A road internal crack recognition post-processing method based on multi-feature score and region fusion

Abstract

The present application belongs to the technical field of intelligent road disease recognition, and specifically relates to a road internal crack recognition post-processing method based on multi-feature scoring and region fusion. The method obtains multiple initial crack candidate boxes output by a target detection model, fuses the initial candidate boxes based on conditions such as road number continuity, channel range similarity and point interval proximity, and constructs a structure-continuous crack complex. A multi-dimensional feature vector of each crack complex is extracted, which fuses radar signal features, image space features and initial detection statistical features, and a pre-trained scoring model is used to generate a credibility score for each complex. According to the score, cracks with high credibility are selected as the final recognition result. The present application effectively enhances the structure continuity and credibility of the crack recognition result, and improves the transparency and engineering practicability of the decision by introducing an explanation mechanism such as SHAP.

Description

Technical Field

[0001] This invention belongs to the field of intelligent road defect identification technology, specifically relating to a post-processing method for identifying internal road cracks based on multi-feature scoring and regional fusion. Background Technology

[0002] In target detection technology, post-processing methods are typically used to filter accurate results from a large number of initial candidate boxes output by the model. A common approach is to cluster the candidate boxes to construct a candidate target set, then extract multi-dimensional features from this set, and finally use an independent scoring model for secondary judgment, replacing the traditional non-maximum suppression (NMS) method. However, directly applying such general post-processing methods to ground-penetrating radar image identification of internal road cracks has significant technical drawbacks:

[0003] Crack signals in radar images are often affected by noise, causing target detection models to output multiple scattered, discontinuous small candidate boxes, a phenomenon known as "fragmentation." Existing post-processing methods mostly rely on inter-overlap (IoU) clustering in two-dimensional image space, which makes it difficult to effectively identify the inherent relationships between these physically continuous fragmented candidate boxes, easily leading to misjudgment or direct filtering of real cracks.

[0004] Most existing methods rely solely on general image features such as the geometric properties of candidate boxes and initial confidence scores, neglecting the multi-channel physical characteristics of ground-penetrating radar (GPR) data. Each channel of radar data contains rich physical information, such as signal amplitude, energy, and gradient. Treating radar images merely as ordinary grayscale images results in the loss of a significant amount of crucial information related to crack authenticity, thus limiting the accuracy of identification.

[0005] While many post-processing techniques employ scoring models (such as deep neural networks) that offer high accuracy, their decision-making processes are opaque and lack interpretability, making them unsuitable for high-reliability applications like road engineering. Therefore, existing technologies fall short in terms of structured fusion, feature extraction dimensionality, and scoring interpretability when handling tasks with unique data structures and high reliability requirements, such as crack identification in radar imagery, thus failing to fully meet practical engineering needs. Summary of the Invention

[0006] To address the aforementioned technical problems, this application provides a post-processing method for identifying internal road cracks based on multi-feature scoring and region fusion. This method includes the following steps:

[0007] S1. Obtain initial crack identification results: Based on the target detection model, process the slice image generated by the ground penetrating radar data to obtain multiple initial crack candidate boxes containing location information and confidence scores;

[0008] S2. Constructing a crack complex: Spatial region fusion is performed on the multiple initial crack candidate boxes to construct a structurally continuous crack complex; the fusion is based on preset conditions of channel number continuity, channel range similarity, and overlapping or adjacent point intervals;

[0009] S3. Extracting multidimensional features: For each of the crack complexes, extract a multidimensional feature vector to characterize its credibility;

[0010] S4. Perform credibility scoring: Input the multidimensional feature vector into a pre-trained scoring model to generate a credibility score for each crack complex;

[0011] S5. Screening and Output: Based on the credibility score, the crack complexes are screened, and crack complexes with scores higher than a preset threshold are retained as the final identification results and output in a structured manner.

[0012] Furthermore, the spatial region fusion in step S2 specifically includes determining whether any two initial crack candidate boxes belong to the same crack complex based on one or more of the following conditions:

[0013] The channel number ranges of the two candidate boxes are adjacent or the interval is within the preset channel number interval threshold;

[0014] The range intersection-over-union (IoU) of the two candidate boxes in the channel dimension is greater than the preset channel IoU threshold;

[0015] Two candidate boxes may overlap in the radar depth point dimension or their distance may be less than a preset point interval threshold.

[0016] Furthermore, the multidimensional feature vector extracted in step C includes at least one of the following types of features:

[0017] Radar signal characteristics, including at least one of amplitude difference, gradient mean, and energy difference ratio in the channel direction;

[0018] Image spatial features, including at least one of the following: grayscale standard deviation ratio and image sharpness index;

[0019] Statistical characteristics of the initial detection results, including the maximum, average, or standard deviation of the YOLO confidence scores of all candidate boxes that make up the complex;

[0020] Structural fusion features include the number of candidate boxes that make up the composite, i.e., the number of fusion attempts.

[0021] Furthermore, the amplitude difference in the channel direction is specifically the average of the difference between the maximum amplitude within the crack region and the maximum amplitude within the adjacent non-crack region;

[0022] The image sharpness index is specifically the image sharpness of the crack region calculated using the Laplacian operator.

[0023] Furthermore, the structural fusion feature also includes a logarithmic transformation term of the number of fusion frames, or an interaction term between the number of fusion frames and other features.

[0024] Furthermore, the scoring model in step S4 is a nonlinear scoring model constructed based on the XGBoost algorithm.

[0025] Furthermore, the scoring model uses the credibility level of expert annotations on the crack complex as training labels.

[0026] Furthermore, the method also includes an interpretation step, specifically:

[0027] The SHAP method is introduced to calculate the marginal contribution value of each feature in the multidimensional feature vector to the credibility score result, so as to realize the interpretability analysis of the score result.

[0028] Furthermore, the structured output in step S5 also includes the scoring dominant features of each crack complex and its SHAP contribution ranking.

[0029] The beneficial effects of this invention are:

[0030] This invention proposes a post-processing method for identifying internal road cracks based on multi-feature scoring and regional fusion, which aims to overcome the problems of fragmented identification results, single feature dimension, and opaque decision-making process in existing technologies when processing crack identification in radar maps.

[0031] This invention proposes a post-processing scheme for constructing a "crack complex," which involves spatially aggregating fragmented initial candidate boxes output by the target detection model based on the continuity of channel numbers, the similarity of channel ranges, and the proximity of depth point intervals. This scheme merges physically related but scattered detection boxes into a structurally complete crack target, thereby solving the problem of false deletions or missed detections caused by discontinuous recognition results. This significantly enhances the structural continuity and regional consistency of the final recognition results, making them more consistent with engineering practice.

[0032] This invention, based on the construction of a crack complex, designs a multi-dimensional feature extraction strategy that integrates the physical characteristics of radar signals and image spatial features. This strategy is no longer limited to traditional image grayscale and texture features, but innovatively introduces physical features that reflect the essence of radar signals, such as amplitude differences, energy difference ratios, and gradient mean values ​​along the channel direction. By constructing comprehensive feature vectors, the model's ability to determine the authenticity of cracks is greatly improved, especially in complex scenarios such as weak signals and strong interference, effectively reducing the false detection rate and increasing the reliability of the recognition results.

[0033] This invention introduces an "XGBoost+SHAP" credibility scoring and interpretation mechanism. This mechanism utilizes an XGBoost model trained with expert-annotated data to generate an accurate quantitative credibility score for each crack complex; then, it combines this with the SHAP method to clearly reveal the key features affecting the score and their contribution. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0036] To make the purpose, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0038] Example 1:

[0039] This embodiment proposes a post-processing method for identifying internal road cracks based on multi-feature scoring and region fusion. This method aims to overcome the problems of fragmented identification results, single feature dimension, and opaque decision-making process in existing technologies when processing crack identification from radar images. The specific steps are as follows:

[0040] S1. Obtain initial crack identification results: Based on the target detection model, the slice image generated by the ground penetrating radar data is processed to obtain multiple initial crack candidate boxes containing location information and confidence scores.

[0041] This embodiment generates sliced ​​images based on raw ground-penetrating radar data and uses a trained target detection model to identify cracks in each slice. YOLOv8 is used as the target detection model in this embodiment. The sliced ​​images are generated by segmenting the radar data along the point count dimension. Each image corresponds to a fixed depth range. The horizontal axis represents the channel number (14 channels in total), and the vertical axis corresponds to the radar depth points (0–256). Each image corresponds to a channel number (i.e., the radar acquisition point number). The YOLO model outputs multiple crack candidate boxes, each containing spatial location information and a confidence score. The spatial location information includes the channel number, channel number range, and point count range. Furthermore, the YOLO model can simultaneously identify non-defect targets such as bridge decks and repair sections. By eliminating these non-target areas, the purity of the crack identification results is effectively improved, providing a reliable foundation for subsequent region fusion and multi-feature scoring.

[0042] S2. Constructing a crack complex: Spatial region fusion of the multiple initial crack candidate boxes to construct a structurally continuous crack complex; the fusion is based on preset conditions of channel number continuity, channel range similarity, and overlapping or adjacent point intervals.

[0043] To enhance the structural continuity and spatial consistency of crack identification results, this invention performs spatial fusion processing on the initial crack identification boxes output by the YOLO model to construct a "crack complex". Each crack identification box... Represented as:

[0044]

[0045] in This indicates the coverage area of ​​the identification box in the radar acquisition channel number dimension; This indicates the range of the recognition box in the channel direction (1~14); This indicates the range of the recognition box in the radar point count (depth direction).

[0046] Two recognition boxes and Whether a structure can be merged into a single fracture complex depends on the following three conditions:

[0047] (1) Condition for continuity of track number

[0048]

[0049] (2) Channel range similarity (IoU) condition

[0050]

[0051] (3) Overlapping or adjacent point ranges

[0052]

[0053] The bounding boxes that meet the above three conditions will be spatially fused using region clustering or disjoint-set data structure algorithms to construct a crack complex. , where 𝑁𝑘 represents the number of slice recognition boxes contained in the 𝑘th complex.

[0054] This regional fusion operation utilizes a two-dimensional "channel number-point number" space for neighborhood search and bounding box aggregation. This ensures that the complex covers the actual longitudinal structural defects present in the radar map, while avoiding fragmented identification results from interfering with subsequent feature extraction and reliable scoring. After fusion, each crack complex contains multiple original slice identification boxes, exhibiting good engineering interpretability and regional consistency, providing processing units for subsequent multi-dimensional feature aggregation and scoring modeling.

[0055] S3. Extracting multidimensional features: For each of the crack complexes, extract a multidimensional feature vector to characterize its credibility.

[0056] This step extracts image spatial features and radar signal features for each slice crack box and its corresponding non-crack region, and aggregates all features to a comprehensive level to form the input vector for the subsequent scoring model. The extracted features include the following:

[0057] (1) Average value of channel direction amplitude difference

[0058] For each channel 𝑖, take the maximum absolute value in the crack region:

[0059]

[0060] Take the maximum value in adjacent non-cracked regions. The difference between the two is Finally, the average value is taken as the amplitude characteristic:

[0061]

[0062] (2) Mean of maximum gradient in the channel direction

[0063] For each channel's grayscale change, calculate the first-order gradient along the channel direction (e.g., using the Sobel operator), and take the maximum value:

[0064]

[0065] The final average value is taken from all 14 channels:

[0066]

[0067] (3) Energy difference ratio

[0068] Calculate the energy (sum of squared gray levels) in the crack region and the background region separately:

[0069]

[0070] Construction energy difference ratio:

[0071]

[0072] (4) Gray standard deviation ratio

[0073] Normalize the grayscale values ​​of the slices to 0–255 and calculate the standard deviation of the grayscale values. The difference ratio is:

[0074]

[0075] (5) Sharpness index (Laplacian)

[0076] Image sharpness of the crack region is calculated using the Laplacian operator:

[0077]

[0078] (6) YOLO confidence score statistics

[0079] This includes the maximum confidence level, average confidence level, and standard deviation of all crack frames within the complex, reflecting the stability and reliability of the initial test results.

[0080] (7) Number of times the fracture complex is fused

[0081] This refers to the number of valid bounding boxes contained within the complex, reflecting the spatial consistency and persistence of the region across slices with different point counts. A larger value indicates that the region is more likely to be a stable, real crack. This value can be used directly or its logarithmic transformation term can be introduced to enhance the model's nonlinear fitting capability, such as:

[0082]

[0083] This feature can be combined with other features as an interaction term (such as confidence score × LogMerge) to enhance the discriminative power of the scoring model.

[0084] In this embodiment, to improve the reliable identification and engineering interpretability of the crack complex, a 10-dimensional feature vector is constructed based on the structural features, image features, and spatial information in the radar image where the identification box is located. Specific features include maximum confidence, average confidence, standard deviation, maximum difference in channel gradients, grayscale standard deviation, maximum difference in amplitude, mean channel fluctuation, number of fusions, and their constructed combination features and interaction terms (such as confidence × LogMerge). The feature calculation process follows the aforementioned formula.

[0085] S4. Perform credibility scoring: Input the multidimensional feature vector into a pre-trained scoring model to generate a credibility score for each crack complex.

[0086] To accurately assess and effectively screen the reliability of crack complex identification results, this embodiment constructs a nonlinear scoring model based on XGBoost and introduces the SHAP (SHapley Additive exPlanations) method to improve the model's interpretability. During the scoring model training phase, each crack complex is first manually scored, and scoring labels are assigned. These correspond to levels ranging from extremely untrustworthy to extremely trustworthy. Then, a multi-dimensional feature vector is extracted for each complex. The main parameters include the mean of the maximum gradient difference in the channel direction, the mean of the channel amplitude difference, the energy difference ratio, the gray-scale standard deviation ratio, the image sharpness index (Laplacian variance), the confidence statistics of the YOLO bounding boxes (such as the maximum value and the mean), and the number of times the bounding boxes of the crack complex are merged, and its logarithmic transformation value.

[0087] Based on the above characteristics, an XGBoost model is constructed, with the objective function being:

[0088]

[0089] Where 𝐹 is the CART (Classification and Regression Tree) function space, and 𝐾 is the number of trees. The model is trained by minimizing the objective loss function.

[0090] During the model application phase, for each new crack complex, the input feature vector is 𝑥, and the XGBoost model outputs a predicted score. To enhance the credibility and interpretability of the scoring results, this invention further introduces the SHAP method, which generates the marginal contribution value of each feature to the current scoring result through a game theory-based feature allocation approach. ,Right now:

[0091]

[0092] in The baseline score (such as the average score of all samples). The SHAP contribution value for the nth feature.

[0093] This explanation mechanism can achieve the following functions: output feature importance ranking for optimizing feature selection strategies; explain the scoring source of each identification result to assist expert review; and generate a crack risk map to achieve regional-level quantifiable ranking of credibility. This scoring mechanism solves the problem of the inability to quantitatively assess the "credibility" of crack identification in existing methods, ensuring the scientific rigor, transparency, and practicality of the post-processing stage, and significantly enhancing the engineering deployment capability of the automatic identification system.

[0094] As an application example of this step, the dataset construction and training configuration selected in this embodiment is as follows:

[0095] ① Number of original images: 2000 radar images, sourced from actual engineering acquisitions.

[0096] ② Image enhancement expanded to: 12,000 images

[0097] ③ Data partitioning: The ratio of training set, validation set, and test set is 8:1:1.

[0098] ④ Number of training epochs: 300 epochs, BatchSize: 1

[0099] ⑤ Tag source: Average of scores from 3 experienced experts.

[0100] After the XGBoost model is trained, the SHAP method is used to evaluate the marginal contribution of each feature to the score output. Based on the feature contribution ranking, features such as "maximum amplitude difference", "inter-channel gradient difference", and "gray-level variance" have high importance in the model and constitute the core basis for confidence determination.

[0101] S5. Screening and Output: Based on the credibility score, the crack complexes are screened, and crack complexes with scores higher than a preset threshold are retained as the final identification results and output in a structured manner.

[0102] After completing the scoring calculation of the crack complex, this invention further constructs a structured output module to record information of high-confidence identification results in a standardized form for subsequent visualization, risk ranking, and maintenance decision support. The system output fields include, but are not limited to, the following: spatial location information (channel number range, corridor range, point range), statistical indicators of YOLO identification confidence (such as maximum value, average value, and standard deviation), complex score, and ranking of the dominant scoring features and their corresponding SHAP contribution values. In addition, it outputs the feature vector of the complex and its top-k explanatory weights, the number of fused boxes (i.e., the number of fused boxes) and their marginal contribution to the scoring results, as well as other extended field information, such as the number of slices, the maximum box area, and structural continuity indicators.

[0103] The final output can be exported in structured formats such as CSV, JSON, or XML, or transmitted to external platforms such as road maintenance platforms, AI annotation systems, and disease databases via system interfaces. Simultaneously, the system supports the automatic generation of visualization results, including: crack complex outline overlaid with slice diagrams, complex scoring heatmaps, and SHAP interpretation diagrams (such as force plots and bar plots).

[0104] In addition to the steps described above, after scoring the fracture complex, this invention uses the predicted score... Set a confidence threshold λ (e.g., λ=4) to filter and classify the identification results. If a crack complex is deemed a credible target, it is retained; otherwise, it is discarded. This screening mechanism effectively removes high-confidence but low-reliability false positives such as bridge deck undulations, radar interference, and repair area residues, thereby improving overall recognition accuracy and engineering applicability. For example, different scoring thresholds (e.g., ≥90, ≥80, ≥70) are set to screen complexes, and the model performance is verified in real labeled data of the Lianyungang-Xuzhou Expressway (a total of 2662 real identifiable cracks). Table 1. Evaluation results of model recognition performance under different scoring thresholds.

[0105] Scoring threshold Precision Recall Number of predicted positive examples TP quantity Total number of true cases ≥90 97.52% 14.76% 403 393 2662 ≥80 91.80% 34.90% 1012 929 2662 ≥70 85.25% 55.60% 1736 1480 2662

[0106] The model scoring mechanism significantly improves recognition accuracy and can flexibly adjust the trade-off between Precision and Recall to meet the screening needs of crack recognition in different engineering applications.

[0107] It should be noted that the above description is merely one specific embodiment of the present invention, intended to help those skilled in the art better understand the present invention, but is not intended to limit the scope of protection of the present invention. Any non-substantial modifications, equivalent substitutions, or improvements made to the above embodiments within the spirit and principles of the present invention and based on the core ideas of the present invention should be included within the scope of protection of the present invention.

Claims

1. A post-processing method for identifying internal road cracks based on multi-feature scoring and region fusion, characterized in that, The method includes the following steps: S1. Obtain initial crack identification results: Based on the target detection model, process the slice image generated by the ground penetrating radar data to obtain multiple initial crack candidate boxes containing location information and confidence scores; S2. Constructing a crack complex: Spatial region fusion is performed on the multiple initial crack candidate boxes to construct a structurally continuous crack complex; the fusion is based on preset conditions of channel number continuity, channel range similarity, and overlapping or adjacent point intervals; S3. Extracting multidimensional features: For each of the crack complexes, extract a multidimensional feature vector to characterize its credibility; S4. Perform credibility scoring: Input the multidimensional feature vector into a pre-trained scoring model to generate a credibility score for each crack complex; S5. Screening and Output: Based on the credibility score, the crack complexes are screened, and crack complexes with scores higher than a preset threshold are retained as the final identification results and output in a structured manner.

2. The method according to claim 1, characterized in that: The spatial region fusion in step S2 specifically includes determining whether any two initial crack candidate boxes belong to the same crack complex based on one or more of the following conditions: The channel number ranges of the two candidate boxes are adjacent or the interval is within the preset channel number interval threshold; The Intersection over Union (IoU) of two candidate boxes in the channel dimension is greater than the preset channel IoU threshold; Two candidate boxes may overlap in the radar depth point dimension or their distance may be less than a preset point interval threshold.

3. The method according to claim 1 or 2, characterized in that: The multidimensional feature vector extracted in step S3 includes at least one of the following types of features: Radar signal characteristics, including at least one of amplitude difference, gradient mean, and energy difference ratio in the channel direction; Image spatial features, including at least one of the following: grayscale standard deviation ratio and image sharpness index; Statistical characteristics of the initial detection results, including the maximum, average, or standard deviation of the YOLO confidence scores of all candidate boxes that make up the complex; Structural fusion features include the number of candidate boxes that make up the composite, i.e., the number of fusion attempts.

4. The method according to claim 3, characterized in that: The amplitude difference in the channel direction is specifically the average of the difference between the maximum amplitude in the crack region and the maximum amplitude in the adjacent non-crack region. The image sharpness index is specifically the image sharpness of the crack region calculated using the Laplacian operator.

5. The method according to claim 3, characterized in that: The structural fusion feature also includes a logarithmic transformation term of the number of fusion frames, or an interaction term between the number of fusion frames and other features.

6. The method according to claim 1, characterized in that: The scoring model in step S4 is a nonlinear scoring model constructed based on the XGBoost algorithm.

7. The method according to claim 6, characterized in that, The scoring model uses the credibility level of expert annotations on the crack complex as training labels.

8. The method according to claim 1 or 6, characterized in that: The method also includes an interpretation step, specifically: The SHAP method is introduced to calculate the marginal contribution value of each feature in the multidimensional feature vector to the credibility score result, so as to realize the interpretability analysis of the score result.

9. The method according to claim 8, characterized in that: The structured output in step S5 also includes the scoring dominant features of each fracture complex and its SHAP contribution ranking.

Images