Efficient and intelligent road disease data labeling method based on large model cooperation

Abstract

The application discloses a large model cooperative road disease data efficient intelligent labeling method, and particularly relates to the technical field of road detection; the application extracts various core features from the disease ontology and the recognizable dimension, comprehensively covers the complexity of the disease itself and the complexity of the recognition difficulty, avoids the limitation of single feature, quantifies the two types of complexity through the disease evaluation coefficient and the disturbance evaluation coefficient, and then fuses them into a complex evaluation coefficient to complete the complexity determination according to the coefficient interval division level, so that the problem that the existing technology mainly adopts a fixed model full quantity calling mode, and the determination of the image complexity mainly depends on a single feature and ignores the cooperative influence of the disease ontology feature and the environmental interference feature is solved.

Description

Technical Field

[0001] This invention relates to the field of road inspection technology, and more specifically, to a method for efficient and intelligent annotation of road defect data using large-scale collaborative models. Background Technology

[0002] As a core component of the road maintenance and management system, road defect detection directly determines the scientific nature of maintenance decisions and the safety of road traffic. With the rapid increase in the mileage of highways and urban roads, the traditional defect detection model that relies on manual inspection can no longer meet the needs of large-scale and high-precision management. Manual inspection is not only inefficient, but also greatly affected by subjective experience. At the same time, in severe weather or complex road environments, there are blind spots and safety risks, resulting in persistently high maintenance costs.

[0003] To overcome the limitations of manual inspection, the industry has gradually introduced computer vision technology to achieve automated disease labeling. However, existing technical solutions still face the following shortcomings in practical applications:

[0004] Current annotation methods mostly adopt a fixed model full call mode, using heavyweight models regardless of the complexity of image diseases. This leads to a waste of resources for simple image annotation, while complex images are rendered unusable due to insufficient model capabilities, creating a contradiction between "using a large amount of material for a small amount of material" and "using a small amount of material for a large amount of material". Furthermore, the determination of image complexity often relies on a single feature, ignoring the synergistic influence of disease features and environmental interference features.

[0005] The existing annotation only outputs basic information such as "category + bounding box" of the disease, without linking it to key road operation attributes and the degree of disease damage. As a result, the annotation data cannot directly support the ranking of maintenance priorities, and maintenance units still need to manually rearrange the priorities, which prolongs the decision-making cycle and has a low level of intelligence.

[0006] To address this, a highly efficient and intelligent annotation method for road defect data based on large-scale model collaboration has been developed. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an efficient and intelligent annotation method for road defect data using large-scale model collaboration.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] Efficient and intelligent annotation methods for road defect data using large-scale collaborative models include:

[0010] Complexity determination: Three sets of complexity level labels are set for image data, and each set of labels corresponds to a pre-built model combination scheme. The complexity levels include simple, medium and complex. The basic features of the image data are extracted using the model combination scheme corresponding to the simple level, including disease ontological features and identifiability features. After determining the complexity evaluation coefficient of the image data to be labeled, the complexity level labels are generated.

[0011] Generate annotations: Based on the complexity level label, if the complexity level label of the image data is simple, the annotation result can be directly output using the model combination scheme corresponding to the simple level. If it is not simple, the corresponding model combination scheme is called to input the image data for annotation and then output the annotation result.

[0012] Priority Classification: After all image data is labeled, the image data is classified into priority queues based on the road features and actual damage characteristics corresponding to the image data; the priority queues include high priority, medium priority and low priority.

[0013] Specifically, the extraction of the disease ontological features from the image data includes:

[0014] The characteristics of the disease itself include the proportion of diseased areas, the number of disease types, and the density of disease distribution.

[0015] Calculate the percentage of the total number of pixels in all diseased areas to the total number of pixels in the image, and use this as the percentage of diseased areas.

[0016] Calculate the number of diseases per unit physical area as the disease distribution density;

[0017] The number of different types of road defects in the statistical images is used as the number of defect types.

[0018] Specifically, the identifiability features extracted from the image data are as follows:

[0019] Identifiability features include edge sharpness, background interference ratio, and interference similarity;

[0020] The diseased areas in the image data are cropped, and the diseased edges are extracted to generate an edge mask. For the gradient value of each edge pixel in the edge mask, those with a gradient value higher than the reference gradient value are selected as high gradient edge pixels. The ratio between high gradient edge pixels and total edge pixels is calculated as the edge sharpness.

[0021] Calculate the percentage of the total number of pixels in all interference regions of the image to the total number of pixels in the image, and use this as the background interference percentage.

[0022] The color space similarity between the diseased area and the interference area is calculated using the Euclidean distance algorithm, and this similarity is used as the interference similarity.

[0023] Specifically, the determination of complex evaluation coefficients for the image data to be labeled:

[0024] For the proportion of diseased areas, the number of disease types, and the distribution density of disease in image data, the disease assessment coefficient of image data is determined by weighted calculation logic in combination with the set reference standards.

[0025] For the edge sharpness, background interference ratio, and interference similarity of image data, the interference evaluation coefficient of the image data is determined by weighted calculation logic in combination with the set reference standard.

[0026] Based on the disease assessment coefficient and disturbance assessment coefficient of image data, a weighted calculation logic is used to determine the complex assessment coefficient.

[0027] Specifically, generating complexity level labels involves setting three sets of coefficient intervals corresponding to the complexity evaluation coefficient, with the three sets of coefficient intervals corresponding to the simple, medium, and complex levels, respectively.

[0028] Specifically, the road features of the image data include:

[0029] The road features include the road type and traffic flow level of the image data;

[0030] Set up a set of road impact scores corresponding to different road types;

[0031] The traffic flow data of the road area where the image data is located within a set time zone before the current time point is acquired, and the traffic flow level of the road area where the image data is located is determined, including high traffic flow, medium traffic flow and low traffic flow.

[0032] Construct traffic flow data intervals corresponding to different traffic flow levels, match the traffic flow data of the road area where the image data is located with the corresponding traffic flow data intervals, and determine the traffic flow level;

[0033] Different traffic flow levels are assigned a set of traffic flow impact scores; the road impact score and traffic flow impact score of the image data are summed to obtain the vehicle impact coefficient of the image data.

[0034] Specifically, the actual characteristics of the disease in the image data include:

[0035] Disease characteristics include the location, type, and area affected;

[0036] The location, type, and area of ​​diseases in statistical image data are determined. Disease areas belonging to the same disease type are accumulated and summed. Weighting coefficients are set for different disease types. The accumulated disease areas of different disease types are multiplied by the corresponding set weights, and then summed to obtain the disease influence coefficient of the image data.

[0037] Specifically, the classification of all image data into priority queues involves:

[0038] The degree coefficient of image data is determined by comprehensively processing the vehicle impact coefficient and the defect impact coefficient based on image data and the reference standards established for different road types.

[0039] The total number of statistical image data is divided into three equal parts by dividing by an integer. All image data are sorted from largest to smallest according to the degree coefficient. After sorting, the three equal parts of image data are extracted from left to right and used as the number of high-priority queues, medium-priority queues, and low-priority queues, respectively.

[0040] The technical effects and advantages of this invention are as follows:

[0041] (1) By extracting core features from the dual dimensions of disease ontology and identifiability, the complexity of the disease itself and the difficulty of identification are fully covered, avoiding the limitation of a single feature. The two types of complexity are quantified by disease evaluation coefficient and disturbance evaluation coefficient respectively, and then fused into a complexity evaluation coefficient. The complexity is determined by classifying the levels according to the coefficient range. This solves the problem that existing technologies often use a fixed model full call mode, and the determination of image complexity often depends on a single feature, ignoring the synergistic influence of disease ontology features and environmental interference features.

[0042] (2) By converting road features into vehicle impact coefficients and disease features into disease impact coefficients, and then combining the two to calculate the degree coefficient, the actual impact of diseases on road operation is quantified. The images are divided into high, medium and low priority queues in descending order of degree coefficients. This solves the problem that the labeled data in the existing technology cannot directly support the maintenance priority ranking, and the maintenance unit still needs to manually rearrange the priority, which prolongs the decision-making cycle. Attached Figure Description

[0043] Figure 1 This is a flowchart of the efficient and intelligent annotation method for road defect data using a large-scale collaborative model, as described in this invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] like Figure 1 As shown, an efficient and intelligent annotation method for road defect data using large-scale model collaboration includes:

[0046] S1: Obtain the road damage image data to be labeled, and use the preprocessed images as the final image data to be labeled;

[0047] Preprocessing includes image denoising, contrast enhancement, and distortion correction;

[0048] Image denoising

[0049] Noise type identification: Identify the type of noise in the image (such as Gaussian noise, salt and pepper noise, motion blur noise) by analyzing gray-level histograms.

[0050] Adaptive filtering

[0051] For Gaussian noise: a bilateral filtering algorithm is used to smooth the noise while preserving the edge information of the defects. The filtering radius is dynamically adjusted according to the noise intensity (range 1-5 pixels).

[0052] For salt and pepper noise: An improved median filter is used to remove outlier pixels by using a sliding window (3×3 to 7×7 pixels), with the window size adaptively changing with the noise density;

[0053] For motion blur: Wiener filtering is used for blind area blurring, and the direction of the blur kernel is estimated by image gradient detection (associated with the road direction) to restore clear texture in a targeted manner.

[0054] Contrast Enhancement

[0055] Illumination unevenness correction: The Retinex algorithm is used to decompose the illumination component and reflection component of the image, and the illumination component is adaptively adjusted to eliminate brightness deviations caused by tree shadows, backlighting, etc.

[0056] Local contrast optimization: For diseased areas (such as cracks and pits), limit contrast adaptive histogram equalization (CLAHE) is applied, and the clipLimit parameter (range 2.0-4.0) is set to enhance local details while avoiding the amplification of background noise;

[0057] Dynamic range compression: A tone mapping algorithm is used for high dynamic range (HDR) images to compress the brightness range to 8 bits (0-255), ensuring that disease features can be identified on conventional display devices.

[0058] Distortion correction

[0059] Camera intrinsic parameter correction: Based on the pre-calibrated camera intrinsic parameter matrix (including focal length, principal point coordinates, and distortion coefficients), radial distortion (k1, k2, k3) and tangential distortion (p1, p2) of the image are corrected using the Zhang Zhengyou calibration method;

[0060] Perspective distortion correction: For tilted images captured by vehicle-mounted cameras, a perspective transformation matrix is ​​established by detecting the vanishing point of the road surface to correct the tilted view image into an approximate orthophoto projection image, ensuring the accuracy of the measurement of the geometric dimensions of the defects.

[0061] Smooth seams: For panoramic images formed by stitching together multiple images, the Poisson fusion algorithm is used to eliminate brightness jumps and texture misalignments at the stitching edges, ensuring the continuity of the diseased area.

[0062] The quality of the preprocessed road defect image data was assessed to determine the final labeled image data:

[0063] The preprocessed color image is converted to grayscale using a 3×3 Laplacian operator (e.g., ...). Perform convolution operation on the grayscale image to obtain the edge enhancement image, and calculate the variance SK of the grayscale values ​​of all pixels in the enhancement image;

[0064] The higher the variance of grayscale values, the higher the image clarity.

[0065] An adaptive threshold segmentation method is used to separate suspected disease areas and background areas from the preprocessed image, based on the average gray value of the suspected disease areas. Standard deviation of the background area ;

[0066] According to the formula Calculate the signal-to-noise ratio (SNR);

[0067] The higher the signal-to-noise ratio, the more likely it is to be classified as a low-noise image;

[0068] A lightweight semantic segmentation model is used to perform preliminary disease segmentation on the preprocessed image to obtain a disease mask; in the mask, 1 represents diseased pixels and 0 represents non-diseased pixels.

[0069] The ratio of the total number of defective pixels in the statistical mask to the total number of pixels in the image is denoted as the defective area percentage (SL).

[0070] The lower the percentage of diseased area, the less complete the diseased area is, which may be due to the deviation of the collection range or the small size of the disease itself, resulting in insufficient annotation information.

[0071] The quality assessment logic is constructed using the gray value variance SK, signal-to-noise ratio SNR, and diseased area percentage SL: ; To label the quality evaluation coefficients; where These are preset passing gray value variance, standard signal-to-noise ratio, and the theoretical minimum proportion of diseased areas, set based on image quality requirements. These are preset weighting coefficients;

[0072] If the quality evaluation coefficient of the image data is lower than the preset quality threshold coefficient, the image data of the target road will be re-acquired; otherwise, it will be marked as a high-quality image and used as the final image data to be labeled.

[0073] S2: Set three sets of complexity level labels for image data, and each set of labels corresponds to a pre-built model combination scheme. The complexity levels include simple, medium and complex. Use the model combination scheme corresponding to the simple level to extract the basic features of the image data, including disease ontology features and identifiability features. After determining the complexity evaluation coefficient of the image data to be labeled, generate complexity level labels.

[0074] Simple-level standard model combinations: lightweight models (such as the MobileNet-SSD detection model) do not require heavyweight components such as large language models (LLM) and visual language models (VLM).

[0075] The standard model combination for medium-level applications: calls the combination of "visual detection model (such as YOLOv8) + lightweight segmentation model (such as U-NetTiny)" without needing to enable the text prompt generation and logic verification functions of LLM.

[0076] Complex-level standard model combination: trigger full model collaboration (LLM generates multi-disease text prompts → VLM combines prompts to achieve multi-target differentiation → segment large models (such as SAM) to generate refined masks → LLM verifies disease correlation relationships).

[0077] Specifically:

[0078] The characteristics of the disease itself include the proportion of diseased areas, the number of disease types, and the density of disease distribution.

[0079] Identifiability features include edge sharpness, background interference ratio, and interference similarity;

[0080] Calculate the percentage of the total number of pixels in all diseased areas to the total number of pixels in the image, and use this as the percentage of diseased areas.

[0081] Acquisition process:

[0082] A lightweight image classification model is used to generate a disease mask (such as MobileSeg output, where "1" represents diseased pixels and "0" represents non-diseased pixels).

[0083] Pixel statistics: Count the total number of pixels with a value of "1" in the mask using a matrix traversal algorithm;

[0084] Calculate the percentage: Read the image resolution, calculate the total number of pixels in the image, and calculate the ratio between the total number of pixels marked "1" and the total number of pixels in the image to obtain the percentage of the diseased area.

[0085] The number of different types of road defects in the statistical images is used as the number of defect types.

[0086] Acquisition process:

[0087] A lightweight image classification model is adopted and pre-trained with a road disease dataset to ensure that the model can distinguish the core disease types;

[0088] Type prediction: Input the preprocessed high-quality image into the classification model, and the model outputs a "type-confidence" list (e.g., "horizontal crack: 0.95, pit: 0.03, crazing: 0.02").

[0089] Threshold screening: Set a confidence threshold (e.g., 0.5) to screen out disease types with a confidence level ≥ 0.5 (excluding false positives due to low confidence).

[0090] Deduplication statistics: Remove duplicates from the filtered types and count the final number of types.

[0091] Calculate the number of diseases per unit physical area as the disease distribution density;

[0092] Acquisition process:

[0093] Disease Counting: A lightweight target detection model is used to detect individual disease entities in the image (such as "3 independent cracks and 2 isolated pits") and output the number of diseases.

[0094] Area conversion: Calculate the physical area corresponding to the image;

[0095] Density calculation: The disease distribution density is obtained by dividing the number of diseases by the physical area.

[0096] For the proportion of diseased areas, the number of disease types, and the distribution density of disease in image data, the disease assessment coefficient of image data is determined by weighted calculation logic in combination with the set reference standards.

[0097] Specifically:

[0098] Calculate the disease assessment coefficient ,in These represent the percentage of diseased areas, the number of disease types, and the density of disease distribution, respectively. The reference values ​​set for the proportion of diseased areas, the number of disease types, and the density of disease distribution are based on the image complexity evaluation criteria. These are the preset weighting coefficients.

[0099] The diseased areas in the image data are cropped, and the diseased edges are extracted to generate an edge mask. For the gradient value of each edge pixel in the edge mask, those with a gradient value higher than the reference gradient value are selected as high gradient edge pixels. The ratio between high gradient edge pixels and total edge pixels is calculated as the edge sharpness.

[0100] Acquisition process:

[0101] Edge extraction: For the preprocessed image, only the diseased area is cropped (based on the disease mask, excluding background interference). The Canny edge detection algorithm (threshold: low threshold 50, high threshold 150) is used to extract the disease edges and generate an edge mask ("1" represents edge pixels, "0" represents non-edge pixels).

[0102] Gradient filtering: Calculate the gradient value of each "1" pixel in the edge mask (using the Sobel operator to calculate the square root of the sum of the squares of the gradients in the x and y directions), and filter out "high gradient edge pixels" with gradient values ​​≥ reference gradient value (e.g., 80) (the higher the gradient, the clearer the edge).

[0103] Ratio calculation: The ratio of the number of high-gradient edge pixels to the total number of edge pixels is calculated.

[0104] Calculate the percentage of the total number of pixels in all interfering regions of the image to the total number of pixels in the image, and use this as the background interference percentage; interfering regions include, for example, shadows, oil stains, fallen leaves, and road markings.

[0105] Acquisition process:

[0106] Interference region segmentation: A lightweight semantic segmentation model is adopted, which is pre-trained with a "road + interference" dataset (containing 50,000+ samples, labeled with 6 types of interference such as shadows and oil stains). The model outputs an "interference mask" ("1" represents interference pixels, and "0" represents non-interference pixels).

[0107] Interference pixel statistics: Count the total number of pixels with "1" in the interference mask;

[0108] Ratio calculation: Calculate the ratio between the total number of pixels with "1" and the total number of pixels in the image;

[0109] Exclusion rule: If the distance between the road markings (such as lane lines) and the defect is greater than the set number of pixels, it will not be included in the interference area (to avoid misjudgment of unrelated interference objects).

[0110] The color space similarity between the diseased area and the interference area is calculated using the Euclidean distance algorithm, and this is used as the interference similarity.

[0111] Acquisition process:

[0112] Color characteristics: Calculate the mean HSV color space value of the affected area and the interference area respectively. and ,

[0113] Color space similarity is calculated using Euclidean distance. .

[0114] For the edge sharpness, background interference ratio, and interference similarity of image data, the interference evaluation coefficient of the image data is determined by weighted calculation logic in combination with the set reference standard.

[0115] Specifically:

[0116] Calculate the disease assessment coefficient ,in These represent edge sharpness, background interference percentage, and interference similarity, respectively. The reference values ​​for edge sharpness and background interference ratio are set based on the image complexity evaluation criteria. These are preset weighting coefficients;

[0117] Based on the disease assessment coefficient and the interference assessment coefficient of image data, a weighted calculation logic is used to determine the complex assessment coefficient;

[0118] The pathology and perturbation coefficients of the image data are multiplied by their respective weighting coefficients, and the result is divided by an integer 2 to obtain the complex evaluation coefficient.

[0119] Establish a mapping rule between complexity evaluation coefficients and complexity levels, that is, set three sets of coefficient intervals corresponding to the complexity evaluation coefficients, with the three sets of coefficient intervals corresponding to the simple level, medium level and complex level respectively;

[0120] S3: Based on the complexity level label, if the image data complexity level label is simple, the labeling result can be directly output using the model combination scheme corresponding to the simple level. If it is not simple, the corresponding model combination scheme is called to input the image data for labeling and then output the labeling result. The corresponding model component is dynamically loaded and the corresponding processing flow is executed according to the call result.

[0121] S4: After all image data has been labeled, the image data is classified into priority queues based on the road features and actual damage characteristics corresponding to the image data; the priority queues include high priority, medium priority and low priority.

[0122] Specifically:

[0123] The road features include the road type and traffic flow level of the image data;

[0124] The actual characteristics of a disease include its location, type, and area.

[0125] The road types include highways, urban roads, and rural roads;

[0126] Different road types are assigned a set of road impact scores; the road impact scores range from 1 to 10, with highways having the highest road impact score, followed by urban roads, which in turn have the highest road impact score, which is the highest road impact score for rural roads.

[0127] The traffic flow data of the road area where the image data is located within a set time zone before the current time point is acquired, and the traffic flow level of the road area where the image data is located is determined, including high traffic flow, medium traffic flow and low traffic flow.

[0128] That is, to construct traffic flow data intervals corresponding to different traffic flow levels, match the traffic flow data of the road area where the image data is located with the corresponding traffic flow data intervals, and determine the traffic flow level;

[0129] For example, high traffic flow corresponds to a traffic flow data range of over 1500 vehicles / hour, medium traffic flow corresponds to a traffic flow data range of 500-1500 vehicles / hour, and low traffic flow corresponds to a traffic flow data range of less than 500 vehicles / hour.

[0130] Different traffic flow levels are set up to correspond to a set of traffic flow impact scores; the vehicle impact score ranges from 1 to 10, and the vehicle impact score for high traffic flow > medium traffic flow > low traffic flow.

[0131] The road impact score and traffic flow impact score of the image data are summed to obtain the vehicle impact coefficient of the image data.

[0132] The location, type, and area of ​​diseases in the statistical image data are determined. Disease areas belonging to the same disease type are accumulated and summed. Weighting coefficients are set for different disease types. The accumulated disease areas of different disease types are multiplied by the corresponding set weights, and then summed to obtain the disease influence coefficient of the image data.

[0133] The degree coefficient of image data is determined by comprehensively processing the vehicle impact coefficient and the defect impact coefficient based on image data and the reference standards established for different road types.

[0134] Specifically:

[0135] Using formula The degree coefficient is obtained by calculation. ;in These represent the vehicle impact coefficient and the defect impact coefficient, respectively. These represent the preset reference vehicle influence coefficient and reference damage influence coefficient for different road types, respectively; the preset values ​​for the reference vehicle influence coefficient and reference damage influence coefficient for highways are relatively low. These are preset weighting coefficients;

[0136] The total number of image data is divided by an integer three to obtain the number of high-priority queues, n (if not divisible by 3, round up). All image data are sorted from largest to smallest according to their degree coefficient. After sorting, n images are selected from left to right to form the high-priority queue. Half of the remaining image data is taken as the number of medium-priority queues, j. The remaining image data is re-sorted, and j images are selected as the medium-priority queue. The rest are used as the low-priority queue.

[0137] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0138] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0139] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0140] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0141] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0142] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0143] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0144] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0145] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for efficient and intelligent annotation of road disease data in collaboration with a large model, characterized by, The method comprises the following steps: Complexity determination: set three sets of complexity level labels for image data, and each set of labels corresponds to a pre-constructed model combination scheme. The complexity level includes simple, medium and complex levels. The basic features of the image data are extracted using the model combination scheme corresponding to the simple level, including disease entity features and recognizable features. After the complexity evaluation coefficient of the image data to be labeled is determined, the complexity level label is generated. The model combination scheme corresponding to the simple level is a lightweight model, which does not need to start a large language model and a visual language model. The model combination scheme corresponding to the complex level triggers full model collaboration, which includes generating a multi-disease text prompt, implementing multi-target differentiation combined with the prompt, generating a refined mask by splitting a large model, and verifying disease correlation. The disease entity features include disease area proportion, disease type number and disease distribution density. The recognizable features include edge sharpness, background interference proportion and interference similarity. For the disease area proportion, disease type number and disease distribution density of the image data, the disease evaluation coefficient of the image data is determined by using the weighting calculation logic combined with the set reference standard. For the edge sharpness, background interference proportion and interference similarity of the image data, the disturbance evaluation coefficient of the image data is determined by using the weighting calculation logic combined with the set reference standard. Based on the disease evaluation coefficient and the disturbance evaluation coefficient of the image data, the complexity evaluation coefficient is determined by using the weighting calculation logic. Generate labels: according to the complexity level label, if the complexity level label of the image data is simple, directly use the model combination scheme corresponding to the simple level to output the labeling result, if it is not simple, call the corresponding model combination scheme to input the image data for labeling and output the labeling result. Priority classification: after all image data are labeled, the image data are classified into priority queues according to the road features and disease actual features corresponding to the image data; wherein the priority queues include high priority, medium priority and low priority. The road features include the road type and traffic level of the image data. Different road types correspond to a set of road impact scores. Different traffic levels correspond to a set of traffic impact scores. The road impact score and the traffic impact score of the image data are added to obtain the vehicle impact coefficient of the image data. The disease actual features include the location, type and area of the disease. The location, type and area of the disease in the image data are counted, the disease area of the same disease type is added, different disease types correspond to a set of weight coefficients, the disease area of different disease types is multiplied by the corresponding set weight, and then summed to obtain the disease impact coefficient of the image data. Based on the vehicle impact coefficient and the disease impact coefficient of the image data, and the reference standard established according to different road types, the degree coefficient of the image data is determined by comprehensive processing. The total number of statistical image data is divided by an integer three to divide into three equal parts, and all image data is sorted in descending order according to the degree coefficient. After sorting, the image data of three equal parts is cut off from left to right as the high priority queue, the medium priority queue and the low priority queue.

2. The method of claim 1, wherein the method is characterized by, The disease ontology features of the image data are extracted, specifically: Calculate the percentage of the total number of pixels in all disease areas in the image to the total number of pixels in the image as the disease area ratio; Calculate the number of diseases in a unit physical area as the disease distribution density; Count the number of different types of road diseases in the image as the disease type number.

3. The method of claim 1, wherein the method is characterized by, Extract the recognizable features of the image data, specifically: Cut off the disease area in the image data and extract the disease edge to generate an edge mask. For each edge pixel in the edge mask, filter out the high gradient edge pixels with a gradient value higher than the reference gradient value, and calculate the ratio between the high gradient edge pixels and the total edge pixels as the edge sharpness. Calculate the percentage of the total number of pixels in all interference areas in the image to the total number of pixels in the image as the background interference ratio. Calculate the color space similarity between the disease area and the interference area using the Euclidean distance algorithm as the interference similarity.

4. The method of claim 1, wherein the method is characterized by, Generate a complexity level label, specifically: Set three groups of coefficient intervals corresponding to the complexity evaluation coefficient. The three groups of coefficient intervals correspond to the simple level, the medium level and the complex level.

5. The method of claim 1, wherein the method is characterized by, Also includes: Obtain the traffic flow data of the road area where the image data is located within a time zone before the current time point to determine the traffic flow level of the road area where the image data is located, wherein the traffic flow level includes high traffic flow, medium traffic flow and low traffic flow; Construct a traffic flow data interval corresponding to different traffic flow levels, match the traffic flow data of the road area where the image data is located to the corresponding traffic flow data interval, and determine the traffic flow level.

Images