Multi-source automated detection data closed loop quality control method

Abstract

This invention discloses a closed-loop quality control method for multi-source automated detection data, comprising: acquiring multi-source raw detection data and performing standardization processing to obtain a standard detection dataset; performing cross-device comparison processing and deviation analysis on the standard detection dataset to obtain a corrected defect detection result set; performing object-level risk analysis on the dataset, selecting objects to be reviewed based on the risk analysis results, and obtaining review and correction results; performing quality evaluation processing based on the corrected defect detection result set and the review and correction results to obtain detection data quality evaluation results; and updating a pre-configured set of device deviation parameters in a closed loop based on the detection data quality evaluation results. This invention, by establishing a closed-loop mechanism of access-correction-review-scoring-update, suppresses spatial and attribute deviations in multi-source heterogeneous data, achieves continuous evolution of the detection system performance, and improves the accuracy and reliability of road detection data.

Description

Technical Field

[0001] This invention relates to automated road maintenance inspection technology, and in particular to a closed-loop quality control method for multi-source automated inspection data. Background Technology

[0002] With the deepening of smart highway construction, the use of automated detection equipment such as lasers and vision systems for pavement defect identification has become a major means of highway maintenance. Achieving accuracy and standardization of detection data is beneficial for improving highway asset management and making maintenance decisions. In a multi-device collaborative operation environment, if the consistency and reliability of data output from different detection platforms can be guaranteed, a digital foundation for highways can be further constructed, enabling quality traceability throughout the entire lifecycle of maintenance operations.

[0003] Currently, quality control of automated testing data mainly relies on independent verification of single devices and subsequent random manual sampling. When processing multi-source heterogeneous data, geographic coordinate projection and attribute field mapping are typically used to achieve data fusion.

[0004] In view of this, existing technologies have the problem of difficulty in achieving spatial alignment and attribute synchronization of heterogeneous detection data in multi-source operating environments, and are unable to establish an adaptive quality control mechanism that can automatically optimize detection parameters based on verification feedback. Summary of the Invention

[0005] Purpose of the invention: This application provides a closed-loop quality control method for multi-source automated detection data, aiming to solve the above-mentioned problems in the prior art.

[0006] Technical solution: Firstly, a closed-loop quality control method for multi-source automated detection data, comprising:

[0007] Obtain multi-source raw detection data, and perform standardization processing on the multi-source raw detection data to obtain a standard detection dataset;

[0008] Cross-device comparison and bias analysis were performed on the standard test dataset to obtain a corrected set of disease detection results.

[0009] A risk analysis at the object level is performed on the corrected disease detection result set. Based on the risk analysis results, objects to be reviewed are selected for review, and the review and correction results are obtained.

[0010] The quality evaluation results of the detection data are obtained by performing quality evaluation processing based on the corrected disease detection result set and the review and correction results.

[0011] Based on the review and correction results and the quality evaluation results of the test data, the set of equipment deviation parameters is updated in a closed loop.

[0012] On the other hand, a computer-readable storage medium includes a stored computer program, wherein the computer program can be executed by an electronic device to perform the method of any one of the first aspects.

[0013] In another aspect, a computer program product includes a computer program that, when executed by a processor, implements the steps of the method in any of the first aspects.

[0014] In another aspect, an electronic device includes a memory and a processor, the memory storing a computer program and the processor being configured to perform the method of any one of the first aspects via the computer program.

[0015] Beneficial effects: This scheme constructs a global joint overdetermined equation system for weighted least squares solution, and dynamically switches spatial matching kernel functions such as one-dimensional overlap, centroid distance or regional intersection based on the aspect ratio of the disease, which suppresses the spatial misalignment of heterogeneous data and improves the cross-device disease matching accuracy from local coarse bounding box overlap to physical morphology level alignment.

[0016] Furthermore, the scheme employs a multi-dimensional conditional cross-grouping and hierarchical backtracking aggregation mechanism, which, while ensuring fine-grained parameter calibration for mainstream diseases, overcomes statistical fluctuations caused by long-tail samples, achieving highly robust attribute synchronization and restoration. The related technical effects will be described in detail below with reference to specific embodiments. Attached Figure Description

[0017] Figure 1 A flowchart of a closed-loop quality control method for multi-source automated detection data provided in an embodiment of this application.

[0018] Figure 2 This is a flowchart illustrating how a standard test dataset is compared and analyzed across devices to obtain a corrected set of disease detection results, as provided in this embodiment of the application.

[0019] Figure 3 This is a flowchart illustrating the process of performing object-level risk analysis on the corrected disease detection result set and selecting objects to be reviewed based on the risk analysis results, as provided in this embodiment of the application.

[0020] Figure 4 This is a flowchart illustrating a closed-loop update of the equipment deviation parameter set based on the review and correction results and the quality evaluation results of the test data, provided for an embodiment of this application.

[0021] Figure 5 This is a flowchart illustrating the quality evaluation results of the detection data based on the corrected disease detection result set and the verification and correction results provided in the embodiments of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a predetermined order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] To address the aforementioned issues, the applicant conducted in-depth searches and analyses, and discovered:

[0025] In practical applications, differences in the accuracy of positioning sensors across different devices and varying sensitivity of detection algorithms to predetermined defect morphologies often lead to spatial discrepancies in detection results for the same road segment, along with inconsistent standards for judging defect size. Furthermore, the lack of a feedback calibration mechanism means that detection parameters are typically fixed after system deployment, making it impossible to dynamically correct system errors based on manual review. This results in detection accuracy being significantly affected by environmental changes and equipment performance drift.

[0026] To solve these problems, combined with Figures 1 to 5 The present invention will be specifically described through the following embodiments.

[0027] For example, before describing the method, some parameters are explained below, as follows:

[0028] Among them, multi-source raw detection data refers to non-standard data collected by automated detection equipment of different brands, models or sensor principles during road operations.

[0029] The equipment can be specifically a visual inspection vehicle, a laser inspection vehicle, or a comprehensive inspection platform that integrates multiple types of sensors.

[0030] A standard detection dataset is a collection of detection results that has undergone field parsing, format conversion, and semantic alignment to eliminate heterogeneity and have a unified data structure.

[0031] The pre-configured set of equipment deviation parameters, which is a set of statistical parameters pre-stored in the system database, is used to describe the systematic error characteristics of each detection device in terms of defect classification, positioning and dimensional measurement.

[0032] Standardization processing includes field parsing and format conversion of multi-source raw detection data sets.

[0033] Quality assessment processing is used to quantify the overall reliability and data quality level of the current testing task.

[0034] Systematic offsets usually originate from initial alignment errors between different equipment positioning systems, and manifest as overall station offsets or lateral offsets.

[0035] Object matching is used to determine whether the detection records output by different devices point to the same physical defect.

[0036] For example, the deviation parameter update is a feedforward correction update based on cross-device comparison evidence of the current batch; the closed-loop update is a backward statistical update based on manual review feedback; both use the same exponential moving average mechanism when merging parameters, and the later update uses the result of the earlier update as the old parameter baseline, without generating version overwrite conflicts.

[0037] The equipment stability index is obtained by calculating the ratio of the standard deviation to the mean of the number of defects output by the corresponding testing equipment in a number of recent testing batches, i.e., the coefficient of variation. The larger the coefficient of variation, the more drastic the fluctuation of the equipment performance and the lower the stability.

[0038] The abnormal risk indicator can be obtained based on the ratio of the number of abnormal markers triggered by the current disease object during the correction process to the total number of preset abnormal marker types. The abnormal marker types include the corrected area being lower than the preset area lower limit, the category replacement confidence being lower than the preset threshold, and the scale correction coefficient being outside the preset normal range.

[0039] In some embodiments, the review determination feature refers to a three-dimensional index used to calculate the object-level review priority score, including object-level uncertainty index, cross-device conflict index, and equipment historical stability index; it reflects the probability influencing factors of each defect object being selected in the review process, and can be used as input features for the selection probability estimation model in the closed-loop update stage.

[0040] The following are six evaluation metrics, for example:

[0041] The detection confidence level, which characterizes the original probability output of the device for disease classification and localization, can be directly obtained from the device's original dataset.

[0042] The consistency index reflects the degree of overlap in the detection results of multiple devices at the same spatial location, and its value is negatively correlated with the degree of cross-device conflict.

[0043] The data integrity index is obtained by calculating the ratio of the number of valid fields in the current disease object record to the total number of fields that should be filled in.

[0044] The manual review pass rate represents the proportion of defects or equipment that are judged to be accurate during the manual inspection process.

[0045] Equipment stability indicators are used to assess the performance fluctuations of equipment during long-term operation.

[0046] Abnormal risk indicators quantify situations such as abnormal correction magnitude, logical conflicts, or abnormal device hardware status.

[0047] In some embodiments, the preset reference scale is defined as the standard deviation of the deviation distribution of all devices within each evaluation batch.

[0048] exp() represents the exponential function.

[0049] For example, the upper limit parameter ω of the weights _max The value can be set to 20. Specifically, a suitable cutoff threshold can be determined using robust statistical methods based on the probability distribution characteristics of the reviewed samples.

[0050] For example, the aspect ratio threshold may be 5, and the area threshold may be 0.5m. 2 It can be adapted to the local road damage distribution characteristics.

[0051] Optionally, the lateral tolerance parameter σ _lat The value is 0.3m, and the distance tolerance σ _p The value is 0.5m. It can be adjusted adaptively according to the positioning accuracy of the detection equipment and the width of the road lanes.

[0052] The preset quality threshold is set based on the image sharpness score, and the specific value can be determined according to the imaging parameters of the detection equipment and the requirements of the model training for sample quality.

[0053] The above parameters and other parameters not mentioned can be flexibly adjusted and optimized by technical personnel according to the on-site working conditions, and this application does not make a unique limitation on them.

[0054] On the one hand, an exemplary scheme for a closed-loop quality control method for multi-source automated detection data is provided. This scheme establishes mechanisms for standardized access, deviation correction, risk review, quality scoring, and closed-loop updates, thereby achieving quality control of multi-source heterogeneous road detection data. The specific steps can be as follows:

[0055] Specifically, the method includes the following steps:

[0056] Step 101: Obtain multi-source raw detection data and standardize the multi-source raw detection data to obtain a standard detection dataset.

[0057] Furthermore, the multi-source raw detection data is uploaded to the cloud server by various automated detection devices. The acquired raw data includes at least device identification data, collection time data, road segment number data, station number interval data, raw image or video frame data, defect detection result data, and device operating status data. Further standardization processing maps the heterogeneous data into a unified structure.

[0058] In the corresponding implementation, the system reads the original file through a pre-configured data adapter module and maps the file format, field name, timestamp format, and spatial coordinate expression of different brand devices into a unified field detection data table.

[0059] For example, convert timestamps of different formats to local time in ISO8601 format;

[0060] Spatial coordinate expressions such as latitude and longitude coordinates and local coordinate system mileage values ​​are uniformly converted into a four-tuple structure consisting of route number, station number, lane number, and lateral offset.

[0061] Standardize the units for the size parameters of the disease.

[0062] Based on this, the acquisition of standard detection datasets can be further used to implement unified semantic labels for diseases, completion of missing fields, removal of abnormal fields, and binding of device metadata.

[0063] Among them, the semantic labels of diseases are uniformly implemented by querying a pre-set category mapping table, which maps the original category codes defined by different manufacturers to a unified category code that conforms to the highway technical condition assessment standard, thereby unifying the structure of the standard detection dataset.

[0064] Step 102: Perform cross-device comparison processing and bias analysis on the standard test dataset to obtain the corrected disease detection result set.

[0065] In this embodiment, cross-device comparison processing involves correlating and comparing detection results from different devices within the same spatial and temporal windows to identify differences in detection of the same disease object by different devices. Deviation analysis, based on the comparison results and combined with a pre-configured set of device deviation parameters, quantifies the positioning offset, dimensional deviation, and category confusion degree of each device.

[0066] By performing deviation analysis, the error patterns of each device relative to the reference baseline can be determined. These patterns are then used to perform feature corrections on the defect records in the standard inspection dataset, including position translation, size scaling, and label verification, resulting in a corrected defect inspection result set.

[0067] Step 103: Perform object-level risk analysis on the corrected disease detection result set, and select objects to be reviewed based on the risk analysis results to obtain the review and correction results.

[0068] Object-level risk analysis involves extracting classification uncertainty, cross-device comparison characteristics, and historical stability characteristics of each disease object, and comprehensively calculating a risk index reflecting the reliability of that disease object. Based on this, high-risk, highly controversial, or abnormally corrected disease objects can be automatically screened from the corrected disease detection results as objects to be reviewed.

[0069] After this, the selected objects to be reviewed are pushed to the manual review terminal, where professionals confirm, correct, or reject them by referring to the original images, thereby generating review and correction results containing real disease attribute information.

[0070] Step 104: Based on the corrected disease detection result set and the verification and correction results, perform quality evaluation processing to obtain the detection data quality evaluation result.

[0071] Accordingly, the corrected test data and the corrected data after manual review are read, and quality evaluation sub-items such as test confidence, consistency index, data integrity, and review pass rate are calculated.

[0072] Weighted fusion calculations are performed on each sub-item to generate quality scores at the disease object level, road section level, or task batch level, and the detection data quality evaluation results are generated in combination with anomaly judgment rules.

[0073] The evaluation results are output in the form of quality scores and anomaly warning tags, which are used for task scheduling and data governance.

[0074] Step 105: Based on the review and correction results and the test data quality evaluation results, perform a closed-loop update on the pre-configured equipment deviation parameter set to obtain the updated equipment deviation parameter set.

[0075] In this step, based on the difference between the actual attributes of the defects reflected in the review and correction results and the original output results of the equipment, and combined with the confidence weight of the quality score, a statistical update algorithm is used to iteratively optimize the set of equipment deviation parameters stored in the database.

[0076] In other words, the deviation parameter can gradually approach the actual physical characteristics of the equipment as the number of test samples accumulates, thus providing a more accurate reference for the calibration of the next round of test data.

[0077] Based on the above embodiments, the specific implementation steps of cross-device comparison processing and deviation analysis are further refined.

[0078] In one possible implementation, the following steps are included:

[0079] Step 201: Extract disease objects from the standard detection dataset to form a comparable disease object set.

[0080] In this embodiment, each disease-related data can be read from a standard detection dataset, and the corresponding disease identifier, spatial coordinates, size parameters, and category label can be extracted.

[0081] Furthermore, to achieve cross-device comparison, it is necessary to perform benchmark unification processing on each defect object. That is, the spatial location information such as the station number and lateral offset of each defect object is converted to a unified road section coordinate system.

[0082] The unified road segment coordinate system uses the centerline direction of the corresponding route as the station axis and the lateral offset direction perpendicular to the centerline as the offset axis. After benchmark unification, all disease objects are integrated to form a comparable set of disease objects under the same geographical benchmark and disease semantic system.

[0083] Step 202: Perform global systematic offset elimination processing on it (the set of comparable disease objects) to obtain the set of comparable disease objects after eliminating systematic errors.

[0084] Optionally, by utilizing the highly consistent anchor point defects detected by each device, the systematic translation vector between devices is estimated; and the coordinates of each device are translated and compensated in the global dimension so that the same defects output by different devices are aligned in spatial position, thereby eliminating systematic errors and obtaining a set of comparable defect objects.

[0085] Step 203: Based on this (the set of comparable defect objects after eliminating systematic errors), perform cross-device object matching to obtain a set of cross-device candidate matching pairs and unmatched objects.

[0086] For example, within a preset spatial overlap threshold and time correlation window, disease objects from different devices are compared pairwise, and a matching correlation score is calculated.

[0087] The optimal matching algorithm is used to find the most relevant combination of objects, forming a cross-device candidate matching pair set;

[0088] For isolated disease objects that cannot be matched and associated in other device results, they are marked as unmatched objects.

[0089] Step 204: Decompose the corresponding detection device's device deviation parameters based on the cross-device candidate matching pair set, and update the pre-configured device deviation parameter set using the device deviation parameters.

[0090] For each group of disease objects in the matching set, the differences in category label, center location, geometric boundary, and severity rating are calculated, which reflect the differences in detection habits and systematic errors of the corresponding equipment relative to the reference equipment or fusion benchmark.

[0091] For example, the differences are statistically analyzed according to equipment, disease type, and size order to obtain the decomposed equipment deviation parameters.

[0092] Furthermore, by using exponential moving average or weighted adjustment algorithms, the newly decomposed deviation parameters are integrated with the original pre-configured set of equipment deviation parameters to complete the real-time update of the parameter library.

[0093] Alternatively, based on the quantitative differences of various category labels, center positions, geometric boundaries, and severity ratings, classification statistics are carried out according to equipment type, disease category, and size magnitude to obtain the decomposed equipment deviation parameters.

[0094] Step 205: Based on the updated pre-configured set of equipment deviation parameters, perform feature correction on the corresponding equipment's defective objects and unmatched objects to obtain a corrected set of defect detection results.

[0095] Alternatively, reverse correction can be performed on the original detection records using the real-time updated deviation parameters.

[0096] Furthermore, the location deviation vector is subtracted from the positioning coordinates of the corresponding equipment, the defect size parameter is divided by the scale deviation ratio coefficient, and the category label is probabilistically corrected according to the category confusion matrix.

[0097] Furthermore, the corrected damage records are reorganized according to equipment and road sections to form a corrected damage detection result set for risk analysis and review.

[0098] Alternatively, for the positioning coordinates, a position deviation vector can be used for correction, specifically by subtracting the corresponding position deviation vector from the positioning coordinates output by the device;

[0099] For the size parameters of the defects, a scaling correction is performed using a scale deviation ratio coefficient. Specifically, the size parameters of the defects output by the equipment are divided by the corresponding scale deviation ratio coefficient.

[0100] For category labels, a probability calculation method is used to correct them based on a preset category confusion matrix to ensure that the category labels remain accurate.

[0101] On the other hand, the specific implementation method of cross-device comparison processing and deviation analysis is further explained, especially for the decomposition and correction of equipment deviation parameters. This embodiment provides a specific technical solution based on conditional grouping and hierarchical backoff mechanism, which specifically includes:

[0102] Step 301: Extract the device identifier, disease category, and size attribute of each matching object in the cross-device candidate matching pair set.

[0103] Furthermore, for the cross-device candidate matching pair set, the matching records are read sequentially. For each matching pair, the corresponding device hardware number or unique identifier is extracted as the device identifier; the unified semantic label mapped during the standardization process is extracted as the disease category; and the area, length, or equivalent diameter is extracted as the size attribute.

[0104] Furthermore, the selection of size attributes is related to the morphology of the disease; for example, length is extracted first for linear diseases, and coverage area is extracted for planar diseases.

[0105] Step 302: Based on the equipment identifier, the disease category, and the size range divided by the size attribute, perform multi-dimensional cross-grouping of the cross-equipment candidate matching pair set to obtain the corresponding condition grouping set.

[0106] Accordingly, a multidimensional conditional probability distribution is constructed. That is, the size attribute is divided into several size intervals using logarithmic or linear equidistant intervals. For example, the area is divided into multiple intervals such as less than 0.1 square meters, 0.1-1.0 square meters, and more than 1.0 square meters.

[0107] Furthermore, by using equipment identification, disease category, and size range as three dimensions, a joint index and cross-segmentation are performed on the data samples in the cross-equipment candidate matching pair set. This ensures that matching pairs with the same equipment origin, similar disease characteristics, and similar geometric dimensions are grouped into the same data subset. These data subsets together constitute the conditional grouping set.

[0108] In some alternative implementations, the size range can be divided using a logarithmic equidistant division method, which makes the range width narrower in the small size area and wider in the large size area, thus better conforming to the error distribution pattern of IoT sensors or vision sensors.

[0109] Step 303: For the target condition group in the condition grouping set, count the number of matching pairs contained in it. In response to the number of matching pairs reaching the preset minimum statistical value, directly evaluate the corresponding category bias, position bias, scale bias and severity bias based on the matching objects in the target condition group, and use them as equipment bias parameters.

[0110] After obtaining the set of conditional groups, iterate through each target conditional group. Calculate the number of cross-device matching pairs contained within that target conditional group. If the number of samples is greater than or equal to a pre-set minimum statistic, such as 30 pairs, the current subset is determined to be statistically significant, and then the bias parameter is calculated at the current fine-grained level.

[0111] The positional deviation can be obtained by calculating the mean vector of the coordinate difference between the center point of the target device and the reference object.

[0112] Scale deviation can be obtained by calculating the mean of the ratios of dimensional attributes;

[0113] Severity bias is obtained by calculating the mean of rating score differences.

[0114] For category bias, a category confusion probability matrix is ​​constructed by statistically analyzing the mapping frequency of each type of label. The statistical results from these four dimensions are combined to form the device bias parameters under the current predetermined conditions.

[0115] Step 304: In response to the number of matching pairs being lower than the preset minimum statistic, a hierarchical backoff mechanism is triggered, and the target condition group is aggregated with other size ranges of the same disease category in turn for analysis.

[0116] If the aggregated count is still lower than the preset minimum statistical value, further aggregate analysis is performed on the global dimension of the corresponding device until the number of associated matching pairs after aggregation reaches the preset minimum statistical value, so as to output the corresponding device deviation parameter.

[0117] In real-world road network detection scenarios, certain pre-defined combinations (such as a device detecting a large area of ​​blocky cracks) may suffer from sample sparsity, meaning the number of matching pairs is lower than the preset minimum statistic. Directly calculating the average error on small samples can easily lead to estimation bias and parameter oscillation.

[0118] Therefore, the sparsity problem can be addressed by triggering a hierarchical fallback mechanism. The specific operation is as follows:

[0119] Remove the size range dimension restriction and merge the target condition group with samples of all size ranges under the same equipment and the same disease category to expand the sample pool for aggregate analysis;

[0120] If the merged sample size still does not meet the requirements, the disease category restriction is further removed, and the system degenerates to the global dimension of the device, using the global average deviation of all disease samples detected by the device to replace the deviation parameter of the current target condition grouping.

[0121] If the number of associated matching pairs is still lower than the preset minimum statistic after aggregation to the global dimension of the corresponding device, then all available samples after aggregation are used for parameter estimation, and the corresponding device deviation parameters are marked as low confidence parameters.

[0122] Based on the aforementioned deviation parameter extraction mechanism, this embodiment further defines the correction action. Specifically, feature correction of the corresponding equipment defects based on the updated pre-configured equipment deviation parameter set includes:

[0123] Optionally, category label replacement is performed based on the corresponding category confusion probability to obtain the corrected category label;

[0124] When performing feature correction, the system queries the category confusion probability matrix corresponding to the pre-configured set of device deviation parameters. If the matrix shows that the original label output by the current device has an extremely high false alarm probability under predetermined conditions, such as being greater than a preset confidence threshold, and the false alarms are highly concentrated in a certain predetermined target category, then the category label of the disease object is automatically replaced with the target category with the highest probability.

[0125] In other embodiments, if there is no obvious single high-probability obfuscation point, but rather a diffuse distribution, the original category label is retained and marked as pending review.

[0126] Optionally, the disease center position is translated according to the position deviation vector, and the coordinates of the vertices of the boundary polygon of the disease object are updated synchronously to obtain the corrected spatial position;

[0127] For spatial position correction, the positional deviation vector corresponding to the current diseased object can be extracted, such as lateral offset or longitudinal offset. This vector is then summed and added to the centroid or center position coordinates of the diseased object to complete the overall translation operation.

[0128] Furthermore, all vertices constituting the geometric polygon of the disease boundary are extracted, and a position deviation vector is added to the coordinates of each vertex to achieve synchronous spatial alignment between the center position and the boundary range.

[0129] Optionally, the length, width, or area of ​​the lesion can be scaled and corrected according to the scale deviation ratio coefficient to obtain the corrected dimensional attributes;

[0130] For size correction, the corresponding scale deviation ratio coefficient can be extracted. The original length, original width, or original area attribute values ​​of the defective object are divided by this ratio coefficient to restore the geometric dimensions.

[0131] In other embodiments, the severity rating of the disease can be corrected based on the severity deviation to obtain a corrected severity rating.

[0132] In one possible implementation, an object-level risk analysis is performed on the corrected disease detection result set, and objects to be reviewed are selected based on the risk analysis results. Specifically, this includes:

[0133] Step 401: Extract the classification uncertainty features and cross-device comparison features of each disease object in the corrected disease detection result set, obtain the pre-stored historical verification features of the corresponding detection equipment, and quantify them into object-level uncertainty indicators, cross-device conflict indicators and equipment historical stability indicators, respectively.

[0134] Furthermore, the corrected disease records are read, and three-dimensional core evaluation metrics are extracted for each disease object. Regarding the classification uncertainty feature, the posterior probabilities of each candidate category output by the detection device that generated the disease object are obtained, and the maximum posterior probability term is extracted. The calculation formula is as follows:

[0135] U(o) = 1 - max(P(c|o)),

[0136] In the formula, U(o) represents the object-level uncertainty index, P(c|o) represents the set of posterior probabilities for each category of the defective object determined by the equipment, and max() corresponds to the maximum value function.

[0137] In other words, 'o' corresponds to the disease object, and 'c' corresponds to the disease category.

[0138] For example, U(o) takes a value between 0 and 1, and the larger the value, the lower the classification confidence of the device itself.

[0139] In another example, if some detection devices only output a single hard label without providing a posterior probability value, the supplementary value of the device's historical detection accuracy for the corresponding disease category can be used to replace the uncertainty indicator.

[0140] Furthermore, based on the historical verification characteristics of the equipment, the performance of all equipment related to the defect in historical tasks was traced and detected. The specific calculation formula can be expressed as:

[0141] D(o) = 1 - (n _pass / n _review );

[0142] In the formula, D(o) represents the historical stability index of the equipment, and n _passn represents the number of defects that have been manually verified and confirmed for the equipment in historical records. _review This represents the total number of defects that the equipment has participated in for manual review. The larger this value is, the higher the proportion of defects that the equipment used to test that object has historically been rejected manually, and the more unstable the equipment's testing performance is.

[0143] For example, for newly accessed devices with no historical review records, their D(o) will be set to a preset initial median of 0.5. The quantitative calculation of cross-device comparison features and their corresponding conflict indicators is further elaborated in steps 4011-4013, specifically including:

[0144] Step 4011: Determine the number of covering devices used to detect the diseased objects in the same spatial location.

[0145] Under a unified road segment coordinate system, based on the spatial boundary range of the defect objects output by each device, the total number of devices that have performed effective detection tasks within the spatially associated area where the target defect object is located is counted; the parameter for the number of covered devices is set as K(o).

[0146] Depending on the value of parameter K(o), cross-device conflict risk can be divided into two different calculation branches to address the evaluation differences under different detection states.

[0147] Step 4012: In response to the number of covered devices being characterized as multi-device coverage, a contradictory conflict assessment is performed. Specifically, based on the degree of disagreement among the category labels output by each participating device for the diseased object, a cross-device conflict index is calculated.

[0148] When the number of covered devices K(o) is greater than or equal to 2, it indicates that multiple devices simultaneously detected the defect at that spatial location. In this case, the risk stems from inconsistencies in the assessment results between the devices. Further conflict assessment is performed, calculated using the following formula:

[0149] C(o) = 1 - (1 / K(o)) 2 )×Σ _a Σ _b (I(ca=cb));

[0150] Where C(o) is the cross-device conflict index, K(o) is the number of covered devices, ca and cb represent the corrected category labels output by the a-th and b-th devices for this object, respectively, I(ca=cb) is the indicator function, which takes a value of 1 when the two category labels are the same and a value of 0 when they are different, Σ _a and Σ _b These represent the double summation over all covered devices.

[0151] When all devices output the same category label, the result of C(o) is 0; the more types of disagreements and the more intense the conflict between devices, the closer the value of C(o) is to 1.

[0152] Step 4013: In response to the number of covered devices being characterized as single-device coverage, a coverage gap risk assessment is performed. Specifically, the assessment of the gap risk is calculated based on a preset gap baseline coefficient and the total number of devices performing detection in the current road segment obtained from the standard detection dataset, resulting in a cross-device conflict index.

[0153] When the number of covered devices K(o) equals 1, it indicates that although multiple devices are operating on the current road section, only one device detected the defect, while the other devices missed detection. In this case, the risk is not a category contradiction, but rather an isolated evidence risk due to a lack of cross-validation.

[0154] Optionally, the missing type calculation formula C(o)=C can be called. _miss ×(1 / n _dev_seg ); where n _dev_seg C represents the total number of devices that performed testing tasks on this road section. _miss This is the preset baseline coefficient for coverage gaps.

[0155] In this embodiment, the preset missing baseline coefficient is lower than the conflict score of the same divergent state in contradictory conflicts, for example, it is set to 0.6.

[0156] In other words, since the risk level of a single device detecting a problem independently but being unable to verify it is lower than the risk level of multiple devices detecting a problem simultaneously but giving completely contradictory conclusions, a classification of different risk causes has been achieved.

[0157] Step 402: The object-level uncertainty index, cross-device conflict index, and device historical stability index are fused together to obtain the object-level review priority score.

[0158] After extracting and quantifying the indicators across the three dimensions, they are weighted and fused. The specific process can be written as follows:

[0159] R(o) = w _U ×U(o)+w _C ×C(o)+w _D ×D(o).

[0160] Where R(o) is the merged object-level review priority score, w _U w _C w _D Each of the three indicators has a preset fusion weight, and the sum of the three weights equals 1.

[0161] Furthermore, since the three input indicators have all been normalized to the dimensionless range of 0-1, and the consistent direction indicates that the larger the value, the higher the risk, the priority score R(o) calculated by fusion is also strictly distributed in the range of 0 to 1, and there is no problem of dimension conflict.

[0162] Step 403: Perform a filtering operation on the object-level review priority score based on the preset review threshold to obtain the objects to be reviewed.

[0163] In this step, a pre-configured review threshold is obtained, for example, a threshold of 0.75 is set. Each disease object in the detection result set is traversed, and the calculated priority score R(o) is compared with the pre-configured review threshold.

[0164] If the priority score is greater than or equal to the preset review threshold, the diseased object can be determined to be a target with high uncertainty or high conflict risk. It is then extracted and written into the task queue to be reviewed, forming a set of objects to be reviewed under normal screening.

[0165] Based on this, the screening of objects to be reviewed also includes:

[0166] Correction anomalies identified during cross-device comparison processing and deviation analysis will be directly written into the task queue to be reviewed.

[0167] Among them, the correction of abnormal objects is not subject to the review threshold.

[0168] In other words, during the feature correction process of cross-device comparison processing and deviation analysis, the validity check of the correction results is performed, and the defective objects that do not meet the preset validity conditions are marked as correction abnormal objects.

[0169] Write the corrected exception objects directly into the queue of tasks awaiting review;

[0170] Among them, the correction of abnormal objects is not subject to the review threshold.

[0171] Furthermore, this embodiment establishes an anomaly reporting mechanism that escalates beyond the appropriate level. During the preceding cross-device comparison and deviation correction phase, a correction validity check is performed to identify over-scaled objects, such as those with a corrected area less than 0.001 square meters, or objects where multiple target categories have similar confusion probabilities, leading to category replacement failure. The underlying abnormal data is then marked as anomaly objects.

[0172] In this step, the corrective anomaly objects are directly added to the final queue of tasks awaiting review. Regardless of whether their priority score R(o) reaches the review threshold, corrective anomaly objects will forcibly trigger manual intervention.

[0173] This embodiment introduces a dichotomous risk quantification assessment of coverage gaps and contradictions to address the inefficiencies in resource allocation for manual review and the inaccuracies in verification feedback under massive data conditions. This approach separates genuine disagreements between devices from the risk of isolated evidence, not only intercepting anomalies in deep algorithmic logic but also improving the recall and utilization rate of high-value disputed data.

[0174] In one possible implementation, a quality evaluation process is performed based on the corrected disease detection result set and the verification and correction results to obtain the detection data quality evaluation result, including:

[0175] Step 501: Extract the detection confidence level, consistency index, data integrity index, manual review pass rate, equipment stability index, and anomaly risk index to form a quality evaluation index set.

[0176] In other embodiments, detection confidence, consistency index and data integrity index are extracted from the corrected disease detection results set, manual review pass rate is extracted from the review and correction results, and pre-stored equipment stability index and abnormal risk index are obtained to form a quality evaluation index set;

[0177] In this step, the corrected disease records and intermediate data generated during the review phase are read, and evaluation indicators in six dimensions are extracted.

[0178] Furthermore, for the large number of defects that did not enter the manual review process, the corresponding manual review pass rate index suffers from a lack of definition under conventional statistical logic. To ensure the completeness of the quality evaluation index set, a historical estimation imputation strategy can be adopted.

[0179] The specific approach is to set the manual review pass rate of disease objects that have not entered the review process as the historical review pass rate of the corresponding testing equipment on the same type of disease, and in the subsequent weight allocation, multiply the corresponding weight of this estimated value by a constant discount factor of less than 1.

[0180] For example, the preset discount factor can be set to 0.5.

[0181] By introducing a discount factor, the credibility of the contribution of the measured verification value and the historical estimate to the final score is distinguished, thereby avoiding the loss of data or the distortion of scores in datasets composed of unverified data.

[0182] Step 502: The quality evaluation index set is dimensionless, and a fusion calculation is performed based on preset weights to generate the quality score of the corresponding disease object, road section or task batch, which serves as the quality evaluation result of the detection data.

[0183] Based on this, all indicators are uniformly converted into comparable scoring directions, meaning that a larger numerical value represents higher quality. In practice, a range normalization operation is applied to positive indicators, with the specific formula as follows:

[0184] x'=(xx) _min ) / (x _max -x _min );

[0185] For negative indicators, the value is first inverted and then normalized. The specific formula is as follows:

[0186] x'=(x _max -x) / (x _max -x _min );

[0187] In the above, x represents the original indicator value, x' represents the converted indicator, and x' represents the original indicator value. _min and x _max These represent the minimum and maximum boundary values ​​of the data distribution for this indicator extracted from the same evaluation batch, respectively.

[0188] For example, after the dimensionless processing is completed, the pre-configured weights of each dimension index are obtained, and a linear weighted sum is performed on each quality evaluation sub-item after conversion to calculate the basic value of the quality score for a single disease object.

[0189] Furthermore, based on system management requirements and according to the spatial affiliation or time batch attributes of the diseased objects, the basic quality score values ​​at the object level are arithmetically averaged or weighted and summarized to generate road segment-level quality score results or task batch-level quality score results.

[0190] Optionally, after generating the quality score, the generated quality score can also be converted into a corresponding disposal mark according to a preset quality score interval mapping rule.

[0191] Furthermore, the quality scoring results are divided into automatic pass intervals, manual sampling intervals, key review intervals, and abnormal warning intervals.

[0192] For example, the range of scores ranking in the top 75% of all samples in the same batch can be set as the automatic pass range, while the range of scores ranking in the bottom 5% can be set as the abnormal warning range.

[0193] Based on this, the detection data quality report is output by combining interval division and abnormal logic recording. The structured data in this report can be directly sent to the task scheduling terminal to guide the dynamic dispatch of periodic review tasks.

[0194] In one possible implementation, an inverse probability weighting mechanism and a normalized convergence criterion are introduced to provide highly stable closed-loop updates, specifically including the following steps:

[0195] Step 601: Based on the review and correction results, determine the selection probability of each disease object selected for the review process.

[0196] Accordingly, the system obtains the manually reviewed and corrected results as well as the status information of the unreviewed data.

[0197] Furthermore, the selection probability of each disease object selected for the review process is determined, specifically including:

[0198] In some embodiments, the object-level uncertainty index, cross-equipment conflict index, and equipment historical stability index corresponding to each disease object are obtained as input features, the review judgment feature of each disease object is used as input features, and whether each disease object is actually reviewed is used as the classification label to construct and fit a binary regression model.

[0199] In some embodiments, the post-hoc probability of each diseased object being selected into the review queue is calculated using a fitted binary regression model, and this probability is used as the selection probability.

[0200] In the specific calculation implementation, the binary regression model specifically adopts the logistic regression model. Its internal calculation formula is expressed as follows:

[0201] psel=1 / (1+exp(-(β _0 +β _1 ×U+β _2 ×C+β _3 ×D))).

[0202] Where psel is the post-hoc probability of the fitted output, i.e., the selection probability; U, C, and D are the object-level uncertainty index, cross-device conflict index, and equipment historical stability index obtained in the previous calculations, respectively; and β _0 To β _3 These are the regression coefficients obtained through maximum likelihood estimation using the full batch of data. This model allows for the extraction of a smoothed actual choice probability surface from the data, thus avoiding the statistical blind spot caused by fixed threshold truncation.

[0203] Step 602: Calculate the corresponding inverse probability weight based on the selection probability, where the inverse probability weight is negatively correlated with the selection probability.

[0204] After obtaining the selection probability of each sample, compensation weights are assigned to them. The baseline calculation formula for the inverse probability weight is: ω = 1 / psel. Further, an upper limit truncation is performed on the calculated initial weights; the specific truncation calculation formula is as follows:

[0205] ω=min(ω,ω _max );

[0206] In the formula, ω is the inverse probability weight of the object. _max The pre-configured weight limit parameter.

[0207] Step 603: Use inverse probability weights to weight the deviation contributions recorded in the review and correction results to obtain the weighted update deviation of each detection device.

[0208] For the positional deviation update operation, extract the original detection position record p of the verification sample. _i The actual location record p after verification _truth The difference between the two is used as the bias contribution for a single sample. Weighted update bias Δp _new The calculation of can be expressed as the following formula:

[0209] Δp _new =Σ(ω×(p _i -p _truth )) / Σ(ω).

[0210] For the weighted extraction of scale bias and severity bias, an isomorphic formula logic can be used for independent calculation.

[0211] Step 604: Determine the corresponding fusion confidence weight based on the quality evaluation results of the detection data, and use the weighted update bias and fusion confidence weight to perform fusion update on the pre-configured set of equipment bias parameters.

[0212] That is, the exponential moving average algorithm is used to smoothly integrate the newly extracted weighted update bias with the existing old parameters in the database. The corresponding update formula is:

[0213] Δp _updated =λ×Δp _new +(1-λ)×Δp _old ;

[0214] In the formula, Δp _updated The corresponding position deviation parameter after fusion and update is Δp _old The corresponding historical position deviation parameter is stored in the parameter library. λ corresponds to the smoothing fusion coefficient, which ranges from 0 to 1 and is used to control the fusion speed of the old and new parameters.

[0215] For example, in actual operation, the quality score of the current batch of detection data is read and used as a fusion speed adjustment factor in the adaptive determination of λ. Furthermore, if the quality score of the current batch is higher than the preset quality score benchmark and the number of reviewed samples is sufficient, the overall reliability of the data in this round is determined to be high, and the value of λ is automatically increased to increase the fusion ratio of the new deviation parameter; when the quality score is lower than the preset quality score benchmark, the value of λ is decreased to maintain the stability of historical parameters and avoid contamination of the parameter database by low-quality data.

[0216] Through the above mechanism, the evaluation results of the detection data quality directly affect the fusion and trade-off strategy of the old and new parameters in the closed-loop update process.

[0217] Based on this, the closed-loop update process also includes convergence determination, specifically including:

[0218] Accordingly, the changes in position deviation, scale deviation, and severity deviation corresponding to the pre-configured equipment deviation parameter set before and after the update are extracted.

[0219] Alternatively, read the new version parameters after the above exponential moving average calculation is completed, subtract them from the old version parameters before the update, and extract the absolute value of the difference or the magnitude of the difference vector to form a multi-dimensional set of parameter changes.

[0220] Accordingly, preset reference scales corresponding to various deviations are obtained, and the changes in position deviation, scale deviation and severity deviation are normalized using the preset reference scales.

[0221] Among them, the dimensional systems of different types of deviations are inconsistent, and direct superposition calculations will result in dimensional conflicts.

[0222] Therefore, normalization is performed, dividing the deviation changes in each dimension by the corresponding preset reference scale to obtain pure numerical variables stripped of their original physical units.

[0223] Accordingly, the normalized changes are weighted and aggregated to obtain a dimensionless convergence index that reflects the magnitude of the parameter set update.

[0224] The weighted aggregation formula is constructed as follows:

[0225] Г=w _p ×(Δp _change / σ _p_ref )+w _s ×(Δs _change / σ _s_ref )+w _g ×(Δg _change / σ _g_ref ).

[0226] In the formula, Δp _change The change in position deviation of the corresponding equipment, σ _p_ref Indicates the corresponding reference scale extracted; Δs _change The corresponding scale deviation change, σ _s_ref Indicates the corresponding reference scale for extraction; w _p w _s For each aggregation weight, Γ corresponds to the dimensionless convergence index, Δg _change σ represents the change in severity deviation. _g_ref This indicates the corresponding reference scale for extraction.

[0227] Accordingly, in response to the dimensionless convergence index falling below the preset convergence threshold, or the cumulative number of maintenance update rounds reaching the preset upper limit, the current parameter version is locked and the current closed-loop update is stopped;

[0228] If the dimensionless convergence index is not lower than the preset convergence threshold and the cumulative update rounds have not reached the preset upper limit, return to continue executing the closed-loop update.

[0229] Furthermore, a preset convergence threshold is obtained, and the dimensionless convergence index is compared with it. If the dimensionless convergence index is lower than the preset convergence threshold, it is determined that the parameter change amplitude is below the acceptable range. A convergence command is triggered, the iteration process is terminated, and the current equipment deviation parameter set is locked to prevent the model from overfitting or parameter oscillation.

[0230] In one alternative implementation, the preset convergence threshold is set to 0.05.

[0231] Based on this, the closed-loop update of the pre-configured set of equipment deviation parameters also includes:

[0232] Extract the diseased objects that meet the preset return conditions from the review and correction results, and write them into the model optimization sample pool;

[0233] The preset return conditions include: the manual review label of the corresponding disease object in the review and correction results is passed, the corrected field in the corrected disease detection result set is complete, and the image quality in the multi-source original detection data is higher than the preset quality threshold.

[0234] For example, in addition to the deviation parameter update business flow, the data feedback processing of the underlying detection algorithm model is executed in parallel, traversing all object records in the review and correction results.

[0235] For example, in addition to performing basic label, field, and image quality judgments, the preset reflux conditions can also add sample distribution balance constraints, that is, determine whether the cumulative number of the same type of disease in the current model optimization sample pool exceeds the preset balance limit.

[0236] For example, when a diseased object meets all the return conditions, its original image data and its corresponding manual verification labels are automatically packaged and stored in a dedicated model optimization sample pool for iterative training.

[0237] For example, samples that do not meet the preset return conditions are transferred to the disputed sample library for isolation and retention according to the missing type.

[0238] Based on this, this scheme employs post-hoc probability fitting and inverse probability weighting algorithms to perform reverse statistical compensation for the bias contribution of the review samples, supplemented by closed-loop convergence judgment logic that removes physical dimensions. This breaks the error amplification chain in the self-evolution of the closed-loop system and maintains the statistical unbiasedness and high stability of the adaptive quality control parameters under long-term iteration of massive data.

[0239] In one possible implementation, this embodiment solves the technical problem of spatial mismatch between multiple devices and difficulty in matching heterogeneous defects, including the following steps:

[0240] Step 701: Extract disease object combinations that meet preset category consistency and confidence conditions from the set of comparable disease objects corresponding to different devices to form a high-confidence anchor point set.

[0241] Before eliminating systematic offsets, it is necessary to find spatial alignment reference points, traverse the comparable defect object sets output by each device, and filter out object pairs that are of the same category, have a detection confidence level higher than the preset threshold, and whose station distance is within the preset coarse matching window.

[0242] These object pairs are preserved and constructed as a set of high-confidence anchor points, serving as the observational basis for solving the equations.

[0243] Step 702: Select a reference device and construct an overdetermined set of equations with the systematic position offset of the other detection devices relative to the reference device as the parameters to be estimated.

[0244] Among them, the spatial location difference of the high-confidence anchor pair set is used as the observation constraint of the overdetermined equation system, and the number of anchor pairs contained between the two detection devices is used as the weight of the corresponding observation constraint.

[0245] When the number of devices is greater than or equal to 3, there may be inconsistencies in the transmission of offset estimates between pairs of devices. Among all devices, the device with the longest detection section or the highest pass rate is selected as the reference device, and the actual station offset of the remaining devices relative to the reference device is set as an unknown.

[0246] For any two devices, the median difference between their anchor point pairs is calculated as the observation offset. Based on the difference relation matrix H, the unknown vector α, and the observation offset vector b, an overdetermined system of equations is constructed. Simultaneously, the number of anchor point pairs between the two devices is extracted to form a diagonal weight matrix W; the more anchor points, the greater the weight of the corresponding observation constraint.

[0247] Step 703: Use the weighted least squares criterion to solve the overdetermined system of equations globally to obtain the global joint offset of each detection device.

[0248] The constructed overdetermined system of equations is solved using matrix operations. Specifically, the closed-form solution formula of the weighted least squares method is used for calculation:

[0249] α _star =(H T ×W×H) -1 ×H T ×W×b.

[0250] Among them, H T Represents the transpose of matrix H. -1 This represents the matrix inversion operation. The vector α obtained by solving this formula is... _star This includes the global joint offset of each detection device. This solution process minimizes the overall alignment error of multiple devices from a global perspective.

[0251] In one alternative approach, if the current detection task involves only two devices, the median difference between the anchor point pairs between the two devices can be directly calculated as the offset, without the need to construct an overdetermined system of equations.

[0252] Step 704: Use the global joint offset to perform reverse compensation processing on the spatial location of the corresponding device's defect objects to obtain a set of comparable defect objects after eliminating system errors.

[0253] Accordingly, the global joint offset obtained from the solution is obtained, and the spatial station coordinates of all defect objects of each non-reference device are subtracted from the offset.

[0254] For the systematic error of the lateral offset, the same joint solution and subtraction operation are used.

[0255] It should be understood that, after reverse compensation, the detection coordinate systems of different devices are aligned, generating a set of comparable defect objects after eliminating systematic errors.

[0256] After eliminating systematic errors, cross-device object matching is performed on the comparable defect object set to obtain a cross-device candidate matching pair set. Specifically, the defect objects are processed in the following way:

[0257] Optionally, the size parameters of the diseased object are extracted, and the aspect ratio and coverage area of ​​the diseased object are determined based on the size parameters.

[0258] After completing coordinate alignment, the process moves to the shape matching stage, where the principal axis length L of the diseased object is extracted. _major and minor axis length L _minor And the coverage area, the aspect ratio AR=L is calculated by division. _major / L _minor ;

[0259] The coverage area can be used as an auxiliary dimensional factor for subsequent morphological feature matching.

[0260] Optionally, based on the aspect ratio and the distribution characteristics of the coverage area, the diseased objects can be classified into linear, compact, or planar forms to obtain corresponding morphological labels.

[0261] Furthermore, preset aspect ratio thresholds and area thresholds are provided.

[0262] For example, in response to the aspect ratio of the diseased object being greater than the aspect ratio threshold, a linear morphology label is assigned to it;

[0263] In response to an aspect ratio less than or equal to the aspect ratio threshold and a coverage area less than the area threshold, a compact shape label is assigned;

[0264] In response to an aspect ratio less than or equal to the aspect ratio threshold and a coverage area greater than or equal to the area threshold, a planar morphology label is assigned.

[0265] Optionally, the spatial matching degree can be calculated by dynamically selecting the corresponding spatial matching strategy based on the priority of the morphological labels in the disease objects of the two parties being matched.

[0266] Specifically, the calculation strategy for linear morphology combines one-dimensional overlap ratio with lateral distance attenuation, the calculation strategy for compact morphology uses centroid Euclidean distance attenuation, and the calculation strategy for planar morphology uses two-dimensional intersection-to-exclusion ratio of the coverage area.

[0267] Accordingly, the priority of the matching morphological labels is comprehensively determined, with planar morphology having the highest priority, followed by linear morphology, and compact morphology having the lowest priority. The kernel function is selected according to the highest priority label. The calculation formulas for linear, compact, and planar morphologies are as follows:

[0268] Linear morphological spatial matching degree S _spatial_elong =max(0,IoU _1D )×exp(-(d _y ) 2 / (2×(σ _lat ) 2 ));

[0269] Compact form space matching degree S _spatial_compact =exp(-distance2 / (2×(σ _p ) 2 ));

[0270] Spatial matching degree S of planar morphology _spatial_distrib =Area _intersect / Area _union ;

[0271] In the formula, IoU _1D Let d be the one-dimensional intersection-union ratio along the station axis. _y σ is the difference in lateral position. _lat σ is the lateral tolerance parameter, distance is the Euclidean distance between the centroids, and σ is the lateral tolerance parameter. _p For distance tolerance, Area _intersect Corresponding intersection area, Area _union The area of ​​the corresponding union.

[0272] Optionally, the size similarity can be calculated based on the size parameters of the diseased objects of both parties;

[0273] A comprehensive matching score is obtained by integrating spatial matching degree and size proximity. A global mapping relationship between different devices is established based on the comprehensive matching score, and object pairs that meet the preset matching threshold are extracted to form an initial one-to-one cross-device candidate matching pair set.

[0274] Further calculate the size similarity of the two diseases, and sum the spatial matching degree and size similarity degree with weight to obtain the comprehensive matching degree score. Construct a bipartite graph with this score as the weight, and use the Hungarian algorithm to solve the globally optimal one-to-one matching. Remove connections with scores below the preset matching threshold, and retain the object combinations to form the initial one-to-one cross-device candidate matching pair set.

[0275] Building upon this, cross-device object matching also includes one-to-many association processing performed on the unmatched objects, specifically including:

[0276] Optionally, among the unmatched objects on the same device, objects that are spatially adjacent and of the same category are locally merged to form a candidate merge group.

[0277] To address the issue where one device detects a large crack as a whole, while another device segments it into fragments, unmatched objects are extracted. Within the unmatched set of a single device, adjacent objects with a station spacing smaller than the preset merging tolerance and consistent semantic categories are searched and merged as a whole to form a candidate merging group. The new boundary and total area after merging are then calculated.

[0278] Optionally, a comprehensive matching score is calculated between the candidate merging group and the unmatched object of another device. When the comprehensive matching score meets the preset judgment criteria, a one-to-many matching relationship is established.

[0279] The generated candidate merge groups are treated as independent objects, and the kernel function is called again to recalculate the overall matching score with the unmatched objects from another device. If the new score exceeds the preset judgment benchmark, the association between the fragment combination and the whole object is confirmed, establishing a one-to-many matching relationship.

[0280] Optionally, a single-round backtracking check is performed on the region where a one-to-many matching relationship is established to determine whether there are any associated objects that have been occupied by the initial one-to-one cross-device candidate matching pair set within the spatial range of the candidate merging group of the corresponding device.

[0281] In other embodiments, to prevent local suboptimal solutions of the Hungarian algorithm from interfering with the merging process, a backtracking check is activated to scan the station intervals covered by the candidate merging groups and determine whether, in the initial one-to-one matching phase, any associated objects that should belong to the fragment group have been assigned to other low-scoring objects.

[0282] Optionally, if there are already occupied associated objects, the comprehensive matching scores of the occupied associated objects in the original mapping relationship and in the candidate merging group are compared. Based on the principle of selecting the best score, it is decided whether to release the occupied associated objects from the original mapping relationship and include them in the candidate merging group, and integrate them to generate the final cross-device candidate matching pair set.

[0283] Furthermore, the old score of the object in the initial one-to-one matching is extracted, the new score that it could contribute if it were added to the current candidate merging group is calculated, the two values ​​are compared, and in response to the higher new score, the old matching connection is cut off, the object is released and assigned to the new candidate merging group, the updated one-to-many mapping is integrated with the remaining one-to-one mapping, and the final cross-device candidate matching pair set is output.

[0284] The optional embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solution of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A closed-loop quality control method for multi-source automated detection data, characterized in that, include: Obtain multi-source raw detection data, and perform standardization processing on the multi-source raw detection data to obtain a standard detection dataset; Cross-device comparison and bias analysis were performed on the standard test dataset to obtain a corrected set of disease detection results. A risk analysis at the object level is performed on the corrected disease detection result set. Based on the risk analysis results, objects to be reviewed are selected for review, and the review and correction results are obtained. The quality evaluation results of the detection data are obtained by performing quality evaluation processing based on the corrected disease detection result set and the review and correction results. Based on the review and correction results and the quality evaluation results of the test data, the equipment deviation parameter set is updated in a closed loop. The process involves cross-device comparison and bias analysis of the standard detection dataset to obtain a corrected disease detection result set. This includes: extracting disease objects from the standard detection dataset to form a comparable disease object set; performing global systematic offset elimination processing to obtain a comparable disease object set after eliminating systematic errors; performing cross-device object matching to obtain a cross-device candidate matching pair set and unmatched objects; decomposing the device bias parameters of the corresponding detection devices based on the cross-device candidate matching pair set, and updating the device bias parameter set using the device bias parameters; and performing feature correction on the disease objects of the corresponding devices based on the updated device bias parameter set to obtain the corrected disease detection result set. The process involves a closed-loop update of the equipment deviation parameter set based on the review and correction results and the quality evaluation results of the test data. Specifically, this includes: determining the selection probability of each defect object selected for the review process based on the review and correction results; calculating the corresponding inverse probability weight based on the selection probability, where the inverse probability weight is negatively correlated with the selection probability; weighting the deviation contribution recorded in the review and correction results using the inverse probability weight to obtain the weighted updated deviation of each testing device; and using the weighted updated deviation to fuse and update the equipment deviation parameter set.

2. The method according to claim 1, characterized in that, A risk analysis at the object level is performed on the corrected disease detection result set. Based on the risk analysis results, objects to be reviewed are selected, specifically including: The classification uncertainty characteristics, cross-equipment comparison characteristics, and equipment historical verification characteristics of each disease object in the corrected disease detection results set are extracted and quantified into object-level uncertainty indicators, cross-equipment conflict indicators, and equipment historical stability indicators, respectively. The object-level uncertainty index, cross-device conflict index, and equipment historical stability index are integrated to obtain the object-level review priority score; Based on the preset review threshold, a filtering operation is performed on the object-level review priority score to obtain the objects to be reviewed.

3. The method according to claim 2, characterized in that, The screening of objects to be reviewed also includes: Correction anomalies identified during cross-device comparison processing and deviation analysis will be directly written into the task queue to be reviewed. Among them, the correction of abnormal objects is not subject to the review threshold.

4. The method according to claim 1, characterized in that, Based on the corrected disease detection result set and the verification and correction results, a quality evaluation process is performed to obtain the detection data quality evaluation results, including: Extract the detection confidence level, consistency index, data integrity index, manual review pass rate, equipment stability index, and anomaly risk index to form a quality evaluation index set; The quality evaluation index set is dimensionless, and a fusion calculation is performed based on preset weights to generate quality scores for corresponding disease objects, road sections, or task batches.

5. The method according to claim 1, characterized in that, Closed-loop updates to the equipment deviation parameter set also include: Extract the diseased objects that meet the reflux conditions and write them into the model optimization sample pool; The return conditions include: the manual verification label of the corresponding disease object is passed, the corrected fields are complete, and the image quality is higher than the preset quality threshold.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein the computer program can be executed by an electronic device to perform the method described in any one of claims 1 to 5.

7. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1 to 5.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 5 via the computer program.

Images