Intelligent identification method and system for internal diseases of road surface based on ground penetrating radar B-scan image

Abstract

The present application relates to the technical field of ground penetrating radar image intelligent interpretation, in particular to a ground penetrating radar B-scan image pavement internal disease intelligent identification method and system, the method obtains multiple frames of B-scan images collected continuously along the survey line by the vehicle-mounted ground penetrating radar, after the pre-processing of the parameters adjusted by the downstream feedback, the deep convolution target detection is used to locate the disease candidate area and extract the hyperbolic curve shape parameter set, the adjacent frame candidate area is clustered along the survey line according to the shape parameter set to obtain the shape parameter evolution track, the track stability is evaluated to eliminate transient interference and invert the dielectric constant, finally, based on the physical corresponding table, the disease entity is divided into four categories of surface layer void, interlayer peeling, base layer loose and water content anomaly, and the mileage projection distribution map is output, the present application can distinguish four types of pavement internal diseases with significant dielectric property differences under the condition of low signal-to-noise ratio and serious aliasing of B-scan image, and significantly suppresses the false alarm of transient interference.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation technology of ground penetrating radar images, specifically to a method and system for intelligent identification of internal road surface defects in ground penetrating radar B-scan images. Background Technology

[0002] Under long-term vehicle loads and alternating water and temperature conditions, four typical hidden defects gradually develop within the structural layers of highway pavements: surface layer delamination, interlayer peeling, base layer loosening, and abnormal water content. These four defects are located between the asphalt surface layer and the various structural layers, and cannot be detected from the road surface appearance. If not identified and treated in time, they may further develop into road collapse accidents, seriously endangering driving safety. Vehicle-mounted ground-penetrating radar continuously emits high-frequency electromagnetic waves along the highway and receives reflected signals from the pavement structural layers. The reflected signals along the survey line are organized into a two-dimensional profile image, i.e., a B-scan image, according to the time delay-amplitude relationship. This is currently the main method for large-scale detection of internal pavement defects. However, B-scan images have a low signal-to-noise ratio, and the reflected signals from defect areas are highly superimposed with normal reflections between pavement layers and interference signals from underground pipelines. This makes the interpretation of B-scan images highly dependent on experienced radar interpretation engineers. The training period is long, and different interpreters have large differences in the defect identification conclusions of the same B-scan image, which restricts the large-scale application of ground-penetrating radar pavement detection technology.

[0003] Chinese patent application CN114355339A discloses a method and system for identifying pavement voids using radar imagery. This method employs a neural network to identify target voids in preprocessed radar images, then projects all voided targets on the ground-penetrating radar antenna channels onto the same plane for aggregation based on pavement location. Voids and non-voids are determined by a spatial aggregation threshold for the void clusters. However, this approach only identifies voids as a single type of defect, failing to address the unified differentiation of four types of defects with significantly different dielectric properties: surface voids, interlayer peeling, base layer loosening, and abnormal water content. Furthermore, its judgment is based on the spatial aggregation degree of void clusters rather than the physical dielectric properties of the defect itself. This makes it impossible to distinguish voids with highly similar hyperbolic reflection morphologies from defects such as abnormal water content and base layer loosening at the physical mechanism level. It also fails to address the core issue of how to utilize multi-frame redundant information continuously collected along the survey line to suppress transient interference false alarms under conditions of high overlap between defect reflections and normal interlayer reflections in a single-frame B-scan image, and pipeline interference.

[0004] Chinese patent application CN112232392B discloses a data interpretation and identification method for 3D ground-penetrating radar (GPR). Addressing the issues of low efficiency and inconsistent standards among different personnel in manually interpreting 3D GPR horizontal, vertical, and oblique multi-profile data, it proposes an automated data interpretation and identification method with high operational efficiency and consistent standards. However, this solution focuses on the standardized interpretation process of 3D data formats, failing to delve into the physical parameterization relationship between the hyperbolic reflection morphology parameters and dielectric constant of defects in B-scan images. It also lacks a feedback adjustment mechanism between preprocessing parameters and downstream detection confidence levels. This forces the fixed preprocessing pipeline to compromise parameters in scenarios where shallow strong reflections and deep weak reflections coexist, leading to widespread missed detection of weakly reflective defects such as loose deep-layer substrates.

[0005] Chinese patent application CN117289261A discloses an automatic identification method for road surface crack detection based on ground-penetrating radar (GPR). This method fuses road surface images captured by a camera with B-scan images from GPR and feeds them into an embedded system's identification model. After obtaining road crack data, it calculates the road damage rate based on a road damage index formula. However, this scheme still belongs to the single-frame B-scan image target detection paradigm. It does not utilize the continuous evolution trajectory of the morphological parameters of the same defect in multiple B-scan images when the vehicle-mounted GPR travels continuously along the highway. Therefore, it cannot distinguish between real internal road surface defects and transient interference such as transient metal pipeline scattering and random point scattering, resulting in a persistently high false alarm rate for road sections with dense pipelines.

[0006] In summary, existing technologies generally treat the high degree of aliasing between the reflection signals of defects and normal reflections between layers and pipeline interference signals in B-scan images as noise that simply needs to be eliminated at the image level. Each frame of a B-scan image is treated as an independent two-dimensional image sample and fed into a general convolutional target detection network for appearance feature classification. This paradigm fails to utilize the neglected discriminative information of the multi-frame morphological parameter evolution trajectory formed by continuous acquisition along the survey line by vehicle-mounted ground-penetrating radar, and also lacks a feedback adjustment mechanism for upstream preprocessing parameters based on downstream detection confidence in the algorithm chain. This results in the inability to accurately distinguish four types of defects with significantly different dielectric constants: surface layer voids, interlayer peeling, base layer loosening, and abnormal water content. Transient interference is frequently falsely reported as internal pavement defects, and deep, weakly reflective defects are frequently missed. There is an urgent need in this field for an intelligent identification method and system that can uniformly distinguish the four types of internal pavement defects and suppress false alarms of transient interference from the perspective of electromagnetic wave propagation physics, even under conditions of low signal-to-noise ratio and severe aliasing in B-scan images. Summary of the Invention

[0007] To address the core bottleneck of existing technologies for identifying internal road surface defects using ground-penetrating radar (GPR) B-scan images, which typically treats multiple B-scans as independent two-dimensional image samples and employs a fixed unidirectional pipeline from unified preprocessing to single-frame target detection, thus failing to uniformly distinguish four types of internal road surface defects with significantly different dielectric properties under conditions of highly overlapping defect reflections, normal interlayer reflections, and pipeline interference, this invention provides an intelligent identification method and system for internal road surface defects using GPR B-scan images. This method deeply couples three elements—clustering the morphological parameter evolution trajectory of multiple B-scans along the survey line, dielectric constant discrimination based on hyperbolic morphological parameter inversion, and feedback adjustment of upstream preprocessing parameters by downstream detection confidence—to form a closed-loop identification link. Without altering the existing vehicle-mounted GPR hardware acquisition method, it achieves unified intelligent identification and transient interference suppression of four types of internal road surface defects—surface layer delamination, interlayer peeling, base layer loosening, and abnormal water content—from the physical mechanism level of differences in dielectric constant during electromagnetic wave propagation and the continuity of multi-frame morphological parameter evolution trajectories.

[0008] The technical solution of this invention is as follows:

[0009] A method for intelligent identification of internal road surface defects from ground-penetrating radar (GPR) B-scan images includes the following steps: Step S1, acquiring multiple frames of B-scan images continuously collected along the survey line by a vehicle-mounted GPR; Step S2, performing preprocessing on the multiple frames of B-scan images, wherein the parameters of the preprocessing are adjusted by feedback from Step S5; Step S3, using a deep convolutional network target detection architecture to locate candidate regions of defects in each frame of the preprocessed B-scan image, and extracting the hyperbolic reflectance morphological parameter set of the candidate regions; Step S4, performing trajectory clustering on adjacent candidate regions along the survey line according to the morphological parameter set to obtain the morphological parameter evolution representing the same defect entity. Trajectory; Step S5, evaluate the stability of the evolution trajectory to obtain a stability index. Trajectories with a stability index below a first threshold are judged as transient interference and are removed. When the stability index is between the first threshold and the second threshold, the preprocessing parameters of step S2 are adjusted accordingly. When the stability index is above the second threshold, the dielectric constant is inverted from the morphological parameter set of the evolution trajectory. Step S6, classify the diseased entities into four categories based on the inverted dielectric constant: surface layer delamination, interlayer peeling, base layer loosening, and abnormal water content. Output the category, burial depth, and horizontal extension range of each diseased entity, and generate an internal disease distribution map by projecting along the mileage coordinates.

[0010] This invention also provides an intelligent identification system for internal road surface defects using ground-penetrating radar B-scan images, comprising: a radar acquisition module for continuously acquiring multiple frames of B-scan images along the survey line; a preprocessing module for performing preprocessing on the multiple frames of B-scan images, wherein the preprocessing parameters are adjusted by feedback and inversion modules; a defect candidate detection module for locating defect candidate regions in each preprocessed B-scan image using a deep convolutional network target detection architecture and extracting a hyperbolic reflectance morphological parameter set of the candidate regions; and a trajectory clustering module for performing trajectory clustering on adjacent candidate regions along the survey line according to the morphological parameter set to obtain the morphological parameter evolution trajectory representing the same defect entity. The feedback and inversion module is used to evaluate the stability of the evolution trajectory and obtain a stability index. Trajectories with a stability index lower than a first threshold are judged as transient interference and are eliminated. When the stability index is between the first threshold and the second threshold, feedback is sent to the preprocessing module to adjust its preprocessing parameters. When the stability index is higher than the second threshold, the dielectric constant is inverted from the morphological parameter set of the evolution trajectory. The disease classification and output module is used to classify disease entities into four categories according to the inverted dielectric constant: surface layer voids, interlayer peeling, base layer looseness, and abnormal water content. It outputs the category, burial depth, and horizontal extension range of each disease entity and generates an internal disease distribution map by projecting along the mileage coordinates.

[0011] The beneficial effects of this invention are as follows:

[0012] First, by clustering candidate regions in adjacent frames along the survey line according to their morphological parameter sets to obtain the technical characteristics of the morphological parameter evolution trajectory representing the same defect entity, a physical-essential distinction is achieved between real defects within the road surface and transient interference. The mechanism is as follows: As a three-dimensional entity with a certain spatial extension, a defect within the road surface, when continuously scanned by a vehicle-mounted ground-penetrating radar at frame intervals of 5cm to 10cm, will inevitably exhibit a continuous evolution of its morphological parameter set in multiple consecutive B-scan images along the survey line direction. The differences in morphological parameters between adjacent frames are bounded and gradually change frame by frame. In contrast, transient interference such as metal pipeline scattering and random point scattering only appears in one frame or a very few adjacent frames, lacking the continuous evolution characteristic across multiple frames. Compared to schemes that rely solely on multi-channel spatial projection aggregation as the basis for determining voids, this scheme significantly reduces the probability of transient interference in densely piped road sections being misjudged as voids. Furthermore, it does not depend on the hardware configuration of a multi-channel antenna array and is equally applicable to single-channel vehicle-mounted ground-penetrating radar.

[0013] Secondly, by extracting the hyperbolic reflection morphological parameter set of candidate regions and inverting the dielectric constant from the morphological parameter set, the physical parameterization of four types of defects with significantly different dielectric properties—layer voids, interlayer peeling, base layer looseness, and abnormal water content—was achieved. The mechanism is as follows: the propagation speed of electromagnetic waves in the pavement structure layer is uniquely determined by the dielectric constant of the medium. The opening angle, vertex amplitude, and tail fin attenuation rate of the hyperbolic reflection in the B-scan image jointly encode the dielectric constant information. Conversely, the physical properties of the medium in the defect area can be obtained by inverting the dielectric constant from the hyperbolic morphological parameters. The dielectric constant of the air-filled void region is close to 1, the dielectric constant of the water-filled abnormal water content region is in the range of 40 to 80, the dielectric constant of the loose base layer region with loose particles is in the range of 3 to 5, and interlayer peeling is characterized by a jump in the dielectric constant of adjacent pavement structure layers. These four types of defects are significantly separable in terms of dielectric constant. Compared to schemes that rely solely on image appearance features for target detection and classification, this scheme has an irreplaceable advantage in distinguishing between diseases with highly similar hyperbolic morphologies, such as voids and water-bearing anomalies, but with significantly different dielectric properties.

[0014] Third, by using the technical feature of adjusting preprocessing parameters based on feedback when the stability index of the evaluation evolution trajectory falls between the first and second thresholds, the missed detection rate of deep weak reflection defects is suppressed. The mechanism is as follows: Ground penetrating radar B-scan preprocessing needs to suppress shallow direct waves and strong ground reflections while enhancing weak reflections in deep structural layers. Fixing the preprocessing parameters requires a compromise, resulting in the hyperbolic characteristics of weak reflection defects such as loose deep base layers being over-suppressed after preprocessing and unable to be captured by the detection network. This scheme uses the stability index of downstream trajectory clustering as a feedback signal. When the stability index falls within the intermediate zone between the first and second thresholds, it triggers an adjustment of the preprocessing parameters and a re-execution of the detection, allowing the preprocessing parameters to adaptively optimize under scenario-driven conditions, significantly reducing the missed detection rate of deep weak reflection defects. This feedback mechanism represents a fundamental improvement over the fixed-parameter preprocessing of existing unidirectional pipeline technologies.

[0015] Fourth, the three technical features mentioned above produce a synergistic effect that significantly exceeds the sum of the individual effects of each feature. Specifically, multi-frame morphological parameter evolution trajectory clustering provides multiple repeated measurements along the trajectory direction for hyperbolic morphological parameters, making the variance of the morphological parameter set obtained by parameterized hyperbolic fitting much smaller than that of any single-frame fitting result, thereby significantly improving the confidence of downstream dielectric constant inversion. The physical results of dielectric constant inversion, in turn, provide objective criteria based on a physical correspondence table for the feedback adjustment of preprocessing parameters, avoiding the risk of the feedback mechanism falling into a data-driven closed loop. The feedback adjustment of preprocessing parameters further improves the stability of hyperbolic morphological parameter extraction in the next round of detection, enabling trajectory clustering and dielectric constant inversion to achieve a virtuous cycle. The closed-loop recognition link composed of the three technical features enables the overall recognition accuracy of this scheme to be significantly better than the level that can be achieved by using any single technical feature independently in typical difficult scenarios such as shallow strong reflection, deep weak reflection, dense pipeline interference, and severe aliasing. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the intelligent identification method for internal road surface defects using ground-penetrating radar B-scan images provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the architecture of the intelligent identification system for internal road surface defects based on ground-penetrating radar B-scan images provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments described are merely exemplary implementations and should not be construed as limiting the scope of protection of the present invention.

[0019] Reference Figure 1 The method for intelligent identification of internal road surface defects in ground-penetrating radar B-scan images provided in this embodiment includes steps S1 to S6.

[0020] Step S1: Acquire multiple frames of B-scan images continuously collected by the vehicle-mounted ground-penetrating radar along the survey line. This embodiment uses a shielded vehicle-mounted ground-penetrating radar with a center frequency of 400MHz to 900MHz and an operating bandwidth of 200MHz to 2000MHz. The radar travels at a constant speed of 30km / h to 60km / h along the road to be inspected. The radar antenna is 5cm to 15cm from the road surface. The transmitting antenna emits high-frequency electromagnetic waves to the road surface at a pulse repetition frequency of 100Hz to 200Hz. The receiving antenna receives echo signals from various reflective interfaces of the road surface structure layer. Each A-scan signal samples 512 to 1024 points, with a time window set to 50ns to 100ns, covering a road surface structure layer detection depth of 0m to 2.5m. A-scan sampling is triggered every 5cm to 10cm along the survey line. The continuously acquired A-scan signals are organized into a two-dimensional B-scan image along the survey line according to time delay and amplitude. Each frame of the B-scan image covers 200 to 400 sampling channels horizontally and time axis sampling points vertically. The radar acquisition system synchronously records the road mileage coordinates of the vehicle at the moment of acquisition of each A-scan signal. The mileage coordinates are obtained by a high-resolution mileage encoder installed on the wheel, with a resolution of 1cm to 5cm.

[0021] The continuously acquired B-scan images are organized into a sequence of images to be identified according to mileage coordinates. Each B-scan image carries the mileage coordinate range of its first and last A-scans as metadata for subsequent trajectory clustering and mileage projection. The output of this step is a sequence of multi-frame B-scan images arranged in mileage order and its frame-by-frame mileage coordinate metadata.

[0022] Step S2: Preprocessing is performed on the multi-frame B-scan images, and the parameters of the preprocessing are adjusted by feedback from step S5. The preprocessing in this embodiment includes three sub-steps: background removal, gain compensation, and bandpass filtering. Background removal uses the mean subtraction method, calculating the A-scan mean at each sampling time along the survey line direction for each frame of B-scan image and subtracting it from the original image to suppress direct waves generated by direct coupling between the transmitting and receiving antennas and near-surface strong reflection stripes generated by stable ground reflection. Gain compensation uses a time-varying gain curve, applying an exponentially increasing gain coefficient to each A-scan signal along the time axis to compensate for the attenuation loss of electromagnetic waves propagating in the road structure layer, thereby improving the contrast of deep weak reflection signals. Bandpass filtering uses a finite impulse response digital filter with a cutoff frequency of [missing value]. The cutoff frequency in the passband is The transition band width is 50MHz to 150MHz to filter out low-frequency drift and high-frequency random noise.

[0023] Preprocessing parameters include background removal window length Gain curve exponential coefficient Bandpass filter cutoff frequency With upper cutoff frequency Four adjustable parameters. The initial preprocessing parameters for this step are set to... Sampling channel , , When the feedback channel in step S5 is triggered, this step receives parameter adjustment instructions from the feedback channel, updates the four preprocessing parameters as instructed, and then re-performs preprocessing on the corresponding multi-frame B-scan images. The preprocessing results are then sent to step S3 for the next round of disease candidate detection. The output of this step is a sequence of preprocessed multi-frame B-scan images.

[0024] Step S3: A deep convolutional network-based target detection architecture is used to locate candidate disease regions in each preprocessed B-scan image and extract the hyperbolic reflectance morphology parameter set of the candidate regions. This embodiment uses a deep convolutional target detection network based on the YOLOv8 architecture, with CSPDarknet as the backbone network and PAN-FPN as the neck network. The output layer resolution of the detection head is 80×80, 40×40, and 20×20. The detection network uses at least 30,000 B-scan images annotated by experienced radar interpretation engineers according to unified annotation specifications as training samples. The annotation categories cover four common pavement internal disease candidate regions: surface layer delamination, interlayer peeling, base layer loosening, and abnormal water content. Training uses the Adam optimizer with an initial learning rate of... The batch size is 32, and the iteration rounds are 300. The trained detection network takes the preprocessed B-scan image as input and outputs the bounding box coordinates, confidence scores, and candidate class priors for all disease candidate regions in each frame.

[0025] To address the issue of unstable morphological parameter extraction caused by the drift of hyperbola vertex pixels due to normal reflections between road layers and pipeline interference under aliased signals, this step applies a parametric hyperbola fitting asymptotic residual minimization algorithm to each candidate defect region to extract a stable set of morphological parameters. Specifically, for candidate regions where the absolute value of the amplitude is greater than the background noise level... set of reflective pixels The hyperbola is then parametrically fitted using the hyperbola equation, which is:

[0026] ,

[0027] The energy of the fitting residual is defined as:

[0028] ,

[0029] Morphological parameter set Iterative updates along the negative gradient direction of the residual energy:

[0030] ,

[0031] Iterate until the change in residual energy between two adjacent iterations is less than the convergence threshold. Stop when the output residual is minimized, and output the set of morphological parameters. This serves as a stable morphological parameter for the candidate region.

[0032] in: To fit the residual energy, is a scalar with a range of values ​​of . The unit is The value calculated by this formula represents the overall deviation between the hyperbolic model and the measured reflective pixels within the candidate region under the current morphological parameter set. The morphological parameter set is a five-dimensional vector, determined by the opening angle. Vertex amplitude Tail fin attenuation rate Horizontal position of the vertex Vertex time position It consists of five components, obtained through an iterative convergence process, which characterize a complete geometric description of hyperbolic reflection; Let be the opening angle of the hyperbola, and be a scalar with a range of values ​​of . The unit is The value is given by the first component of the morphological parameter set and is related to the propagation speed of electromagnetic waves in the medium. Related, the relationship is ; Let be the amplitude at the vertex of the hyperbola, and be a scalar with a range of values ​​of . The unit is dimensionless (normalized amplitude), given by the measured reflection intensity at the vertex, and is related to the reflection coefficient of the diseased body; Let be the hyperbolic tail fin attenuation rate, and be a scalar with a value range of . The unit is It is given by the attenuation slope of the amplitude of the reflection at the far end of the vertex with time delay, and is related to the dielectric loss characteristics; Let be the horizontal position of the hyperbola vertex, and let be a scalar with a range of values. The unit is The value is given by the horizontal coordinate of the vertex pixel, which corresponds to the center position of the disease in the direction of the survey line; Let be the time position of the hyperbola vertex, and be a scalar with a range of values. The unit is The time coordinate of the vertex pixel is given, which corresponds to the two-way reflection delay of the defect in the pavement structure layer. The value represents the number of reflective pixels participating in the fitting within the candidate region; it is a positive integer and its value ranges from 1 to 2. In this embodiment , The value is determined by the size of the candidate region; subscript The value is the number of the reflective pixel. ; For the first The fitting weights for each reflective pixel are scalars, with values ​​ranging from 1 to 2. The dimensionless signal-to-noise ratio (SNR) of the pixel amplitude and the background noise is normalized, so that high SNR pixels can make a greater contribution in the fitting. For the first The measured time delay of a single reflective pixel is a scalar quantity, with units of . It is given by the vertical coordinate of the pixel; For the first The hyperbolic model prediction delay of a reflective pixel under the current morphological parameter set is a scalar value, in units of . Substituting the hyperbola equation into Calculated; For the first The horizontal coordinate of each reflected pixel is a scalar, in units of . The value is given by the horizontal position of the pixel. Let be the speed of electromagnetic wave propagation in the medium, and be a scalar with a range of values ​​of . The unit is From the set of morphological parameters and Provided jointly; Let be the speed of light in a vacuum, and be a constant with values ​​ranging from 1 to 10. The unit is ; Here, is the regularization term, and is a scalar with a range of values. , dimensionless, is defined as a penalty function for the deviation of the morphological parameter set from the physically reasonable range, so that the iteration converges to a physically interpretable solution; is the regularization coefficient, is a constant, and takes values ​​of . Its unit is To balance the relative importance of data fitting terms and regularization terms; For the first The set of morphological parameters for each iteration is a five-dimensional vector, given by the result of the previous iteration; superscript... The iteration round number, with a value of This embodiment ; Let be the gradient operator over the set of morphological parameters, and be a five-dimensional vector operator, defined as follows: ; Let be the iteration step size, and be a constant with values ​​ranging from 1 to 10. The unit is (and Same dimensions so that the update quantity is the same as (Same dimensions), determined empirically; too large will lead to iteration divergence, too small will lead to slow convergence; The convergence threshold is denoted by , which is a constant and takes the value . The value is determined empirically. The left-hand side of the fitted residual energy is in units of... In the items on the right Units are , Dimensionless Units are The units on both sides are consistent; the iterative update formula has five-dimensional vectors on both sides and the units of each component are consistent with the units on both sides. The corresponding components are consistent.

[0033] The output of this step is the set of all disease candidate regions and their stable morphological parameters for each frame in the preprocessed multi-frame B-scan image sequence. Confidence score, candidate category prior.

[0034] Step S4: Cluster the candidate regions of adjacent frames along the survey line according to the morphological parameter set to obtain the morphological parameter evolution trajectory representing the same disease entity. This step aggregates discrete disease candidate regions in multiple frames of B-scan images along the survey line direction into evolution trajectories representing the same disease entity. Specifically, for the first... The first frame Candidate region morphological parameter set With the The first frame Candidate region morphological parameter set The morphological similarity between the two is defined as the weighted Euclidean distance:

[0035] ,

[0036] in , , .when Less than the set threshold And the difference in mileage coordinates between the two Less than the set threshold At that time, Aggregation into The next node in the evolutionary trajectory starting from point A. Iterate the above aggregation process frame by frame until no next node that meets the conditions can be found, thus obtaining the complete evolutionary trajectory. ,in This represents the total number of nodes contained in the trajectory.

[0037] in: This is a morphological similarity measure between candidate regions in adjacent frames, and is a scalar with a value range of [value range missing]. The unit is dimensionless (normalized by the weight of each component), calculated by this formula, and represents the probability that two candidate regions belong to the same disease entity. The smaller the value, the more likely they are to belong to the same disease entity. For the first Frame number The stable morphological parameter set of each candidate region, a five-dimensional vector, is output from step S3 and is defined as in the aforementioned formula; superscript... Frame number, value , The total number of frames in the image sequence to be identified; subscript The candidate region is the number within this frame, and its value is [value]. , For the first The total number of candidate regions detected in the frame; , , These are the differences in aperture angle, vertex amplitude, and tail fin attenuation rate, respectively, with units of... Dimensionless It is obtained by subtracting the corresponding components of adjacent frames; , , These are the weighting coefficients for the difference in aperture angle, the difference in vertex amplitude, and the difference in tail fin attenuation rate in the metric, respectively, and are constants, with units of [missing values]. Dimensionless In this embodiment, , , Determined by experience, the aim is to make the three components in It has comparable contributions; The aggregation threshold is a constant with a value of [value missing]. It is dimensionless and determined by experience. If it is too large, different disease entities will be incorrectly aggregated into the same trajectory. If it is too small, the evolution trajectory of the same disease entity will be interrupted. The threshold for determining the mileage coordinate difference is a constant, and its value is [value missing]. The unit is It is determined empirically based on vehicle speed and frame rate; The morphological parameter evolution trajectory obtained from aggregation is a node sequence with the number of nodes being [missing information]. ; The total number of nodes in the evolutionary trajectory is a positive integer, determined by the aggregation iteration process. The three terms inside the square root are respectively... dimension Dimensionless Dimensionless dimension Dimensionless, the sum of the three terms is still dimensionless, and the square root is obtained after taking the square root. It is dimensionless, consistent with the left side.

[0038] To address the issue of spatial misalignment between adjacent frames due to cumulative drift of the odometer trigger during vehicle operation, this step further uses the normal interlayer reflection strips of the road surface in the B-scan image as odometer self-calibration anchor points. Specifically, for each frame of the B-scan image, the horizontally stable strip with the largest amplitude along the time axis and the longest extension in the survey line direction is located as the interlayer reflection strip of the road surface, and its position is recorded as the [missing information - likely a specific point or value]. The horizontal pixel position sequence of the frame is Perform normalized cross-correlation on the horizontal pixel position sequence of the interlayer reflection stripe between two adjacent frames:

[0039] ,

[0040] Inter-frame fine-tuning correction is performed on the mileage coordinates of adjacent frames based on the position of the normalized cross-correlation peak.

[0041] ,

[0042] in: For adjacent Frame and the The mileage coordinate correction between frames is a scalar, and its value range is... The unit is The value calculated by this formula represents the cumulative drift of the odometer trigger between adjacent frames; For the first The lateral pixel position sequence of the normal reflection strip between the pavement layers in a frame is a vector with a length equal to the number of pixels in the survey line direction of that frame, in pixels. It is obtained by locating the horizontally stable strip along the time axis of the preprocessed B-scan image. For the first Horizontal pixel position sequence of normal reflection strips between road surface layers Translate the entire pixel along the horizontal direction The displacement vector obtained after 1 pixel. The normalized cross-correlation function is defined as follows: The output is a scalar with a range of values. Dimensionless; For vectors The mean, operators L2 norm, operators Transpose of a vector; The search offset is an integer, with a value range of 1. The unit is pixels; The upper limit of the search range is denoted by , which is a constant and takes the value . Pixels are determined empirically based on vehicle speed and frame rate. The physical length corresponding to a single pixel is a constant, and its value ranges from [value missing]. The horizontal pixel resolution of the radar acquisition system is determined. For the revised first Frame mileage coordinates are scalars, with units of . Used to replace the original Participation in trajectory aggregation determination; The first record of the original odometer encoder Frame mileage coordinates are scalars, with units of . Mileage correction equal to pixel position offset Multiply by the physical length of a single pixel The unit is pixels. , consistent with the left end.

[0043] After performing the self-calibration, the regions of discontinuity in interlayer reflection stripes detected during the self-calibration process will be output as early warning markers for interlayer stripping defects. The output of this step is multiple sets of morphological parameter evolution trajectories. And early warning markers for interlayer peeling disease.

[0044] Step S5: Evaluate the stability of the evolutionary trajectory to obtain a stability index. Trajectories with a stability index below a first threshold are identified as transient disturbances and discarded. When the stability index is between the first and second thresholds, the preprocessing parameters of step S2 are adjusted accordingly. When the stability index is above the second threshold, the dielectric constant is retrieved from the morphological parameter set of the evolutionary trajectory. This step applies this method to each morphological parameter evolutionary trajectory output in step S4. Calculate the normalized variance of its morphological parameters along the trajectory direction: The reciprocal of the normalized variance is defined as the trajectory stability index:

[0045] ,in: For evolutionary trajectory The normalized variance of the morphological parameter is a scalar with a range of values ​​of . , dimensionless, is calculated by this formula, and characterizes the relative fluctuation of the morphological parameters of each node in the evolutionary trajectory from the mean of the trajectory; For the evolutionary trajectory, define the same step S4; Define the total number of nodes contained in the trajectory, as in step S4; subscript The value is the node's index within the trajectory. ; For the trajectory number The stable morphological parameter set of each node is a five-dimensional vector, defined as in the aforementioned formula; Let be the mean vector of the set of morphological parameters of each node of the trajectory, and be a five-dimensional vector. Calculated; The weighting matrix for the morphological parameter components is... The diagonal matrix, whose diagonal elements are derived from the weighting coefficients of the morphological similarity measurement formula in step S4, ensures that each component has a comparable contribution to the normalized variance; operators Transpose of a vector; To prevent small constants with denominators of zero, the value is taken as follows: ,and Same dimension; Let be the trajectory stability index, and let be a scalar with a value range of . , dimensionless, is calculated by this formula, and characterizes the overall consistency of the morphological parameters of each node in the evolutionary trajectory. The larger the value, the more stable the trajectory. To prevent the stability index from becoming an infinitely large small constant, a value is selected. , dimensionless. The relative variance is dimensionless. The reciprocal of the product is the same as the constant, and the reciprocal is also dimensionless. Both the left and right sides are dimensionless.

[0046] stability index With the first threshold Second threshold Comparison to determine subsequent processing: when When the trajectory is determined to be transient interference (such as scattering from metal pipelines, random point scattering, etc.), it is discarded; when When the feedback channel is triggered, the preprocessing parameters of step S2 are adjusted according to the feedback control law described below, and then steps S3 to S5 are re-executed; when When the trajectory is determined to be a reliable pathological entity evolution trajectory, it proceeds to dielectric constant inversion and soft classification output.

[0047] The first threshold in this embodiment Second threshold The stability index is determined empirically from the distribution of the stability index of the training sample set. The feedback control law will reduce the stability index of this round to below a certain value. The set of frame numbers occupied by the trajectory is denoted as the feedback trigger set. ,right In reverse adjustment of preprocessing parameters is performed in each frame: ,in: The preprocessing parameter vector is adjusted as feedback; it is a four-dimensional vector, with components representing the background removal window length, etc. Gain curve exponential coefficient Bandpass filter cutoff frequency Bandpass filter upper cutoff frequency The units are, in order, sampling channels, , , The result is calculated by this formula and used as the input parameter for the next round of preprocessing. The preprocessing parameter vector before feedback adjustment is defined as above and is given by the preprocessing parameters of the previous round. The feedback step size vector is a four-dimensional vector, and its components take values ​​of... Determined by experience; For a symbolic function, the value is... or Dimensionless; For feedback trigger set The mean value of the trajectory stability index for each frame is a scalar, dimensionless. The direction vector for parameter adjustment is a four-dimensional vector, determined based on the feedback triggering mode (when deep, weak reflections dominate). To enhance deep signals; when shallow strong reflections dominate, take (To strengthen suppression), dimensionless. The terms on the right-hand side are... Each component has the same dimension, and the left end All components have consistent dimensions. After the feedback adjustment is completed, steps S2 to S5 are re-executed for the corresponding frame, with a maximum of 3 iterations to avoid a feedback dead loop.

[0048] For stability indicators higher than the second threshold The reliable evolution trajectory is obtained by inverting the dielectric constant from its set of morphological parameters. Based on the physical relationship between the propagation speed of electromagnetic waves in a medium and the dielectric constant, the inversion formula is: ,in: For evolutionary trajectory The inverted relative permittivity corresponding to the diseased entity is a scalar, with a range of values ​​of 1000. , dimensionless, is calculated by this formula and characterizes the relative permittivity of the medium in the region where the diseased entity is located; Let be the estimated electromagnetic wave propagation velocity in the medium containing the diseased entity, where is a scalar and its value ranges from . The unit is The slope of the hyperbola is calculated by substituting the mean of the morphological parameters of each node of the trajectory into the formula for the slope of the asymptote of the hyperbola. For the time position of each vertex of the trajectory The mean is a scalar value, in units of . The mean value is obtained by calculating the mean value from the trajectory nodes; The opening angle of each node of the trajectory The mean is a scalar value, in units of . The mean value is obtained by calculating the mean value from the trajectory nodes; The speed of light in a vacuum is defined as in the aforementioned formula, and its value is... . The right-hand unit is Dimensions and It needs to be homogenized through pixel-to-physical quantity conversion. Specifically... The given value is the reciprocal of the slope of the asymptote of the hyperbola in the (t,x) plane. Its units need to be converted using the B-scan coordinate system. In this embodiment, the horizontal pixel physical length is scanned by multiplying by B. Units are homogenized with the time axis unit conversion factor to achieve this. The final unit is ; The right side is () / () = dimensionless square, which is also dimensionless, consistent with the left side. The output of this step is the reliable disease entity evolution trajectory and its inverted dielectric constant after stability screening. .

[0049] Step S6: Based on the inverted dielectric constant, classify the diseased entities into four categories: surface layer voids, interlayer peeling, base layer looseness, and abnormal water content. Output the category, burial depth, and horizontal extension range of each diseased entity, and generate an internal disease distribution map by projecting along the mileage coordinates.

[0050] To address the generalization requirement for rare disease morphologies not covered by the training set, this step performs constraint correction based on a physical correspondence table when classifying the four types of diseases according to the inverted dielectric constant. The physical correspondence table is defined as follows: Surface voids correspond to dielectric constants at... The interval, the dielectric constant corresponding to the water content anomaly is in In the range where the base layer is loose, the dielectric constant is within a certain range. In the interval, the dielectric constant jump of the adjacent pavement structural layers corresponding to the interlayer stripping is greater than... The classification criteria for the physical correspondence table are:

[0051] ,

[0052] in: For evolutionary trajectory The category determination result corresponding to the diseased entity is taken as surface void. Interlayer peeling Loose grassroots Abnormal water content one; This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. It is dimensionless. For the first The physical reasonable range of dielectric constants corresponding to different types of diseases is given by the physical correspondence table; , Defined as in the aforementioned formula, this provides both physical criteria and confidence weighting. When Not falling into any When this happens, the feedback channel is triggered to return to step S5 for re-evaluation, to prevent the classification output from deviating from the physically reasonable range.

[0053] The estimated burial depth of each diseased entity is as follows: The horizontal extension length is estimated as the difference between the mileage coordinates of the first and last nodes of the trajectory. The horizontal extension width is estimated as the median of the horizontal pixel span of the candidate regions for each node in the trajectory multiplied by... After all reliable pavement defects are summarized in mileage order, they are projected along the mileage coordinates: the highway is divided into segments according to kilometer markers, and for each segment, the types, quantities, and cumulative horizontal extension length of defects falling within that segment are counted. Each segment is then colored according to defect density, and a heat map of defect density along the highway mileage is output. The output of this step is a map showing the types, depths, horizontal extension ranges, and mileage projection distribution of pavement defects, serving as the basis for pavement maintenance decisions.

[0054] This concludes the description of all steps in the intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images provided in this embodiment. This method deeply couples multi-frame morphological parameter evolution trajectory clustering, hyperbolic morphological parameter inversion of dielectric constant, and downstream detection confidence feedback adjustment of upstream preprocessing parameters to form a closed-loop identification link. This enables the unified differentiation of four types of internal road surface defects with significantly different dielectric properties and the suppression of transient interference false alarms, even under conditions of low signal-to-noise ratio and severe aliasing in B-scan images.

[0055] Example 2: Intelligent Identification System for Internal Road Surface Defects Based on Ground Penetrating Radar B-Scan Images

[0056] Reference Figure 2 The intelligent identification system for internal road surface defects provided in this embodiment using ground-penetrating radar B-scan images includes a radar acquisition module, a preprocessing module, a defect candidate detection module, a trajectory clustering module, a feedback and inversion module, and a defect classification and output module. The six modules are electrically connected in sequence to form a closed-loop identification link. The feedback and inversion module has a feedback path to the preprocessing module, forming a system-level feedback closed loop.

[0057] The radar acquisition module includes a shielded vehicle-mounted ground-penetrating radar antenna, a pulse generator, a receiver, an analog-to-digital converter (ADC), and a mileage encoder. The shielded vehicle-mounted ground-penetrating radar antenna is installed at the rear of the detection vehicle, 5cm to 15cm from the road surface, and transmits high-frequency electromagnetic pulses to the road surface at a center frequency of 400MHz to 900MHz and an operating bandwidth of 200MHz to 2000MHz. The pulse generator triggers transmission at a pulse repetition frequency of 100Hz to 200Hz. The receiver receives echo signals from various reflective interfaces of the road surface structure layer. The ADC digitizes each A-scan signal with 512 to 1024 sampling points. The mileage encoder is installed on the wheels of the detection vehicle, with a resolution of 1cm to 5cm, and synchronously acquires the road mileage coordinates of the vehicle with each A-scan signal. The radar acquisition module continuously acquires multiple frames of B-scan images along the survey line at frame intervals of 5cm to 10cm. Each B-scan image carries the mileage coordinate range of its first and last A-scans as metadata, which is output to the mileage coordinate receiving end of the preprocessing module and the defect classification and output module.

[0058] The preprocessing module includes a background removal submodule, a gain compensation submodule, and a bandpass filtering submodule. The background removal submodule processes each frame of B-scan image along the survey line according to the background removal window length. The mean value of the A-scan at each sampling time is calculated and subtracted from the original image to suppress strong signals from direct waves and ground reflections. The gain compensation submodule applies an exponentially increasing gain curve along the time axis to each A-scan signal, with the exponential coefficient being... This is to compensate for electromagnetic wave propagation attenuation. The bandpass filter submodule uses a finite impulse response digital filter with a cutoff frequency of [frequency value missing] in the passband. The upper cutoff frequency is This is to filter out low-frequency drift and high-frequency random noise. The initial preprocessing parameters of the preprocessing module are: Sampling channel , , It is equipped with a parameter adjustment interface connected to the feedback path of the feedback and inversion modules. When the feedback path is triggered, the four preprocessing parameters are updated according to the feedback control law, and the preprocessing is re-executed on the corresponding frame. The preprocessed multi-frame B-scan image sequence is output to the disease candidate detection module.

[0059] The disease candidate detection module includes a deep convolutional target detection submodule based on the YOLOv8 architecture and a parametric hyperbola fitting submodule. The backbone network of the deep convolutional target detection submodule is CSPDarknet, and the neck network is PAN-FPN. The output layer resolution of the detection head is three scales: 80×80, 40×40, and 20×20. At least 30,000 frames of B-scan images labeled by experienced radar interpretation engineers are used as training samples. Training employs the Adam optimizer with an initial learning rate of... The batch size is 32, and the iteration rounds are 300. This submodule takes the preprocessed B-scan image as input and outputs the bounding box coordinates, confidence scores, and candidate category priors for all disease candidate regions in each frame. The parameterized hyperbolic fitting submodule performs iterative fitting by minimizing the asymptotic residual for each disease candidate region. For the set of reflective pixels in the candidate region whose absolute amplitude is greater than the background noise level, parameter fitting is performed according to the hyperbolic equation. The morphological parameter set is iteratively updated along the negative gradient direction of the residual energy. Iteration stops when the change in residual energy between two adjacent rounds is less than the convergence threshold, and the morphological parameter set with the minimum residual is output. The stable morphological parameters of the candidate region are used as the basis for the detection of the disease. The output of the disease candidate detection module is all disease candidate regions and their stable morphological parameter set, confidence score, and candidate category prior for each frame, which are then fed into the trajectory clustering module.

[0060] The trajectory clustering module includes a morphological similarity measurement submodule, an odometry self-calibration submodule, and an aggregation iteration submodule. The morphological similarity measurement submodule calculates the morphological similarity measure between the morphological parameter sets of candidate regions in adjacent frames using a weighted Euclidean distance formula, with weight coefficients set to... , , The mileage self-calibration submodule locates the horizontally stable strip with the largest amplitude and longest extension along the survey line direction along the time axis in each frame of B-scan image as the normal reflection strip between pavement layers. It then performs normalized cross-correlation calculations on the lateral pixel position sequence of the interlayer reflection strips in adjacent frames to obtain the mileage coordinate correction. The interlayer reflection continuity interruption regions detected during the self-calibration process are used as early warning markers for interlayer stripping diseases and output to the disease classification and output module. The aggregation iteration submodule uses morphological similarity as a threshold. Threshold for difference between mileage coordinates To establish the aggregation criteria, candidate regions that meet the criteria between adjacent frames are aggregated into the same evolutionary trajectory. This process is iterated frame by frame until no next node that meets the criteria can be found. The trajectory clustering module outputs multiple sets of morphological parameter evolution trajectories. The data is fed into the feedback and inversion module.

[0061] The feedback and inversion module includes a stability evaluation submodule, a feedback control submodule, and a dielectric constant inversion submodule. The stability evaluation submodule calculates the normalized variance of the morphological parameters of each evolution trajectory and its reciprocal as a trajectory stability index. , with the first threshold Second threshold A comparison is made to determine the subsequent processing path. The feedback control submodule calculates the direction and magnitude of preprocessing parameter adjustments for trajectories where the stability index is between the first and second thresholds using a feedback control law, and outputs parameter adjustment instructions to the preprocessing module along the feedback path, iterating a maximum of three times to avoid feedback loops. The dielectric constant inversion submodule inverts the dielectric constant for reliable evolution trajectories where the stability index is above the second threshold, based on the physical relationship between the slope of the hyperbola asymptote and the electromagnetic wave propagation speed. The output of the feedback and inversion module is the reliable disease entity evolution trajectory and its inverted dielectric constant, which have passed stability screening. This is sent to the disease classification and output module, and at the same time, parameter adjustment instructions are sent to the preprocessing module through the feedback path.

[0062] The disease classification and output module includes a physical correspondence table constraint classification submodule, a geometric parameter estimation submodule, and an odometer projection mapping submodule. The physical correspondence table constraint classification submodule maps the inverted dielectric constant into four disease categories based on the physical correspondence table: dielectric constant at... The interval is determined to be a surface layer void, and is in The interval was determined to be a water content anomaly, and is in The interval is determined to be loose in the base layer, and the dielectric constant jump of adjacent pavement structural layers is greater than 1. When the inversion dielectric constant does not fall into any range, the feedback channel is triggered to return to the feedback and inversion module for re-evaluation. The geometric parameter estimation submodule estimates the burial depth of each defect entity by half the product of the two-way time difference and the electromagnetic wave propagation speed, estimates the horizontal extension length by the difference in mileage coordinates between the first and last nodes of the evolution trajectory, and estimates the horizontal extension width by multiplying the median lateral pixel span of the candidate region of each node by the physical length of a single pixel. The mileage projection mapping submodule divides the highway into segments according to the kilometer markers, counts the types, quantities, and cumulative horizontal extension lengths of defects falling into each segment, colors each segment according to defect density, and outputs a heat map of defect density along the highway mileage. The final output of the defect classification and output module is a map showing the types, burial depths, horizontal extension ranges, and mileage projection distribution of defects within the pavement, serving as the basis for pavement maintenance decisions.

[0063] The six modules mentioned above are sequentially connected via a data bus to form a closed-loop identification link. The feedback path from the feedback and inversion modules to the preprocessing module constitutes a system-level feedback loop, allowing the preprocessing parameters to be adjusted inversely by the downstream detection confidence level, significantly reducing the false negative rate of deep, weakly reflective defects. This system can be deployed on an in-vehicle industrial control computer or an embedded GPU platform. On the NVIDIA Jetson AGX Orin platform, the processing speed can reach 30 frames per second of B-scan images, meeting the real-time detection requirements of a vehicle-mounted GPR traveling at 60 km / h along a highway.

[0064] To verify the effectiveness of this invention, a vehicle-mounted ground-penetrating radar scan was performed on a 30km test section of a highway in a certain province, collecting approximately 60,000 B-scan images. These included 42 real defects verified by on-site core drilling (11 surface layer voids, 9 interlayer peelings, 14 loose base layers, and 8 abnormal water content), and 17 metal pipeline crossings. Using the method of this invention, 40 defects were identified. Compared with the core drilling results, 38 defects were correctly identified, with all categories correctly determined, resulting in an overall identification accuracy of approximately 90.5%. Only one pipeline crossing was misclassified as a void, with a transient interference false alarm rate of approximately 5.9%. In contrast, using the method described in CN114355339A on the same dataset, 36 defects were identified, with only 29 correctly identified, resulting in an overall identification accuracy of approximately 69.0%. Furthermore, the method could not classify the four types of defects, and 6 pipeline crossings were misclassified as voids, with a transient interference false alarm rate of approximately 35.3%. The experimental results show that the method of the present invention has significant advantages over the existing technology in two dimensions: the ability to uniformly distinguish four types of internal diseases and the ability to suppress transient interference.

[0065] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of protection of this invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of this invention.

Claims

1. A method for intelligent identification of internal road surface defects from ground-penetrating radar B-scan images, characterized in that, Includes the following steps: S1. Acquire multiple frames of B-scan images continuously collected by the vehicle-mounted ground-penetrating radar along the survey line; S2. Perform preprocessing on the multi-frame B-scan images; S3. A deep convolutional network is used to locate the disease candidate region in each frame of the preprocessed B-scan image and extract the hyperbolic reflectance morphology parameter set of the candidate region. S4. Along the survey line, the candidate regions of adjacent frames are clustered according to the morphological parameter set to obtain the morphological parameter evolution trajectory representing the same disease entity. S5. Evaluate the stability of the evolution trajectory to obtain a stability index. Trajectories with a stability index lower than a first threshold are determined to be transient interference and are eliminated. When the stability index is between the first threshold and the second threshold, the preprocessing parameters of step S2 are fed back for adjustment. When the stability index is higher than the second threshold, the dielectric constant is inverted from the morphological parameter set of the evolution trajectory. S6. Based on the inverted dielectric constant, the diseased entities are divided into four categories: surface layer voids, interlayer peeling, base layer looseness, and abnormal water content. The category, burial depth, and horizontal extension range of each diseased entity are output, and an internal disease distribution map is generated by projecting along the mileage coordinates.

2. The intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images according to claim 1, characterized in that, The trajectory clustering in step S4 includes: using the weighted Euclidean distance of the difference in opening angle, difference in vertex amplitude, and difference in tail fin attenuation rate in the set of morphological parameters as the morphological similarity measure of candidate regions in adjacent frames, and aggregating candidate regions in adjacent frames with morphological similarity measures less than a set threshold and continuous mileage coordinates into an evolutionary trajectory node sequence of the same disease entity, thereby identifying the spatially recurring morphological parameter change sequence among multiple frames of B-scan images as a continuous evolutionary representation of the same disease entity.

3. The intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images according to claim 2, characterized in that, Step S4 further includes using the normal reflection strips between pavement layers in the B-scan image as mileage self-calibration anchor points: extracting the horizontal pixel positions of the normal reflection strips between pavement layers in adjacent frames, calculating the difference between the horizontal pixel positions of adjacent frames as the horizontal pixel position difference, using the horizontal pixel position difference to perform inter-frame fine-tuning correction on the mileage coordinates of adjacent frames before performing the morphological similarity measurement, and outputting the interlayer reflection continuity interruption areas detected during the self-calibration process as early warning markers for interlayer stripping defects.

4. The intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images according to claim 3, characterized in that, In step S3, the extraction of the hyperbolic reflection morphological parameter set of the candidate region adopts the asymptotic residual minimization algorithm of parameterized hyperbolic fitting: the set of reflective pixels in the candidate region is parameterized according to the hyperbolic equation, and the gradient of the fitting residual with respect to the morphological parameter set is iteratively converged in reverse, and the morphological parameter set with the minimum residual is output as the stable morphological parameters of the candidate region.

5. The intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images according to claim 4, characterized in that, The evaluation of the stability of the evolutionary trajectory in step S5 specifically involves: calculating the normalized variance of the morphological parameters of each node in the evolutionary trajectory, and using the reciprocal of the normalized variance as the trajectory stability index; when the stability index is lower than the first threshold, it is determined to be transient interference and is removed; when the stability index is between the first threshold and the second threshold, feedback is triggered to adjust the preprocessing parameters in step S2, and steps S3 to S5 are re-executed; when the stability index is higher than the second threshold, the stability index is used as the inversion confidence to perform soft classification output for the four types of diseases.

6. The intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images according to claim 5, characterized in that, In step S6, when classifying the four types of defects based on the inverted dielectric constant, a constraint correction based on a physical correspondence table is also included: the physical correspondence table defines that surface layer voids correspond to dielectric constants in the range of 1 to 1.8, abnormal water content corresponds to dielectric constants in the range of 40 to 80, loose base layer corresponds to dielectric constants in the range of 3 to 5, and interlayer peeling corresponds to dielectric constant jumps in adjacent pavement structural layers that are greater than a set value. The dielectric constant ranges corresponding to each type of defect defined in the physical correspondence table are called the physical reasonable ranges. The dielectric constants inverted in step S5 are compared with the physical correspondence table, and the inversion results that deviate from the physical reasonable ranges are triggered to reprocess the feedback channel, so that the classification output is constrained by the physical laws of electromagnetic wave propagation.

7. The intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images according to claim 1, characterized in that, The burial depth estimation of the diseased entity in step S6 is as follows: the burial depth value is calculated by half of the product of the two-way time difference of the hyperbola vertex time position on the time axis of the B-scan image and the propagation speed of the electromagnetic wave in the medium, based on the morphological parameter set extracted in step S3. The propagation speed of the electromagnetic wave in the medium is obtained by substituting the dielectric constant obtained in step S5 into the wave velocity formula.

8. The intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images according to claim 1, characterized in that, The horizontal extension range of the disease entity in step S6 is estimated as follows: the difference between the mileage coordinates corresponding to the first node and the last node of the evolution trajectory is taken as the horizontal extension length of the disease entity along the longitudinal direction of the road, and the median of the lateral pixel span of the candidate region of each node in the evolution trajectory is taken as the lateral extension width of the disease entity along the road surface.

9. The intelligent identification method for internal road surface defects in ground-penetrating radar B-scan images according to claim 1, characterized in that, The step S6, which involves generating an internal defect distribution map by projecting along the mileage coordinates, is as follows: the highway is divided into segments according to the kilometer markers; for each segment, the types, quantities, and cumulative horizontal extension lengths of defect entities falling within that segment are statistically analyzed; each segment is colored according to defect density; and a heat map of defect density along the highway mileage is output.

10. A ground-penetrating radar B-scan image pavement internal defects intelligent identification system, used to implement the ground-penetrating radar B-scan image pavement internal defects intelligent identification method according to any one of claims 1-9, characterized in that, include: The radar acquisition module is used to continuously acquire multiple frames of B-scan images along the survey line; The preprocessing module is used to perform preprocessing on the multi-frame B-scan images, and its preprocessing parameters are adjusted by feedback from the feedback and inversion module. The disease candidate detection module is used to locate disease candidate regions in each frame of preprocessed B-scan image using a target detection architecture based on a deep convolutional network, and to extract the hyperbolic reflectance morphology parameter set of the candidate regions. The trajectory clustering module is used to perform trajectory clustering on candidate regions of adjacent frames along the survey line according to the morphological parameter set, so as to obtain the morphological parameter evolution trajectory representing the same disease entity. The feedback and inversion module is used to evaluate the stability of the evolution trajectory to obtain a stability index. Trajectories with a stability index lower than a first threshold are identified as transient disturbances and are eliminated. When the stability index is between the first threshold and a second threshold, the module is fed back to adjust its preprocessing parameters. When the stability index is higher than the second threshold, the dielectric constant is inverted from the morphological parameter set of the evolution trajectory. The disease classification and output module is used to classify disease entities into four categories based on the inverted dielectric constant: surface layer voids, interlayer peeling, base layer looseness, and abnormal water content. It outputs the category, burial depth, and horizontal extension range of each disease entity and generates an internal disease distribution map by projecting it along the mileage coordinates.

Images