A highway subgrade damage type AI identification classification method, system and device

Abstract

This application provides an AI-based method, system, and device for identifying and classifying roadbed damage types. Applied to the field of artificial intelligence technology, it acquires road surface images, extracts micro-texture features, identifies abnormal areas, performs unsupervised clustering, and evaluates and confirms new damage types. This enables the automatic discovery of unknown damage and has the ability to automatically identify unknown damage types. It can also promptly detect specific damage in road sections with new materials, avoiding maintenance delays and increased costs.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an AI-based method, system, and device for identifying and classifying roadbed damage types. Background Technology

[0002] Ensuring the integrity and safety of the roadbed structure is crucial in the routine maintenance of highway infrastructure. With the rapid development of artificial intelligence technology, automated systems have made significant progress in the identification and classification of highway roadbed damage, greatly improving inspection efficiency and accuracy. These systems typically utilize high-resolution image acquisition equipment to continuously acquire visual information about the road surface and roadbed edges in real time, and employ technologies such as deep learning to effectively identify and classify known and common types of highway roadbed damage. This provides maintenance departments with standardized damage reports, enabling data-driven maintenance decisions and significantly reducing reliance on human experience.

[0003] However, with the application of new materials, such as new recycled asphalt mixtures with a high proportion of recycled asphalt pavement materials, highway pavements may exhibit specific damage that is difficult for traditional systems to identify after long-term service. Specifically, road sections paved with new recycled asphalt mixtures begin to show a damage pattern that has never appeared in the past learning data of existing intelligent recognition programs. This damage manifests as an unevenly distributed, hair-like network of microcracks, with extremely small widths, typically less than 0.1 mm, making it difficult to directly identify with the naked eye from a high-speed inspection vehicle. Its texture characteristics are significantly different from conventional fatigue network cracks. For example, conventional fatigue network cracks usually exhibit a relatively regular polygonal or irregular blocky structure with relatively clear crack edges; while this new microcrack network is more like randomly distributed, interwoven fine lines with blurred edges, and its visual appearance varies significantly under different lighting conditions.

[0004] Because this novel microcrack network differs significantly in visual characteristics from any known damage types learned by existing intelligent recognition programs during the learning phase, these programs cannot match it with any known damage patterns. Consequently, they are highly likely to misclassify it as normal asphalt texture or as visual interference caused by environmental factors such as uneven lighting or shadow variations. This systematic underreporting results in significant limitations for intelligent recognition systems in detecting damage on these new material road sections, making it impossible to provide accurate early warnings.

[0005] Because the intelligent recognition system failed to effectively warn of this early, minute damage, maintenance departments missed the optimal window for preventative maintenance. Subsequent rainwater erosion caused these microcracks to become channels for moisture penetration, accelerating the softening of the roadbed and reducing its load-bearing capacity. Simultaneously, continuous traffic loads repeatedly acted on these damaged areas, causing the microcrack network to rapidly expand and deepen, eventually developing into large, visible network cracks and aggregate spalling. This evolution of damage meant that a situation initially manageable with low-cost surface sealing or crack sealing eventually deteriorated to the point where costly milling and repaving were required for complete repair.

[0006] Faced with the challenge of specific damage caused by the application of new materials that existing intelligent identification systems cannot recognize, the traditional method of constantly creating labels for newly discovered damage types and retraining the program is too slow, and the cost of data labeling and program training is extremely high. Therefore, the industry urgently needs a more forward-looking intelligent identification and classification method. This method should have the ability to autonomously learn the characteristics of unknown or rare damage, and be able to proactively identify these new, subtle damages with unique physical mechanisms that appear in actual operation. This would ensure the timeliness and accuracy of highway maintenance decisions and avoid huge economic losses and safety hazards caused by information lag.

[0007] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0008] In view of the shortcomings of the prior art, this application provides an AI-based method, system and equipment for identifying and classifying roadbed damage types. It has the ability to automatically identify unknown damage types and can promptly detect specific damage in road sections with new materials, thus avoiding maintenance delays and increased costs.

[0009] Firstly, an AI-based method for identifying and classifying highway subgrade damage types, the method comprising the following steps:

[0010] S1: Acquire a road surface image, divide the road surface image into local regions, and extract micro-texture feature vectors from the local regions;

[0011] S2: Identify anomalous regions with potential unknown structures based on the microtexture feature vectors of all the local regions;

[0012] S3: Perform unsupervised clustering on the micro-texture feature vectors corresponding to the abnormal regions to form clusters;

[0013] S4: Evaluate the size and intra-cluster feature consistency of the clusters, and identify the clusters that simultaneously meet the size requirements and intra-cluster feature consistency requirements as newly discovered specific damage types;

[0014] S5: Generate a report of the newly discovered specific damage type.

[0015] Furthermore, step S2 includes:

[0016] S11: Acquire a road surface image and perform multi-scale local region division on the road surface image;

[0017] S12: Extract multi-scale, multi-directional texture features from the multi-scale local region;

[0018] S13: Combine the multi-scale, multi-directional texture features of the multi-scale local region to form the micro-texture feature vector.

[0019] Furthermore, step S2 includes:

[0020] S21: Compare the micro-texture feature vector with a normal road surface texture reference standard to identify areas that deviate significantly from the normal state;

[0021] S22: Input the micro-texture feature vector into a known damage classifier to determine whether it conforms to any known damage type;

[0022] S23: Regions that simultaneously deviate significantly from the normal state and do not conform to any known damage type are marked as the anomalous regions with potential unknown structure.

[0023] Furthermore, step S21 includes:

[0024] S211: Construct the feature space boundary of the normal road surface texture reference standard;

[0025] S212: Compare the micro-texture feature vector with the feature space boundary;

[0026] S213: Identify the region where the micro-texture feature vector is located outside the feature space boundary as a region that deviates significantly from the normal state.

[0027] Furthermore, step S3 includes:

[0028] S31: Calculate the Euclidean distance between each of the micro-texture feature vectors and other micro-texture feature vectors in the feature space; and calculate the local density of each micro-texture feature vector within a preset radius centered on it.

[0029] S32: If there exists a Euclidean distance between any two of the micro-texture feature vectors that is less than a preset distance threshold, and their respective local densities are greater than a preset density threshold, then construct the feature adjacency relationship of the micro-texture feature vectors.

[0030] S33: Based on the feature adjacency relationship, identify the connected set of the micro-texture feature vectors to form an initial cluster;

[0031] S34: Evaluate the geographical distribution of the initial clusters, and merge the initial clusters with continuous geographical distributions to form the clusters.

[0032] Furthermore, step S34 includes:

[0033] S341: Calculate the geographical envelope of the initial cluster;

[0034] S342: Determine the geographical continuity of the initial cluster based on the overlapping or proximity relationships of the geographical range envelope;

[0035] S343: Based on the result of the geographical location continuity judgment, merge the initial clusters with continuous geographical locations to form the cluster.

[0036] Furthermore, step S4 includes:

[0037] S41: Calculate the size index and intra-cluster characteristic consistency index of the clusters;

[0038] S42: Determine the initial judgment value of the cluster based on the scale index and the intra-cluster feature consistency index;

[0039] S43: When the initial determination value is within an uncertain range, obtain the image sequence of the region where the cluster is located;

[0040] S44: Analyze the feature changes of the clusters in the image sequence, and update the determination value of the clusters based on the feature changes;

[0041] S45: Compare the updated judgment value with the confirmation threshold to confirm the cluster as a newly discovered specific damage type.

[0042] Furthermore, step S43 includes:

[0043] S431: Calculate the feature differences of the clusters at adjacent time points in the image sequence;

[0044] S432: Accumulate the aforementioned feature differences to obtain the total change;

[0045] S433: Update the cluster determination value based on the total change.

[0046] Secondly, a system for AI-based identification and classification of highway subgrade damage types, characterized in that it operates the method described in any one of the above claims, the system comprising:

[0047] Acquisition module: Acquires road surface image, divides the road surface image into local regions, and extracts micro-texture feature vectors from the local regions;

[0048] Identification module: Based on the microtexture feature vectors of all the local regions, identify abnormal regions with potential unknown structures;

[0049] Clustering module: performs unsupervised clustering on the micro-texture feature vectors corresponding to the abnormal regions to form clusters;

[0050] Classification module: Evaluates the size and intra-cluster feature consistency of the clusters, and identifies the clusters that simultaneously meet the size and intra-cluster feature consistency requirements as newly discovered specific damage types;

[0051] Report generation module: Generates a report on the newly discovered specific damage type.

[0052] Thirdly, a method and device for AI-based identification and classification of roadbed damage types includes a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps in any of the methods described above are performed.

[0053] Beneficial effects: The method, system and equipment for identifying and classifying roadbed damage types proposed in this application can automatically discover unknown damage by acquiring road surface images, extracting micro-texture features, identifying abnormal areas, performing unsupervised clustering and evaluation to confirm new damage types. It has the ability to automatically identify unknown damage types and can promptly discover specific damage in road sections with new materials, thus avoiding maintenance delays and increased costs. Attached Figure Description

[0054] Figure 1 This is a flowchart of an AI-based method for identifying and classifying roadbed damage types proposed in this application.

[0055] Figure 2 This is a structural diagram of an AI-based classification system for identifying and classifying roadbed damage types proposed in this application.

[0056] Figure 3 This is a simplified schematic diagram of an AI-based method for identifying and classifying roadbed damage types proposed in this application.

[0057] Labeling Explanation: 201. Acquisition Module; 202. Identification Module; 203. Clustering Module; 204. Classification Module; 205. Report Generation Module. Detailed Implementation

[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and marked in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0059] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0060] Please refer to Figure 1 , Figure 3 A method for AI-based identification and classification of roadbed damage types, comprising the following steps:

[0061] S1: Obtain the road surface image, divide the road surface image into local regions, and extract the micro-texture feature vector from the local regions;

[0062] S2: Identify anomalous regions with potential unknown structures based on the micro-texture feature vectors of all local regions;

[0063] S3: Perform unsupervised clustering on the micro-texture feature vectors corresponding to the abnormal regions to form clusters;

[0064] S4: Evaluate the size and intra-cluster consistency of clusters, and identify clusters that simultaneously meet the size and intra-cluster consistency requirements as newly discovered specific damage types.

[0065] S5: Generate reports of newly discovered specific damage types.

[0066] The workflow of this method begins with data acquisition and preprocessing. High-resolution image acquisition equipment deployed on inspection vehicles continuously captures high-definition images of the road surface. Unlike traditional methods that treat the entire image as the unit of analysis, this method first subdivides each acquired road surface image into a series of non-overlapping or partially overlapping local regions, for example, dividing them into several image blocks of size 64 pixels by 64 pixels. The purpose of this is to shift the granularity of analysis from the macroscopic overall road surface to the microscopic local texture, because the novel microcrack network may not be noticeable at the macroscopic scale, but it will inevitably change the original texture distribution of the road surface at the microscopic scale. For each subdivided local region, the system extracts its microscopic texture feature vector. This feature vector is a set of values ​​used to mathematically describe the texture characteristics of the local region, which can quantitatively reflect information such as texture roughness, directionality, and periodicity. In this way, the raw and unstructured image data is transformed into structured numerical data that can be analyzed in depth by computers.

[0067] After obtaining a set of micro-texture feature vectors covering the entire road surface image, the next step is to identify anomalous regions with potential unknown structures. This step is the starting point for discovering new damage, and its core task is to filter out regions exhibiting distinctive characteristics from a massive amount of normal road surface texture. These regions are called anomalous because their micro-texture features do not conform to the patterns of normal, healthy road surfaces, and may not match any known damage types. This step defines the target range for subsequent in-depth analysis, avoids the enormous overhead of indiscriminate calculations on all road surface data, and allows the system to focus its attention on the regions most likely to contain new damage information.

[0068] After identifying anomalous regions, unsupervised clustering is performed on the microtexture feature vectors corresponding to these regions. Since the goal is to discover unknown damage types, supervised learning methods that require pre-provided labels cannot be used. Unsupervised clustering is an ideal solution, enabling grouping based solely on the intrinsic similarity of the data without prior knowledge. The system takes the microtexture feature vectors of all anomalous regions as input, calculates their relationships in the feature space, and groups vectors with similar features together to form clusters. Each cluster represents an anomalous pattern with shared microtexture features and is a candidate for a newly discovered specific damage type.

[0069] However, not all clusters represent meaningful new damage. Some clusters may simply consist of random image noise or isolated, unrepresentative outliers. To ensure the reliability of the findings, this method rigorously evaluates and confirms the clusters. The evaluation is based on two core dimensions: size and intra-cluster feature consistency. Size requirement means that a cluster must contain a sufficient number of local regions or cover a sufficiently large geographical area to exclude sporadic, accidental outliers and ensure that the discovered pattern has a certain degree of universality or continuity. Intra-cluster feature consistency requirement means that all local regions constituting the same cluster must have highly similar microtextural feature vectors. This ensures that the cluster represents a stable and uniform damage pattern, rather than a hodgepodge of various different anomalies. Only when a cluster simultaneously meets both the preset size requirement and intra-cluster feature consistency requirement will the system ultimately confirm it as a newly discovered specific damage type.

[0070] Once a newly discovered specific type of damage is confirmed, a corresponding report is generated. This report records detailed information about the new damage type, such as its typical microtextural features, geographical distribution on the highway, and extent of impact, and is then sent to highway maintenance personnel. This report plays a crucial early warning role, compensating for the inability of existing systems to identify such damage. It enables maintenance departments to be aware of new problems on the road surface as soon as possible and to develop targeted investigation and maintenance plans accordingly, thereby effectively avoiding serious consequences caused by information delays.

[0071] Furthermore, step S2 includes:

[0072] S11: Acquire road surface image and perform multi-scale local region segmentation on the road surface image;

[0073] S12: Extract multi-scale, multi-directional texture features from multi-scale local regions;

[0074] S13: Combine multi-scale, multi-directional texture features of multi-scale local regions to form a micro-texture feature vector.

[0075] Because damage at different scales presents different visual characteristics in images, a tiny crack might appear as a sharp line at very high resolution, but only as a slightly blurred area at lower resolution. Therefore, employing multi-scale local region segmentation means that the system analyzes images not only at a single, fixed size, but also at multiple different size levels. For example, the system can simultaneously segment and analyze road surface images at three scales: 32 pixels by 32 pixels, 64 pixels by 64 pixels, and 128 pixels by 128 pixels. This multi-level perspective ensures that regardless of the physical size of the damage, there is always an analysis scale that can most effectively capture its features, thus avoiding feature omissions that might occur due to single-scale analysis.

[0076] At each scale, the system further extracts multi-scale, multi-directional texture features. Road surface damage, especially cracks, exhibits a clear directionality. To comprehensively describe this characteristic, the system employs tools such as Gabor filter banks. Gabor filters can be configured to be particularly sensitive to textures in specific directions and frequencies. By applying a set of Gabor filters covering multiple directions, such as 0 degrees, 45 degrees, 90 degrees, and 135 degrees, and multiple frequencies, the system can extract a series of response values ​​from each local region. These response values ​​collectively constitute a comprehensive description of the texture of that region.

[0077] Finally, the system combines the texture features extracted from all scales and directions to form a higher-dimensional, more information-rich micro-texture feature vector. This final feature vector is like a detailed identity file for a local area, recording not only its texture information at a specific scale and direction, but also the comprehensive representation of this information across different scales and directions. This comprehensive feature representation greatly enhances the ability to identify complex and subtle damage patterns, providing a solid data foundation for subsequent anomaly area identification, enabling more accurate differentiation between normal road surface textures and novel, specific damage.

[0078] Furthermore, step S2 includes:

[0079] S21: Compare the micro-texture feature vector with a normal road surface texture reference standard to identify areas that deviate significantly from the normal state;

[0080] S22: Input the micro-texture feature vector into the known damage classifier to determine whether it conforms to any known damage type;

[0081] S23: Regions that simultaneously deviate significantly from the normal state and do not conform to any known damage type are marked as potentially unstructured anomalous regions.

[0082] The system internally constructs a texture model representing a healthy, intact road surface, which is learned from a large number of normal road surface images. When a microscopic texture feature vector of a new local region is input, the system compares it with this normal standard. If the difference between the feature vector and the normal standard exceeds a preset threshold, the region is initially marked as significantly deviating from the normal state. The purpose of this step is to filter out all potentially problematic regions, whether known or unknown damage, from the large number of normal regions, thereby avoiding the fundamental error of misclassifying novel damage as normal road surface texture.

[0083] The known damage classifier is trained using traditional supervised learning methods and internally stores feature patterns of all known and common roadbed damage types, such as longitudinal cracks, transverse cracks, alligator cracks, and potholes. The system inputs the feature vectors of areas deviating from the normal state into this classifier. The classifier then determines whether the feature vector matches any known damage pattern. If a match is found, it indicates that the damage in that area is a type already recognized by the system and does not belong to the target of this discovery task, therefore it is excluded.

[0084] Ultimately, only regions that simultaneously meet two conditions—significantly deviating from the normal state and not conforming to any known damage type—are ultimately marked by the system as anomalous regions with potential unknown structures. This rigorous dual-screening mechanism acts like a sophisticated filter, effectively eliminating interference from normal and known damage regions. This ensures that the final set of anomalous regions represents a truly high-value candidate set containing novel, unknown damage patterns. This provides clean and targeted input data for subsequent unsupervised clustering analysis, directly addressing the technical challenge of existing systems' inability to distinguish novel damage from the background.

[0085] Furthermore, step S21 includes:

[0086] S211: Construct the feature space boundary of a normal road surface texture reference standard;

[0087] S212: Compare the micro-texture feature vector with the feature space boundary;

[0088] S213: Identify regions where the micro-texture feature vectors are located outside the feature space boundary as regions that deviate significantly from the normal state.

[0089] To make the comparison with normal road surface texture reference standards more objective and accurate, a discrimination method based on feature space boundaries can be adopted. Here, feature space is an abstract multi-dimensional mathematical space whose dimensions are equal to the dimensions of the micro-texture feature vectors. Each local region's micro-texture feature vector can be considered a point in this space. To construct the feature space boundary of normal road surface texture, the system first collects a large number of images of normal road surfaces confirmed to be healthy and undamaged, and extracts their micro-texture feature vectors. These points representing normal road surfaces will cluster in the feature space to form one or more high-density regions, forming a specific distribution pattern. The system will use techniques such as single-class support vector machines or Gaussian mixture models to learn the contour of this distribution, thereby constructing a boundary that can encompass the vast majority of normal sample points.

[0090] This boundary becomes the quantitative standard for judging whether a region is normal. When the micro-texture feature vector of a new local region is extracted, the system locates it in this feature space. Then, the system determines whether this point is located inside or outside the constructed boundary.

[0091] If the point is within the boundary, it indicates that its microscopic texture features are highly similar to known normal road surface samples, and therefore the area is classified as normal. Conversely, if the point falls outside the boundary, it means that its texture features differ significantly from all known normal patterns. The system will then identify this area as significantly deviating from the normal state. This discrimination method based on feature space boundaries transforms the originally vague similarity comparison into a clear and quantifiable geometric location judgment, greatly improving the accuracy and reliability of abnormal area identification and laying a solid foundation for subsequent discovery of unknown damage.

[0092] Furthermore, step S3 includes:

[0093] S31: Calculate the Euclidean distance between each micro-texture feature vector and other micro-texture feature vectors in the feature space; and calculate the local density of each micro-texture feature vector within a preset radius centered on it.

[0094] S32: If there exists any two micro-texture feature vectors whose Euclidean distance is less than a preset distance threshold and whose local density is greater than a preset density threshold, then construct the feature adjacency relationship of the micro-texture feature vectors.

[0095] S33: Based on the adjacency relationship of features, identify the connected sets of micro-texture feature vectors to form an initial cluster;

[0096] S34: Evaluate the geographical distribution of the initial clusters and merge the initial clusters with continuous geographical distribution to form a new cluster.

[0097] Euclidean distance is a direct indicator of the similarity between two feature vectors in the feature space. A smaller distance indicates greater similarity in texture between the two local regions. Euclidean distance is calculated by summing the squares of the differences between the two vectors in each dimension and then taking the square root. Local density, on the other hand, refers to the number of other feature vectors contained within a specific radius around a feature vector. A higher local density indicates a more crowded region in the feature space, making it more likely to be a core part of a cluster.

[0098] The conditions for establishing adjacency relationships are twofold: for any two feature vectors, their Euclidean distance must be less than a preset distance threshold, and their respective local densities must be greater than a preset density threshold. The distance threshold ensures that the connected vectors are similar in features, while the density threshold effectively filters out noise points located in sparse regions, ensuring that connections only occur in high-density core regions of the feature space. This dual constraint based on density and distance makes the clustering process highly adaptable to irregularly shaped clusters, making it very suitable for identifying crack networks with varied morphologies.

[0099] Once the feature adjacency relationships are fully established, the system can use connected component search methods in graph theory to find all interconnected feature vector sets. Each such connected set constitutes an initial cluster. These initial clusters represent groups of anomalous regions that are tightly clustered in the feature space and have similar texture features.

[0100] However, simply clustering in the feature space is insufficient. Road surface damage is a continuous entity in the physical world. Due to image noise or subtle variations in texture, a physically continuous damage region may have its feature vectors segmented into multiple neighboring initial clusters in the feature space. To address this issue, this method evaluates and merges geographically contiguous initial clusters. It examines the actual geographic coordinates of the local regions contained in each initial cluster on the road. If two or more distinct initial clusters are found to be geographically closely connected or partially overlapping, the system merges them into a larger, final cluster. This step combines the abstract feature space analysis results with the real physical spatial distribution, ensuring that the final clusters fully and accurately reflect the actual extent and morphology of a continuous damage region.

[0101] To make the process of merging initial clusters with continuous geographical locations more concrete and operable, a method based on geographical extent envelope can be used. Specifically, step S34 includes:

[0102] S341: Calculate the geographic extent envelope of the initial clusters;

[0103] S342: Determine the geographical continuity of the initial clusters based on the overlap or proximity of the geographical envelope;

[0104] S343: Based on the results of the geographical continuity judgment, merge the initial clusters with continuous geographical locations to form a new cluster.

[0105] Each initial cluster consists of a series of local regions, each corresponding to a specific geographic coordinate. The geographic envelope can be the smallest bounding rectangle enclosing the coordinates of all local regions of the cluster, or a more compact convex hull polygon. This envelope, in the geographic information system, indicates the approximate area and location of the initial cluster on the road surface.

[0106] After obtaining the geographic envelopes of all initial clusters, all envelope pairs are further traversed to check for overlap or proximity. Overlap refers to two envelopes sharing a common geometric region, indicating that the damaged areas they represent are physically directly connected. Proximity refers to two envelopes that, while not directly overlapping, have a shortest distance less than a preset geographic proximity threshold. This threshold is set based on the GPS accuracy of the inspected images, typically 1 meter, but if the GPS error is greater than 0.5 meters, the threshold is set to twice the GPS error. This indicates that although the damaged areas they represent may have slight physical discontinuities, they are still very likely to belong to the same overall disease.

[0107] Finally, all initial clusters identified as having overlapping or proximity relationships are merged into a single set. All initial clusters in this set are then combined into a single, larger final cluster. In this way, damage patterns that might have been fragmented in the feature space are reintegrated in the physical space dimension, forming a complete clustering result that accurately reflects the actual damage extent. This ensures that the system's identification of novel damage patterns is not only accurate in terms of features but also conforms to the actual spatial morphology.

[0108] Furthermore, step S4 includes:

[0109] S41: Calculate the size index of the cluster and the consistency index of intra-cluster characteristics;

[0110] S42: Determine the initial judgment value of the cluster based on the size index and the intra-cluster characteristic consistency index;

[0111] S43: When the initial judgment value is in an uncertain range, obtain the image sequence of the region where the cluster is located;

[0112] S44: Analyze the feature changes of clusters in the image sequence, and update the cluster determination value based on the feature changes;

[0113] S45: Compare the updated decision value with the confirmation threshold to confirm the cluster as a newly discovered specific damage type.

[0114] The initial judgment value is a quantitative score derived by combining the size index and the intra-cluster feature consistency index. This score aims to assess whether a cluster meets the preliminary criteria for a newly discovered specific damage type. For example, a function can be set up that takes the size index and the intra-cluster feature consistency index as input and outputs an initial judgment value between 0 and 1, where a higher value indicates a greater likelihood of a new damage type. In a specific embodiment, the normalized size index and the intra-cluster consistency index can be directly substituted into the geometric mean-hyperbolic tangent function: the square root of their product is calculated, and then compressed using tanh, resulting in an initial judgment value between 0 and 1. This function is entirely data-driven and sensitive to bottlenecks, ensuring that the judgment value rapidly approaches zero when either the size index or the intra-cluster consistency index is low, and smoothly approaches one when the index is high, thus satisfying the automatic quantification of the dual thresholds of size index and intra-cluster consistency.

[0115] For clusters with very high or very low initial judgment values, a direct judgment can be made. However, for clusters whose initial judgment values ​​fall within a preset uncertainty range, this method will not make a hasty conclusion. For example, if the uncertainty range is 0.4-0.7 for the initial judgment value, values ​​below 0.4 are directly excluded, and values ​​above 0.7 are directly confirmed. In this case, the method will record the precise geographical location of this uncertain cluster and retrieve an image sequence of that location from a historical database or continuously acquire it in future inspection tasks. An image sequence refers to a series of images taken of the same road surface area at different points in time, such as over several weeks or months.

[0116] After acquiring the image sequence, the changes in the cluster's features over time are analyzed. Real road surface damage, especially cracks under load and environmental influences, typically exhibits a progressive development trend, such as crack lengthening, widening, or an increase in the density of the microcrack network. Random noise or stains, on the other hand, usually show random fluctuations or remain constant at different time points. This method also tracks the changes in the microtexture feature vector of the cluster over time and updates its judgment value based on this trend. If the analysis shows that the cluster's features exhibit a consistently deteriorating pattern, its judgment value will be significantly increased. Conversely, if the features are stable or show irregular changes, the judgment value will be decreased.

[0117] Finally, the updated decision value, after dynamic evolution analysis, incorporates richer and more reliable temporal dimension information. This method compares this updated decision value with a final confirmation threshold. Only when the updated decision value exceeds this high-confidence threshold is the cluster officially confirmed as a newly discovered specific lesion type. This dynamic-static combined confirmation mechanism significantly improves the accuracy of identifying early, subtle, and slowly developing novel lesions, effectively avoiding false positives and false negatives.

[0118] Furthermore, step S43 includes:

[0119] S431: Calculate the feature differences of clusters at adjacent time points in an image sequence;

[0120] S432: Accumulate feature differences to obtain the total change;

[0121] S433: Update the cluster determination value based on the total change.

[0122] Specifically, once an image sequence of an uncertain cluster is obtained, such as inspection images from four consecutive weeks, the micro-texture feature vectors of that cluster region are first extracted from the images of each week. Then, the feature differences between adjacent time points are calculated. For example, the difference in feature vectors between the first and second weeks, the second and third weeks, and the third and fourth weeks are calculated. This difference can be obtained by calculating the Euclidean distance or other distance metrics between the feature vectors of two time points; its magnitude reflects the drastic change in damage features within that time period.

[0123] Next, these feature differences calculated over consecutive time periods are accumulated to obtain a total change. This total change reflects the overall trend and magnitude of the cluster feature evolution throughout the entire observation period. For example, if the feature differences calculated each week are all significantly positive, the accumulated total change will be large, strongly indicating that the damage in the region is continuing to develop.

[0124] Finally, the cluster's decision value is updated based on this quantified total change. A function can be defined to map the total change to an adjustment in the decision value. A large, consistently positive total change corresponds to a large decision value gain, thus pushing the cluster's decision value above the confirmation threshold. Conversely, a total change close to zero or with random fluctuations may lead to a lower decision value. In this way, the system transforms the intuitive judgment of damage evolution into an objective and repeatable calculation process, making the final confirmation of novel damage more scientific and reliable.

[0125] Please refer to Figure 2 A system for AI-based identification and classification of highway subgrade damage types, characterized in that it operates as described in any of the above methods, and the system includes:

[0126] Acquisition Module 201: Acquires road surface images, divides the road surface images into local regions, and extracts micro-texture feature vectors from the local regions;

[0127] Identification module 202: Identifies abnormal regions with potential unknown structures based on the micro-texture feature vectors of all local regions;

[0128] Clustering module 203: Performs unsupervised clustering on the micro-texture feature vectors corresponding to the abnormal regions to form clusters;

[0129] Classification module 204: Evaluate the size and intra-cluster feature consistency of clusters, and identify clusters that simultaneously meet the size and intra-cluster feature consistency requirements as newly discovered specific damage types;

[0130] Report generation module 205: Generates reports on newly discovered specific damage types.

[0131] In this system, the acquisition module 201 is the starting point of the data processing flow, responsible for transforming the raw road surface images collected by inspection vehicles into structured micro-texture feature data, providing high-quality input for subsequent analysis. The identification module 202 undertakes the first-stage screening task, accurately locating potential unknown damage areas by comparing them with normal standards and a known damage database. The clustering module 203 plays a core role in unsupervised learning, organizing scattered outliers into pattern clusters with inherent consistency, a crucial step in discovering new damage types. The classification module 204 plays the role of quality control and final decision-making, ensuring that only true and meaningful clusters are identified as new damage through a rigorous evaluation and confirmation mechanism, guaranteeing the system's reliability. Finally, the report generation module 205, as the system's output, presents the findings to the user in a clear and detailed report, realizing the transformation from data to information and providing direct support for highway maintenance decisions.

[0132] Furthermore, this application also provides a device for an artificial intelligence-based identification and classification method for highway subgrade damage types. This device includes a processor and a memory. The memory stores computer-readable instructions, and when the processor executes these instructions, it runs the steps of any of the aforementioned methods. This device provides a physical platform for the implementation of the aforementioned methods; it can be a dedicated computing unit integrated into a highway inspection vehicle or a server deployed in a data center. The processor is responsible for executing all computational tasks within the method, while the memory stores the execution instructions and data generated during processing. Through the combination of hardware and software, this device can automatically and efficiently complete the autonomous discovery and classification of unknown subgrade damage, providing a powerful intelligent tool for the highway maintenance industry.

[0133] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0134] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for AI-based identification and classification of highway subgrade damage types, characterized in that, The method includes the following steps: S1: Acquire a road surface image, divide the road surface image into local regions, and extract micro-texture feature vectors from the local regions; S2: Identify anomalous regions with potential unknown structures based on the microtexture feature vectors of all the local regions; S3: Perform unsupervised clustering on the micro-texture feature vectors corresponding to the abnormal regions to form clusters; Step S3 includes: S31: Calculate the Euclidean distance between each of the micro-texture feature vectors and other micro-texture feature vectors in the feature space; and calculate the local density of each of the micro-texture feature vectors within a preset radius centered on it. S32: If there exists a Euclidean distance between any two of the micro-texture feature vectors that is less than a preset distance threshold, and their respective local densities are greater than a preset density threshold, then construct the feature adjacency relationship of the micro-texture feature vectors. S33: Based on the feature adjacency relationship, identify the connected set of the micro-texture feature vectors to form an initial cluster; S34: Evaluate the geographical distribution of the initial clusters, and merge the initial clusters with continuous geographical distributions to form the clusters; Step S34 includes: S341: Calculate the geographical envelope of the initial cluster; S342: Determine the geographical continuity of the initial cluster based on the overlapping or proximity relationships of the geographical range envelope; S343: Based on the result of the geographical location continuity judgment, merge the initial clusters with continuous geographical locations to form the cluster; S4: Evaluate the size and intra-cluster feature consistency of the clusters, and identify the clusters that simultaneously meet the size requirements and intra-cluster feature consistency requirements as newly discovered specific damage types; S5: Generate a report of the newly discovered specific damage type.

2. The method for AI-based identification and classification of highway subgrade damage types according to claim 1, characterized in that, Step S1 includes: S11: Acquire a road surface image and perform multi-scale local region division on the road surface image; S12: Extract multi-scale, multi-directional texture features from the multi-scale local region; S13: Combine the multi-scale, multi-directional texture features of the multi-scale local region to form the micro-texture feature vector.

3. The method for AI-based identification and classification of highway subgrade damage types according to claim 2, characterized in that, Step S2 includes: S21: Compare the micro-texture feature vector with a normal road surface texture reference standard to identify areas that deviate significantly from the normal state; S22: Input the micro-texture feature vector into a known damage classifier to determine whether it conforms to any known damage type; S23: Regions that simultaneously deviate significantly from the normal state and do not conform to any known damage type are marked as anomalous regions with potential unknown structures.

4. The AI-based identification and classification method for highway subgrade damage types according to claim 3, characterized in that, Step S21 includes: S211: Construct the feature space boundary of the normal road surface texture reference standard; S212: Compare the micro-texture feature vector with the feature space boundary; S213: Identify the region where the micro-texture feature vector is located outside the feature space boundary as a region that deviates significantly from the normal state.

5. The AI-based identification and classification method for highway subgrade damage types according to claim 1, characterized in that, Step S4 includes: S41: Calculate the size index and intra-cluster characteristic consistency index of the clusters; S42: Determine the initial judgment value of the cluster based on the scale index and the intra-cluster feature consistency index; S43: When the initial determination value is within an uncertain range, obtain the image sequence of the region where the cluster is located; S44: Analyze the feature changes of the clusters in the image sequence, and update the determination value of the clusters based on the feature changes; S45: Compare the updated judgment value with the confirmation threshold to confirm the cluster as a newly discovered specific damage type.

6. The AI-based method for identifying and classifying highway subgrade damage types according to claim 5, characterized in that, Step S43 includes: S431: Calculate the feature differences of the clusters at adjacent time points in the image sequence; S432: Accumulate the aforementioned feature differences to obtain the total change; S433: Update the cluster determination value based on the total change.

7. A method system for AI-based identification and classification of highway subgrade damage types, characterized in that, The system comprising: (1) the method of any one of claims 1-6; Acquisition module: Acquires road surface image, divides the road surface image into local regions, and extracts micro-texture feature vectors from the local regions; Identification module: Based on the microtexture feature vectors of all the local regions, identify abnormal regions with potential unknown structures; Clustering module: performs unsupervised clustering on the micro-texture feature vectors corresponding to the abnormal regions to form clusters; Classification module: Evaluates the size and intra-cluster feature consistency of the clusters, and identifies the clusters that simultaneously meet the size and intra-cluster feature consistency requirements as newly discovered specific damage types; Report generation module: Generates a report on the newly discovered specific damage type.

8. A method and device for AI-based identification and classification of highway subgrade damage types, characterized in that, It includes a processor and a memory, the memory storing computer-readable instructions that, when executed by the processor, perform the steps of the method as described in any one of claims 1-6.

Images