respectively.
[0050] First, the network output coordinates are inversely mapped back to the original image pixel coordinates. The inverse scaling mapping formula is: ; ; in, ; and These are the horizontal and vertical inverse scaling factors, respectively.
[0051] After obtaining the original image coordinates, the pixel coordinates of the original image are converted into actual physical coordinates. For horizontal detection distance... The calculation formula is: .
[0052] in, This refers to the channel spacing in a ground-penetrating radar scan.
[0053] speed of electromagnetic waves in concrete Based on the relative permittivity of concrete Given a relative permeability of approximately 1, the propagation speed is: .
[0054] in, It is the speed of light in a vacuum.
[0055] Using this propagation speed, calculate the actual burial depth of the defect: .
[0056] in The time sampling interval corresponding to a single pixel. This is the zero-point correction value for ground-penetrating radar. Coefficient 2 is used to convert the two-way propagation time of electromagnetic waves into one-way propagation distance.
[0057] Finally, the recovered target detection box is overlaid on the corresponding original grayscale image, and the defect category and confidence label are displayed simultaneously to achieve visualized and accurate localization of internal defects in reinforced concrete.
[0058] To further verify the effectiveness of this invention, comparative experiments were conducted under the same training and test sets. The pseudo-color three-channel fusion method proposed in this invention was compared with the following two benchmark methods: using only the original grayscale image (single channel) and using only the clutter-suppressed image (single channel). The experiments used the mean accuracy (...). The average positioning error is used as an evaluation indicator.
[0059] The evaluation indicators are defined as follows: Average accuracy when the intersection-union ratio threshold is 0.50; The average AP obtained when the crossover ratio threshold is from 0.50 to 0.95 and the step size is 0.05; Average accuracy when the intersection-union ratio threshold is 0.75; Arithmetic mean of AP for each defect category; mean absolute error (ARR) ): Divided into horizontal average positioning error ( ) and depth average positioning error ( All dimensions have been converted to actual physical dimensions using ground-penetrating radar physical parameters, with units in millimeters (mm). The horizontal positioning error is calculated as the horizontal distance error between the center of the predicted bounding box and the center of the actual bounding box; the depth positioning error is the physical depth error between the top edge of the predicted bounding box and the top edge of the actual bounding box. Comparative experimental results are shown in Table 2.
[0060] As shown in Table 2, compared with the original grayscale image and clutter-suppressed image as single-channel input methods, the pseudo-color three-channel fused data constructed in this invention achieves higher detection accuracy in various defect detection tasks. In terms of intersection-to-union (IoU) ratio... When the threshold is 0.50, the overall average precision is calculated using the pseudo-color three-channel fused image as input. The accuracy reached 0.9234, which is higher than 0.7803 and 0.8232 when using the original grayscale image and clutter-suppressed image as single-channel inputs, respectively. Meanwhile, the positioning error... and The thicknesses were reduced to 3.16 mm and 1.43 mm respectively, indicating that the proposed multi-feature channel-level fusion strategy can effectively improve the defect detection and localization performance in complex reinforced concrete environments.
[0061] The above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion, characterized in that: Includes the following steps, And the following steps are performed in sequence. Step 1: Use ground-penetrating radar equipment to scan the reinforced concrete structure and obtain the original ground-penetrating radar B scan data. The original ground-penetrating radar B scan data includes a radar echo data matrix and an original grayscale image generated from the radar echo data matrix. Step 2: Based on Radon transform and Curvelet transform, clutter suppression processing is performed on the radar echo data matrix. Then, combined with region of interest extraction and spatial domain Gaussian smoothing fusion reconstruction, ground penetrating radar data with steel reinforcement clutter suppression and the corresponding clutter suppression image are obtained. Step 3: Execute a multi-feature channel-level fusion strategy. Perform pixel-by-pixel difference and absolute value operation between the original grayscale image and the clutter suppression image to obtain the clutter image. Map the original grayscale image, the clutter suppression image, and the clutter image to different color channels of the target three-channel image, and cascade them along the channel dimension to construct a pseudo-color three-channel fused image and obtain pseudo-color three-channel feature fused ground-penetrating radar data. Step 4: Scale the pseudo-color three-channel fused image to the input size required by the target detection network, and input it into the pre-trained target detection network to output the defect category and the corresponding bounding box coordinates; Step 5: Inversely scale and map the bounding box coordinates back to the coordinate system of the original grayscale image, then convert them into the actual physical coordinates of the defect, and overlay the corresponding target detection box onto the original grayscale image to realize the localization of internal defects in reinforced concrete.
2. The method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion according to claim 1, characterized in that: The Radon transform is used for direct wave removal, specifically as follows: a. Mapping radar echo data to the slow-time domain using a linear Radon transform: ； in: Indicates slowness and intercept time The corresponding linear Radon transform result; Indicates along a linear trajectory The intercepted ground-penetrating radar echo signal; This represents the slowness parameter, which is the reciprocal of the speed. This represents the intercept time, which is the time when the trajectory is at zero horizontal position; This indicates the horizontal scanning position of the ground-penetrating radar antenna along the survey line, i.e., the spatial position of each radar channel; b. By setting a slowness threshold Extract the direct wave component; c. Subsequently, the direct wave is reconstructed through inverse transformation and removed from the original data to obtain the de-direct wave signal. : ； in: This is the radar echo data matrix; The direct wave signal component reconstructed by the inverse Radon transform; The Curvelet transform is used for multi-scale decomposition and directional constraints, specifically: a. to Perform a fast discrete Curvelet transform based on the Wrapping algorithm to obtain multi-scale, multi-directional coefficients. : ； In the formula, For scale ,direction Spatial location The corresponding Curvelet base functions; For scale parameters, For direction parameters, Spatial location parameters; b. Based on the correspondence between the Curvelet directional sub-bands and the rebar reflection directions, construct a set of rebar clutter directions. ; c. For Curvelet coefficients belonging to the direction set of reinforcement clutter, directional suppression is achieved by proportional attenuation in the non-interest region: ； in: The attenuation coefficient is... ; These are the Curvelet coefficients after direction selection and proportional attenuation suppression; d. Construct statistically adaptive thresholds for each scale: ； in: For the first The mean of the magnitudes of the Curvelet coefficients; For the first Standard deviation of Curvelet coefficient amplitude; To adjust the parameters; For the first Statistical adaptive threshold for scale; e. For coefficients that satisfy the following formula: If so, the corresponding coefficient is set to zero to remove residual weak clutter and random noise; For a strong reflection coefficient that satisfies the following formula: Then further proportional compression is applied: ； in: The strong reflection determination coefficient; For the remaining coefficients, their processed amplitudes remain unchanged.
3. The method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion according to claim 2, characterized in that: The extraction of the region of interest and the spatial domain Gaussian smoothing fusion reconstruction include: a. Region of Interest (ROI) Extraction and Constraints: To avoid excessive attenuation of the defect signal, Hilbert transform and two-dimensional spatial smoothing are used to calculate the local energy of the B-scan data after removing the direct wave. Statistical thresholding is applied to separate high-energy scattering regions above the statistical threshold as the Region of Interest (ROI), forming an ROI mask. Threshold constraints are applied within the ROI, and the initial filtered signal is obtained through inverse transform reconstruction. ; b. Spatial Domain Smoothing Fusion: Introducing a Gaussian Smoothing-based Weighting Function By mitigating abrupt changes at the ROI boundary and achieving a smooth transition, ground-penetrating radar data with reinforced concrete clutter suppressed is obtained. : 。 4. The method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion according to claim 1, characterized in that: The pre-trained target detection network employs a Faster R-CNN (Fast Region Convolutional Neural Network) and uses a 50-layer ResNet-50 residual network as the backbone feature extraction network. Its training process is as follows: Acquire historical sample ground-penetrating radar data and corresponding real target bounding box annotation information; The historical sample ground-penetrating radar data includes measured data collected from actual reinforced concrete specimens and simulation data generated based on electromagnetic simulation software; among them, the measured data includes B-scan data collected from the top and side scans through different scanning directions; After processing the historical sample ground-penetrating radar data according to steps one through three, a three-channel feature fusion training dataset is constructed; and manual defect bounding box annotation is performed on the training data. The constructed training samples are uniformly scaled to the input size required by the object detection network, and the coordinates of the corresponding real object bounding boxes are adjusted synchronously according to the image scaling ratio; In the data preprocessing stage, after data augmentation, random affine transformation is introduced, including horizontal mirror flipping with a set proportional probability and random pixel translation with a set pixel value. Invalid bounding boxes that have moved out of the image boundary after translation are simultaneously screened out using the intersection-union threshold. The initial target detection network is then fed into the network, and the network is iteratively trained using a stochastic gradient descent optimizer (SGDM) with momentum. The selected learning rate adjustment strategy is then used to optimize the training, enabling the target detection network to learn the distinction between clutter and defects in multiple feature channels until the target detection network converges, and finally, an internal defect localization model is established.
5. The method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion according to claim 1, characterized in that: The inverse scaling mapping formula is: ； ； In the formula, , These are the required input width and height for the object detection network, respectively. , These are the width and height of the original grayscale image, respectively; These represent the pixel coordinates of the top-left or bottom-right corner of the bounding box, respectively. ,in, and These are the pixel coordinates of the top left and bottom right corners of the bounding box, respectively. and These are the inverse scaling factors for the horizontal and vertical directions, respectively; This represents the original bounding box pixel coordinates after inverse scaling. Similarly, ,in, and These are the pixel coordinates of the top left and bottom right corners of the original bounding box, respectively.
6. The method for locating defects in reinforced concrete based on ground-penetrating radar feature fusion according to claim 5, characterized in that: The conversion of the actual physical coordinates of the defect is as follows: based on the channel spacing of the ground penetrating radar scan, the original horizontal pixel coordinates in the pixel coordinates of the upper left and lower right corners of the original boundary box are converted into horizontal detection distance; based on the relative permittivity of concrete, the propagation speed of electromagnetic waves in the medium is determined, and combined with the propagation speed, the single pixel sampling interval of the ground penetrating radar time window, and the time zero point correction amount, the vertical coordinates representing the original longitudinal time pixels in the original boundary box are converted into the actual burial depth of the defect. Wherein, the horizontal detection distance The calculation formula is: ; In the formula, The spacing between tracks in a ground-penetrating radar scan; The propagation speed of the electromagnetic waves in concrete Based on the relative permittivity of concrete It is determined that, under the condition that the relative permeability approaches 1, the propagation speed is: ; In the formula, The speed of light in a vacuum; The actual burial depth of the defect : ; In the formula, The time sampling interval corresponding to a single pixel. This is the zero-point correction value for ground-penetrating radar. Coefficient 2 is used to convert the two-way propagation time of electromagnetic waves into one-way propagation distance.
7. A reinforced concrete defect localization system based on ground-penetrating radar feature fusion, employing the reinforced concrete defect localization method based on ground-penetrating radar feature fusion as described in claim 1, characterized in that, It includes a ground-penetrating radar acquisition unit, a clutter suppression unit, a multi-feature channel-level fusion unit, a defect detection unit, and a coordinate mapping unit; The ground-penetrating radar acquisition unit is used to acquire the original ground-penetrating radar B scan data; The clutter suppression unit is used to obtain a clutter suppression image; The multi-feature channel-level fusion unit is used to construct a pseudo-color three-channel fused image and obtain pseudo-color three-channel feature fused ground-penetrating radar data; The defect localization unit is used to output the defect category and the corresponding bounding box coordinates; The result mapping unit is used to locate internal defects in reinforced concrete.