A Road Crack Detection Method and System Based on Multi-Party Soft Consensus and Visual Language Calibration

By employing a multi-person soft consensus and visual language calibration method, the problems of varying consensus strength and boundary ambiguity among multiple annotators were resolved. This enabled multi-level road crack detection, meeting diverse inspection needs, suppressing the response of crack-like interference objects, and maintaining the connectivity and accuracy of the detection.

CN123118570APending Publication Date: 2026-11-17NANJING UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611377547.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively utilize the varying strengths and boundaries of consensus among multiple annotators, making it difficult to unify the requirements for multi-strictness inspections in binary supervised compression during road crack detection, and resulting in insufficient suppression of crack-like interference.

Method used

By employing a multi-person soft consensus and visual language calibration method, independent annotations from multiple annotators are obtained. The original voting consensus is calculated and a continuous crack support field is obtained through structural regularization. A calibration correction field is generated by combining multi-scale visual coding and visual language calibration. During training, soft reconstruction, structure-aware ranking, and multi-operating point cutting constraints are applied. During inference, multi-rigor detection results are output by cutting according to different operating point thresholds.

Benefits of technology

It achieves road crack detection with multiple levels of stringency, avoids the loss of boundary ambiguity under binary mask supervision, maintains the connectivity and response hierarchy of crack detection, suppresses the response of crack-like interference objects, and meets the inspection requirements of different stringency levels.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118570A_ABST
    Figure CN123118570A_ABST
Patent Text Reader

Abstract

This invention discloses a road crack detection method and system based on multi-party soft consensus and visual language calibration. The method includes: acquiring a road image and a binary mask; calculating the original voting consensus and obtaining a continuous crack support field through structural regularization; inputting the road image into a crack detection network; outputting a basic segmentation response field through multi-scale visual encoding and soft decoding; then selecting positive crack concepts and negative crack-like concepts by a pre-trained visual language encoder, generating a calibration correction field, and fusing it to obtain a continuous crack support prediction field; during training, jointly applying soft reconstruction, structure-aware ranking, and multi-operating-point cutting constraints on the same prediction field; and during inference, cutting according to at least two preset operating-point thresholds to obtain multi-rigor detection results. This invention incorporates the consensus strength and boundary ambiguity of multi-party interpretations into the supervisory signal, suppressing crack-like interference responses such as road markings, shadows, seams, and tire tracks, facilitating the reading of crack evidence of different rigor according to maintenance needs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent road defect detection and pavement asset management technology, and in particular to a road crack detection method and system based on multi-person soft consensus and visual language calibration. Background Technology

[0002] Road surface crack detection is a fundamental step in road condition assessment and preventative maintenance. Vehicle-mounted camera systems can now capture road images at a high frequency, so image acquisition itself is no longer a major constraint. The main constraint lies in interpreting faint, intermittent, and disturbed traces in the images as evidence of cracks usable for maintenance management. Experienced inspectors, when interpreting these images, must weigh the differences between hairline cracks and asphalt texture, exclude structures similar in appearance to cracks such as road joints, road shadows, lane markings, tire tracks, and repair marks, and apply varying levels of stringency depending on the purpose—whether it's screening, routine inspection, or high-priority confirmation.

[0003] Most existing deep learning-based crack segmentation methods compress the supervised target into a binary mask, simplifying training and evaluation to determining whether a pixel overlaps with this mask. While this approach facilitates the use of common segmentation loss and intersection-over-union (IoU) metrics, it loses information shared by multiple trained annotators on the same road surface. Annotation discrepancies often occur in areas with high interpretation difficulty, including blurred boundaries, hairline crack segments, closed crack edges, the boundary between pavement joints and shadows, and pixels at the far end of the perspective. Binary masks force each pixel to take only two values: yes or no, while on-site interpretation more closely approximates the graded support strength. For slender structures, a difference of one or two pixels in annotation can alter connectivity, centerline continuity, and length estimation, thus affecting subsequent maintenance sequencing.

[0004] Techniques such as soft labeling, probabilistic segmentation, and multi-annotator fusion attempt to preserve the uncertainty of interpretation, but common practices still treat label discrimination as statistical noise, rarely compiling multi-vote data into structured evidence that can directly serve inspection operations, and rarely incorporating different working points such as sensitivity, standard, and strictness into the same continuous field training objective. Another type of cue-based segmentation method and visual-language segmentation method can refer to the target using natural language, but they focus on locating the object described by the cue; thin and high-contrast crack-like interference objects are easily amplified along with real cracks, lacking negative calibration methods for named interference objects. In addition, the above methods usually only output binary results at a single threshold, making it difficult to provide crack evidence of different strictness levels in the same inference, and in engineering, it is often necessary to train and maintain a separate segmentation model for each strictness level.

[0005] Therefore, there is a need for a road crack detection method and system that can organize multiple independent annotations into a structured regularized continuous crack support field, simultaneously constrain the relative order of responses, the operability of multiple working points, and the connectivity of curve structures on a single continuous prediction field, and suppress crack-like interference responses with the help of visual language concepts, so that the output evidence can be read with different degrees of strictness according to maintenance needs. Summary of the Invention

[0006] The purpose of this invention is to provide a road crack detection method and system based on multi-person soft consensus and visual language calibration, in order to alleviate the problems of consensus strength and boundary ambiguity among multiple annotators in binary supervision compression, the difficulty of unified continuous prediction in meeting the inspection requirements of multiple strictnesses, and the insufficient suppression of crack-like interference objects corresponding to lane markings, road shadows and road joints.

[0007] The technical solution to achieve the purpose of this invention is: a road crack detection method based on multi-person soft consensus and visual language calibration, comprising:

[0008] Obtain road images and binary masks independently annotated by multiple annotators, calculate the original voting consensus, and obtain a continuous crack support field through structural regularization;

[0009] The road image is input into the crack detection network. The basic segmentation response field is output by multi-scale visual encoding and soft decoding. Then, the pre-trained visual language encoder dynamically selects the positive crack concept and the negative crack-like concept according to the image, generates the calibration correction field and fuses it to obtain the continuous crack support prediction field.

[0010] During training, soft reconstruction, structure-aware ranking, and multi-operating-point cutting constraints are jointly applied on the same prediction field.

[0011] During inference, the results are cut according to at least two preset working point thresholds to obtain multi-rigidity detection results.

[0012] Furthermore, the threshold values ​​of each working point used when applying the multi-working-point cutting constraint correspond one-to-one with the threshold values ​​of each working point used when cutting during the inference phase and have the same value. This allows the crack detection network to use the working-point threshold values ​​actually used in the inference phase as the cutting target during the training phase, and to derive the crack detection results of each strictness from the same continuous crack support prediction field, without the need to train a segmentation head separately for each strictness.

[0013] This invention also provides a method for detecting road cracks, comprising: acquiring a road image to be detected; inputting the road image into a trained crack detection network, wherein a multi-scale visual encoder and a soft decoder output a basic segmentation response field, and a visual language calibration branch generates a visual language calibration correction field based on selected positive and negative concepts, and the two are fused to obtain a continuous crack support prediction field; segmenting the continuous crack support prediction field according to at least two preset operating point thresholds, and outputting multi-rigor crack detection results; wherein the trained crack detection network is trained using the continuous crack support field obtained by the aforementioned method as a soft supervision target, and according to the aforementioned training target; thus, the invention can be implemented independently by deploying only the inference stage.

[0014] A road crack detection system based on multi-user soft consensus and visual language calibration is provided to implement the above method. The system includes:

[0015] The multi-annotation acquisition module is used to acquire road images and obtain the annotations from... The annotation module obtains a set of binary crack masks by annotators independently annotating the same road image; the multi-annotation acquisition module returns and re-annotates annotation results that are incomplete or whose annotation positions are significantly off-target, and stores and outputs the road image that has passed the verification and the set of binary crack masks according to the image index;

[0016] A continuous crack support field construction module is used to receive the binary crack mask set, calculate the original voting consensus score per pixel, and then... After applying the structure regularization operator, the output is a continuous crack support field. The structure regularization operator includes at least a monotonically decreasing shaping of the consensus response based on the distance from the pixel to the center line of the crack, and normalization based on the boundary calibrated and fixed on the training set.

[0017] The basic segmentation response prediction module receives the road image, extracts at least four levels of feature maps with different strides using a multi-scale visual encoder of the crack detection network, and then merges these feature maps by a soft decoder of the crack detection network after mapping them to a uniform channel width, outputting an unnormalized basic segmentation response field. With decoder projection features and will Send to the visual language calibration module;

[0018] The visual language calibration module is used by a pre-trained visual language image encoder and text encoder to dynamically select positive and negative concept subsets from a fixed visual language concept base based solely on the appearance of the road image. Calculate the dense concept similarity with the selected concept, and output a single-channel visual language calibration correction field via the calibration head. Then combine it with Fusion, outputting continuous cracks to support the predicted field ;

[0019] Joint training module for continuous crack support field The crack detection network is trained for a soft-supervised objective, the training objective of which includes at least constraints. and Soft reconstruction constraints and constraints for value differences and Structure-aware sorting constraints based on the relative order of pixel pairs, and constraints and The network parameters after training are output by multi-operating-point cutting constraints where the cutting results are consistent at at least two preset operating-point thresholds.

[0020] The multi-working-point output module is used to load the trained network parameters, predict the road image window by window in a sliding-window manner, and perform Gaussian weighted fusion on the overlapping areas of adjacent windows to obtain the full image. Then, according to at least two preset working point thresholds, The cracks are cut separately, and the multi-strictness crack detection results are output, which correspond one-to-one with the threshold values ​​of each working point.

[0021] Furthermore, the visual language calibration module caches the selection results of the positive concept subset and the negative concept subset on the entire road image, and makes each sliding window crop of the same road image share the cached selection results; the system also includes a maintenance decision interface module, which performs connected component analysis on the mask cut from the multi-strictness crack detection results under the selected working point threshold, obtains the regional crack index by weighted summation of road segment-level statistics including connected skeleton length, average crack width and regional area density, and outputs the results after dividing each road segment into high, medium and low maintenance priorities according to a preset grading threshold.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) The present invention aggregates the independent binary annotations of multiple annotators on the same road image according to voting and then performs the structure regularization operator. Organized into a continuous fracture support field This allows the strength of consensus and boundary ambiguity in multi-person interpretation to be continuously entered into the supervision signal, thereby avoiding the loss of interpretation discrepancies on blurred boundaries, hairline crack segments and far-end pixels in binary mask supervision, making the supervision signal closer to the graded crack evidence strength in on-site interpretation.

[0024] (2) The present invention constructs the support field The system employs response shaping based on centerline distance and a normalized boundary that is calibrated and fixed once on the training set, thereby enabling... The value of the threshold has the same caliber across different road images, thus ensuring that the threshold for the same working point has consistent semantics across all road images. This provides a prerequisite for the establishment of multi-working-point segmentation constraints and avoids the situation where the soft label of image-by-image normalization is difficult to bear the semantics of the operational threshold.

[0025] (3) The present invention constrains the relative order of the prediction field in three directions: response level, center line to boundary, and crack relative to crack-like interference objects by the structure-aware sorting constraint, thereby maintaining the response hierarchy and connectivity from the core to the edge on the slender curve structure, alleviating the problem of the labeling difference of one or two pixels under binary supervision, that is, changing connectivity, center line continuity and length estimation, and thus affecting the maintenance sorting.

[0026] (4) In this invention, the actual operating point threshold used in the inference phase is written into the training target in a differentiable form, and the operating point thresholds of the training phase and the inference phase are in one-to-one correspondence, so that the prediction field is supported by the same continuous crack. Export crack detection results at multiple levels of severity, including sensitive, standard, and strict, without the need to train segment heads separately for each severity level. This allows for the on-demand reading of crack evidence at the appropriate severity level between screening, routine inspections, and high-priority confirmation.

[0027] (5) The present invention utilizes a pre-trained visual language encoder to dynamically select negative concepts based on image appearance, and generates a pixel-by-pixel calibration correction field based on dense concept similarity to suppress the basic segmentation response. This reduces the response of crack-like interference objects corresponding to lane markings, road shadows, road joints, tire marks and repair marks while preserving the support of real cracks. This changes the existing use of text prompts to find target objects, which also enhances the interference response on slender, high-contrast structures. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0029] Figure 2 A schematic diagram illustrating the construction of a continuous crack support field obtained through structural regularization for multi-vote consensus.

[0030] Figure 3 This is a schematic diagram of the overall architecture of the crack detection network and the visual language calibration branch.

[0031] Figure 4 This is a schematic diagram of sample pair construction in structure-aware sorting supervision.

[0032] Figure 5 This is a schematic diagram of the dynamic calibration process for visual language-related crack interference.

[0033] Figure 6 This is a schematic diagram of the module composition of the system of the present invention.

[0034] Figure 7 A schematic comparison of the outputs of the prediction field supporting the same continuous crack at sensitive, standard, and stringent operating points. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings. The specific values, network structures, and data scales described below are illustrative and do not constitute a limitation on the scope of protection.

[0036] Combination Figure 1 This embodiment presents a road crack detection method based on multi-user soft consensus and visual language calibration, including steps S1 to S6. The method aggregates binary crack masks provided by multiple annotators for the same road image into a continuous crack support field. Using this as a soft supervision target, the crack detection network predicts the continuous crack support prediction field on the same continuous dimension. The visual language calibration branch suppresses the response at crack-like interference points based on the image appearance. During inference, it derives multi-strictness detection results according to the sensitive operating point, standard operating point, and strict operating point, which correspond to three levels of strictness from low to high.

[0037] S1. Obtain road images and acquire a preset number of images. Binary crack mask sets obtained by renowned annotators from independently annotating the same road image. .

[0038] S1-1. Acquisition of road images. In one embodiment, the road images are vehicle-view images acquired by an onboard camera system along the driving direction. The acquired scenes cover perspective distortion, lighting changes, and scene elements that present a slender, high-contrast appearance, such as road shadows, lane markings, repair areas, road joints, asphalt textures, vehicle occlusion, tire tracks, and road edge boundaries. The road images can also be acquired by drones, handheld devices, or fixed-position devices.

[0039] S1-2, Labeling Protocol. For each of the aforementioned road images, by... Each annotator, trained in road inspection, independently generates a binary crack mask. Annotation is performed at full resolution with local zoom-in allowed; annotation begins on a blank canvas, without automatic proposals or pre-set masks to avoid inter-annotator interference. The annotators label visible or defensible crack support pixels as... Pixels lacking evidence of cracks are labeled as , record the The annotation specialist is in the pixel The label at the location is ; Pick to Name, in this embodiment .

[0040] S1-3, Annotation Quality Control. Each submitted image is reviewed; empty annotations or annotations that are significantly off-target to the image content are returned and re-annotated; intermediate results are not shared among the annotators.

[0041] S2. Calculate the original voting consensus graph based on the voting scores. Then, through the structure regularization operator The value obtained is in Continuous crack support field The structure regularization operator It includes at least a monotonically decreasing shaping of the consensus response based on the distance from the pixel to the center line of the crack, and normalization of the shaped response based on the boundary calibrated and fixed on the training set.

[0042] like Figure 2 As shown, this step aggregates independent annotations from multiple users into a soft-supervised target located on a continuous dimension. The structure regularization operator... The internal structure includes six deterministic steps, from S2-2 to S2-7.

[0043] S2-1, Calculation of the original voting consensus.

[0044]

[0045] in, For pixels The original consensus score at the location takes a value of 100. ; Number of independent annotators; For the first The annotation specialist is in the pixel Binary crack annotation at the location.

[0046] In one type of measured annotation, with Aspect mask union The pixel neighboring bands are the statistical range, and the average Dice coefficient between each pair of annotators is approximately Fleiss consistency coefficient Approximately ; in at least obtain In the pixels of the ticket, approximately Received a unanimous vote, Only obtained to The discrepancies are concentrated at blurred boundaries, hairline cracks, the junction of seams and shadows, and distant pixels in perspective. If... Direct majority voting yields a binary mask, but weak consensus and boundary transition information will be lost, therefore... Feeding a deterministic structure regularization operator .

[0047] S2-2, Gaussian smoothing.

[0048]

[0049] in, The consensus response after Gaussian smoothing; The standard deviation is equal to Two-dimensional Gaussian kernel; Represents two-dimensional convolution; For Gaussian smoothing, take to Pixels, one method of selection is Pixels are used to suppress isolated noise tickets.

[0050] S2-3, supports mask extraction and skeleton refinement.

[0051]

[0052] in, To support the mask; This is an indicator function; it takes the value when the condition is true. Otherwise take ; To support mask extraction threshold, take to One method is to select... The support mask is then refined to obtain a set of centerlines with a single pixel width. The thinning process can employ morphological thinning algorithms or median transformation. When the support mask is empty, let... And this concludes the construction of this image.

[0053] S2-4, Morphological Refinement.

[0054]

[0055] in, For morphological refinement of response; and These represent grayscale closing and grayscale opening operations, respectively. As a structural element, take to A square structural element, one method of taking is .

[0056] S2-5, Centerline distance attenuation.

[0057]

[0058] in, This is the centerline distance attenuation factor, used to reflect the attenuation of tubular cracks from the centerline to the boundary; For pixels Euclidean distance to the set of centerlines; As the centerline distance attenuation scale, take to Pixels, one method of selection is Pixel.

[0059] S2-6, Response Fusion.

[0060]

[0061] in, This is the intermediate support response after fusion; As the fusion weight for shape response and distance attenuation, take to One method is to select... .

[0062] S2-7, Fixed quantile normalization. This method satisfies the condition that the entire training set satisfies this condition. The quantile boundaries of the crack positive pixels are statistically normalized and fixed, and then cropped and normalized.

[0063]

[0064] in, For continuous crack support field; and These are the low-order and high-order normalized quantile boundaries, respectively, marked and fixed on the training set. to and to One method for determining quantiles is... , ; This indicates that the value will be truncated to... .

[0065] The support field It is a relative consensus support metric that can be compared between images, rather than a metric of the physical width or depth of the crack; both training and evaluation use it as a reference target. The above... Structural element dimensions , , Both the normalized quantile boundary and the normalized quantile boundary are fixed after being calibrated once on the training set. The validation set and the test set use the same set of values ​​without readjustment, thus ensuring that the threshold values ​​of each working point have consistent semantics across images.

[0066] Furthermore, the structure regularization operator This includes the following three processes, and the execution order of these three processes is not limited: Spatial smoothing is performed to suppress noise responses caused by isolated markers; the smoothed response is then weighted with a monotonically decreasing attenuation based on the distance from the pixel to the crack centerline, ensuring that the response of pixels at the crack core is higher than that of pixels at the crack edge; the weighted response is then linearly mapped and truncated to a normalized boundary that is calibrated and fixed once on the training set. To make different road images They have the same value caliber.

[0067] Furthermore, the structure regularization operator Implemented as follows: For Gaussian smoothing yields consensus responses. Extraction threshold based on support mask Take support mask The skeleton is then refined to obtain the set of centerlines. ;right The morphologically refined response is obtained by sequentially performing grayscale closing and grayscale opening operations. ;according to Calculate the centerline distance attenuation factor; according to the fusion weight. Receive intermediate support response Then, the quantile boundaries are statistically determined and fixed on the positive pixels of the cracks in the training set. and By performing truncation and normalization, we obtain ;when Empty time command The field is zero. All the above parameters are calibrated once on the training set and then fixed.

[0068] S3. Input the road image into the crack detection network, extract multi-level features by a multi-scale visual encoder, and output the basic segmentation response field by a soft decoder after fusing the multi-level features. , This is the unnormalized split output.

[0069] like Figure 3 As shown, the crack detection network includes a multi-scale visual encoder, a soft decoder, and a visual language calibration branch, the first two of which are described in this step.

[0070] S3-1, Multi-scale visual encoder. In one embodiment, the multi-scale visual encoder is a SegFormer-B2, with an output step size of... , , , The four-level feature maps, corresponding to the number of channels are respectively , , , The weights are initialized using weights pre-trained on a general semantic segmentation dataset (one option is the ADE20K dataset).

[0071] S3-2, Soft Decoder and Basic Segmentation Response Field. The soft decoder refers to a decoding section that projects feature maps at each level onto a unified channel width and performs multi-level fusion, then directly regresses the continuously valued gap support response on a single output channel. The output does not have a rigid binary decision stage, therefore its output can be directly used as the support strength on a continuous dimension. Specifically, the four-level feature maps are respectively processed... The convolution is projected onto a uniform decoding channel width, which is taken as... to One method is to select... Then upsample the projection features at each level to a step size of ; The resolution is calculated and fused step-by-step in a feature pyramid manner; the fused features are output through the prediction head. Another path is fed into the projection layer of the visual language calibration branch to obtain... .

[0072] in, Based on the segmented response field (the unnormalized segmented output, i.e., logits); These are decoder projection features, whose channel dimension matches the text embedding dimension. Without calibration, they can be used... To make a preliminary estimate, It is a sigmoid activation function.

[0073] S3-3, Training Cropping and Data Augmentation. During training, the road image is randomly cropped, and the cropping size is taken as... to One method is to select... and in the cropped image and the corresponding support field At least one of the following is applied simultaneously: random scaling, horizontal flipping, brightness and contrast perturbation, motion blur, and slight perspective perturbation.

[0074] S4. Based on a fixed visual language concept library, the visual language calibration branch of the crack detection network, using a pre-trained visual language image encoder and text encoder, dynamically selects positive and negative concept subsets from the positive and negative concept libraries of the fixed visual language concept library based solely on the appearance of the road image. It calculates the dense concept similarity of each selected concept on the road image and maps it to a single-channel visual language calibration correction field via a calibration head. ;Will and Fusion yields a continuous crack support prediction field. , .

[0075] like Figure 5 As shown, this step includes four stages: concept library construction, dynamic concept selection and caching, dense similarity calculation and pooling, and calibration head fusion. The visual language calibration branch focuses on suppression rather than seeking, and the final mask is still controlled by the soft decoder.

[0076] S4-1. Composition of the Concept Base. The concept base is a fixed text library that remains consistent throughout the training and inference phases. It includes a positive concept base and a negative concept base. The positive concept base includes road surface cracks, hairline cracks, longitudinal cracks, transverse cracks, branching cracks, network cracks, and repaired cracks. The negative concept base includes lane markings, road surface shadows, tire tracks, road surface joints, repair marks, asphalt textures, vehicle edges, curb edges, road edges, oil stains, and water stains. These correspond to actual crack-like interferences on the road surface and are used to identify the objects that need to be suppressed.

[0077] S4-2. Dynamic Selection and Caching of Concepts. The visual language encoder using frozen parameters, i.e., both the image encoder and the text encoder, do not participate in gradient updates. The image embedding is calculated after scaling the entire road image, and global similarity is calculated between it and the text embeddings of all concepts in the concept library. The one with the highest similarity is selected. A positive concept and the past Each negative concept constitutes a subset of the concepts in the current image; Pick to One method is to select one. , Pick to One method is to select one. Each sliding window cropping of the same image inherits the concept selection result of the entire image. The selection process is based solely on the color image and the fixed text library, without using the support field. Hard negative sample region annotations, validation set labels, test set labels, or current model predictions. During deployment, the entire image selection results are cached according to image labels.

[0078] S4-3, Dense Concept Similarity and Pooling. For each selected concept... Embed its text as After normalization and Calculate pixel-by-pixel cosine similarity.

[0079]

[0080] in, For the soft decoder at the pixel The decoder projection features output at the location; For pixels With concept The degree of close conceptual similarity between them; For the concept Text embedding; express Norm; This indicates transpose.

[0081] Furthermore, for the selected set of positive concepts With the selected set of negative concepts According to temperature coefficient Perform logarithmic and exponential pooling.

[0082]

[0083] in, and These are the positive concept pooling response field and the negative concept pooling response field, respectively. For the concept pooling temperature coefficient, ,Pick to In this embodiment, we take .

[0084] S4-4, Calibration Head and Response Fusion. , and After splicing along the channel dimensions, input the calibration head.

[0085]

[0086] in, A calibration correction field for visual language; The calibration head, in one embodiment, has two layers. convolution; This indicates splicing along the channel dimension.

[0087] Furthermore, the calibration correction field is added to the basic segmentation response field and then passed through a sigmoid activation function to obtain the continuous crack support prediction field.

[0088]

[0089] in, To support the predicted field for continuous cracks, the value is taken from... ; The learnable calibration scaling factor is initially set to... to One method is to select... The correction is used to suppress crack-like interference responses corresponding to the selected negative concept in the current image, while preserving support for real cracks.

[0090] S5, with the aforementioned support field The crack detection network is jointly trained for soft supervision purposes. The support field... The supervisory information mainly manifests in the relative order between pixels, the response attenuation from the centerline to the boundary, and the cutting semantics at each working point. Therefore, in addition to soft reconstruction constraints, relational constraints and cutting consistency constraints should also be applied. The training objective includes at least soft reconstruction constraints, structure-aware ranking constraints, and multi-working-point cutting constraints; wherein, the soft reconstruction constraints... and The structure-aware sorting constraint constrains the differences in values ​​at each pixel. and The relative order on the same pixel pair is consistent, and the multi-working-point cutting constraint is constrained. and The cutting results are consistent at at least two preset working point thresholds.

[0091] S5-1, Soft Reconstruction Constraint. This constraint is used to constrain... and The differences in values ​​across pixels can be represented in one form as follows.

[0092]

[0093] in, This is a soft reconstruction loss; and These are obtained over all pixels. Normative terms and Norm squared terms; The soft Dice is in product form, and its molecule is The denominator is This approach is relatively stable for classes with extreme imbalances and is used solely as a training objective. In another implementation, the soft reconstruction constraint may include only one or two of the three terms in the above formula.

[0094] S5-2, Structure-Aware Ordering Constraints. For example... Figure 4As shown, this constraint is used to constrain... The relative order between pixel pairs and the support field The relative order is consistent, including three categories: response level sorting, centerline-to-boundary sorting, and crack-to-crack interference sorting. At least two of these categories are applied during implementation. The response level sorting is based on... The value is used to divide the response partition: high response partition Response partition With low response partition ,in, , and The first, second, and third preset partition thresholds are, in order, the first, second, and third. ,make and A buffer zone that does not participate in pairing is left between them; and Between and and Sample pairs are constructed between these pairs, and a hinge loss with spacing is used to constrain the predicted values ​​of pixels in high-response partitions to be higher than those in low-response partitions. The spacing between the sample pairs is... With weight All follow The increase is due to the increase in the centerline to the boundary sorting constrains the predicted value of pixels closer to the centerline of the crack within the local crack neighborhood to be higher; the crack relative crack interference sorting constrains the predicted value of candidate crack pixels to be higher than the predicted value of interference-side pixels.

[0095] The interval and the weight are determined by the following formula:

[0096]

[0097] in, and For two pixels in the same sample pair, Based on the interval, This is the upper limit of the interval. and The adjustment coefficient is positive, and ;

[0098] The interfering object side pixels are taken from at least one of the hard negative sample region and the negative concept high response region; the hard negative sample region is the non-crack region on the road image of the training set after the region where the crack-like interfering object is located is marked and verified, and the negative concept high response region is the region determined according to the dense concept similarity and whose response to the selected negative concept is higher than a preset response threshold; each training clipping is sampled in layers according to a fixed quota. right , right There are four types of sample pairs: center line to boundary, crack to crack-like interference. When there are no usable pixels on the side of the interference during training cropping, the quota of the sample pair of the type is replenished to the other three types of sample pairs.

[0099] In this embodiment, ;in, , and These are the values ​​of the first, second, and third preset partition thresholds, respectively, and the values ​​for each are taken as follows: to , to and to The third preset partition threshold is less than the second preset partition threshold, and there is a buffer zone between them. Pixels falling into the buffer zone do not participate in the construction of sample pairs.

[0100] exist and between, and Sample pairs are constructed between them, and a hinge loss with gaps is used.

[0101]

[0102] in, To compensate for the loss due to hierarchical ordering; For a set of sample pairs, Its quantity; Indicated in the support field The above satisfies Sample pairs; and These are the interval and weight of the sample pair, respectively, and both increase as the difference in support strength between the two pixels increases.

[0103]

[0104] in, and These are the upper limits of the interval. Spacing with base One method of selection, and They are the interval slopes. With weighted slope One way to choose, and satisfying In one implementation method Pick to , Pick to .

[0105] Centerline-to-boundary sorting constructs centerline-to-boundary sample pairs within the local crack neighborhood, constraining the distance. Closer pixels achieve higher prediction values, and their spacing... Pick to One method is to select... The crack-to-crack interference ranking constructs crack and crack-like interference sample pairs between crack candidate pixels and interference side pixels, constraining the former's predicted value to be higher than the latter's, with a spacing between them. Pick to One method is to select... Weight Pick to One method is to select... The interfering pixel is taken from at least one of the hard negative sample region and the negative concept high response region confirmed during the training period, wherein the negative concept high response region is based on the region obtained in step S4-3. Confirm, i.e., take Those sorted at the top within the current cropping and Pixels not exceeding the third preset partition threshold. The total sorting loss is calculated using the following formula.

[0106]

[0107] in, For structure-aware ranking loss; and These are the centerline-to-boundary sorting loss and the crack-relative-crack-like interference sorting loss, respectively. and This is one way to choose the coefficient for the combination of the two.

[0108] To control computational load, each training clipping sample draws a fixed number of sample pairs using stratified sampling, for a total of [number missing]. to Yes, one way to choose is Yes, and they are allocated according to four categories: high response zone to medium response zone, medium response zone to low response zone, centerline to boundary, and cracks and crack-like interference objects. Yes; when there are no available hard negative sample regions during pruning and the negative concept high response regions are insufficient to fill the quota, the remaining quota of the fourth category will be proportionally replenished to the first three categories.

[0109] S5-3. Acquisition and Verification of Hard Negative Sample Regions. The hard negative sample regions used during training are obtained according to the following procedure. First, the annotator marks the areas containing crack-like interference objects such as lane markings, road shadows, road joints, repair marks, tire tracks, asphalt textures, vehicle edges, and curb edges on the training set images using rectangular or polygonal frames. The marked areas must meet the following requirements: The pixel sets do not intersect. The second step involves a second annotator reviewing each area individually. This review includes verifying the correctness of the category labeling of the bounded object, whether the bounding box boundary extends beyond the crack's neighborhood, and whether the bounded area truly exhibits a slender, high-contrast crack-like appearance. Areas failing the review are returned for re-annotation, and those with category disputes remaining after discussion are discarded. The third step involves recording the set of regions that pass the review as... The images are taken only from the training set and are used solely for sampling the interference-side pixels in step S5-2. It should be noted that the hard negative sample region annotation on the test set is completed independently of the training set and is used only for calculating the hard negative sample false response rate during the evaluation phase. It does not participate in training or in the concept selection in step S4-2.

[0110] S5-4, Multi-operating-point cutting constraint. For each preset operating point threshold... Micro-cutting is possible.

[0111]

[0112] in, For the operating point threshold to The result obtained by performing a differentially sized section; , and These are the sensitive, standard, and strict operating point thresholds, respectively, to meet the requirements. The three correspond to three levels of stringency, from low to high; the soft reconstruction constraint adopts... and Between item, At least one of the following: terms and product-form soft Dice terms; in this embodiment, respectively, are taken as follows: , and ; For the cutting temperature coefficient, take to One method is to select... .

[0113] Target cutting The threshold values ​​for each of the aforementioned operating points are calculated separately. and The multi-operation point cut loss is obtained by averaging the cross-entropy loss and Dice loss. The threshold values ​​used for each working point in this constraint are the same as those used for cutting in step S6. The threshold values ​​used for each working point are one-to-one and consistent, thus ensuring that the training target directly covers the actual judgment position used during inference. The crack detection network uses the actual working point threshold values ​​used in the inference phase as the cutting target during the training phase, and is composed of the same... Export crack detection results for each severity level without needing to train a separate segmentation head for each severity level.

[0114] S5-5, Curved Structure Topology Constraints and Gradient Consistency Constraints. Curved structure topology constraints apply to the standard operating point... and Apply a soft centerline Dice loss to constrain the consistency of the centerline connectivity between the two.

[0115]

[0116] in, For topological loss of curved structures; The overlap is based on the centerline. Gradient consistency constraints are taken as follows: , This is the spatial gradient operator.

[0117] S5-6, Total Loss and Training Settings. The total loss for joint training is as follows.

[0118]

[0119] in, For the total losses of the joint training, , , , and The loss terms, in order, are the soft reconstruction constraints, structure-aware sorting constraints, multi-operating-point cutting constraints, curve structure topology constraints, and gradient consistency constraints. to The weights are for five loss components: soft reconstruction, structure-aware ranking, multi-operating-point cutting, curve structure topology, and gradient consistency. All are positive numbers. to , to , to , to and to For example, this embodiment takes .

[0120] The optimizer uses AdamW; the learning rate is... to One method is to select... Weight decay The learning rate is adjusted using cosine annealing and several warm-up rounds are set. The total number of training rounds is taken as... to Wheel, one way to take it is Round, global batch size retrieval to One method is to select... Throughout the training process, the visual language encoder remains frozen, with only the multi-scale visual encoder, the soft decoder, and the calibration head remaining active. With the calibration scaling factor Participate in the update.

[0121] S6. Output during the reasoning phase and according to at least two preset working point thresholds The cracks were cut separately to obtain multi-level crack detection results.

[0122] S6-1, Sliding Window Prediction and Gaussian Weighted Fusion. A sliding window method is used for prediction on high-resolution full images. The sliding window size is consistent with the training cropping size, and the overlap width between adjacent windows is set to... to Pixels, one method of selection is Pixels; the prediction results of each window are weighted and fused using a two-dimensional Gaussian weight with the window center as the peak value. That is, the weighted prediction value is accumulated pixel by pixel and then divided by the accumulated weight value to obtain the whole image. The sliding windows of the same image share the concept selection result cached in step S4-2.

[0123] S6-2, Exporting results with multiple stringency and selecting operating points. For example... Figure 7 As shown, for the same According to , and The process involves cutting to obtain three binary detection results at sensitive, standard, and stringent operating points. The output at the sensitive operating point retains weak support and hairline crack evidence, suitable for wide-area screening; the output at the standard operating point strikes a balance between recall and precision, suitable for routine inspections; and the output at the stringent operating point retains high-confidence crack evidence, suitable for high-priority confirmation. All three are derived from the same... The output is optimized so that a separate segmentation head does not need to be trained for each strictness level when switching between strictness levels. The number of working points is not limited to three and can be adjusted according to business needs. and Add a working point threshold between them.

[0124] S6-3, Generation of Road Segment Maintenance Priorities. The detection results are divided into several road segments according to mileage or frame sequence. For each road segment... Generate maintenance priority information according to the following procedure. First, perform a connected component analysis on the cutting mask at the selected working point to obtain the set of connected components. For each connected component Refine the skeleton to obtain the skeleton, and count the lengths of the connected skeletons. Average width With average support strength Secondly, the regional crack index is calculated using the following formula.

[0125]

[0126] in, For road section The regional crack index, with values ​​ranging from ; It is the ratio of the sum of the lengths of all connected skeletons in the road segment to the total number of road surface pixels, after being truncated and normalized by the upper limit; This is the ratio of the total number of crack pixels to the total number of road surface pixels, truncated and normalized to the upper limit. and These are the average width weighted by the area of ​​the connected components, normalized after upper limit truncation, and the average support strength, respectively. to For the corresponding weights, satisfying One method is to take .

[0127] Finally, output priority according to the preset grading threshold: High maintenance priority is output at the same time. Maintenance priority in time output. When outputting low maintenance priority, among which One method is to , One method is to The weights and grading thresholds mentioned above are example values ​​and can be recalibrated according to the road network level and maintenance procedures during implementation. Furthermore, connected domains that still exist at strict operating points can be counted as high-confidence crack evidence, while connected domains that only appear at sensitive operating points can be counted as crack evidence to be observed. These two types of connected domains are then included in the regional crack index with different weights.

[0128] Furthermore, the structure-aware ranking constraint includes at least two of the following: response hierarchy ranking, centerline-to-boundary ranking, and crack-to-crack interference ranking. Each type of sample pair is sampled hierarchically according to a fixed quota and a quota replenishment mechanism is provided. The interference-side pixels in the crack-to-crack interference ranking are taken from at least one of the hard negative sample region confirmed by verification and the negative concept high response region determined according to the dense concept similarity, thereby enabling the visual language calibration branch and the structure-aware ranking constraint to cooperate in the suppression of crack-like interference.

[0129] Furthermore, the multi-scale visual encoder outputs at least four levels of feature maps with different strides. Each level of feature map is mapped to a uniform channel width via a projection layer, and then fused level by level at the highest resolution level using a feature pyramid approach to obtain a fused feature. The fused feature is then passed through the output layer of the soft decoder to obtain... The other path, through the projection layer of the visual language calibration branch, yields the decoder projection features. , The dense concept similarity in step S4 is used to calculate the similarity. During training, at least one of random scaling, flipping, brightness and contrast perturbation, motion blur, and perspective perturbation is applied to the training cropped from the road image. During the inference stage, sliding window prediction is used on the high-resolution road image, and Gaussian weighted fusion is performed on the overlapping areas of the windows.

[0130] like Figure 6 As shown, the road crack detection system that implements the above method includes the following modules, which are connected in sequence according to the input-output relationship, and can be implemented by the processor executing the computer program stored in the memory.

[0131] The multi-annotation acquisition module, corresponding to step S1, uses the acquired road image as input to organize... Annotators independently complete full-resolution annotations on a blank canvas, rejecting unqualified annotations and requiring re-annotation, and outputting a set of binary crack masks.

[0132] The continuous crack support field construction module corresponds to step S2, which uses the mask set as input to calculate... And call the structure regularization operator. The process sequentially performs Gaussian smoothing, support mask extraction and skeleton refinement, morphological refinement, centerline distance decay, response fusion and fixed quantile normalization. When the support mask is empty, a zero field is output, and the support field is output. .

[0133] The basic segmentation response prediction module, corresponding to step S3, takes the road image as input, and extracts feature maps of at least four different scales by the multi-scale visual encoder. These feature maps are then projected onto a unified channel width by the soft decoder and fused using a feature pyramid method, resulting in the output. and .

[0134] The visual language calibration module corresponds to step S4, using the road image and... For input, dynamic selection A positive concept and Negative concepts, calculation And perform logarithmic and exponential pooling, through the calibration head Output and Output after fusion Furthermore, the concept selection results are cached as a whole graph.

[0135] The joint training module corresponds to step S5, with and As input, manage stratified sampling and quota replenishment of sorted sample pairs, maintain the verified hard negative sample region, and calculate the total loss. Optimize the crack detection network.

[0136] The multi-working-point output module corresponds to step S6. It takes the road image to be detected and the trained crack detection network as input, performs sliding window prediction and Gaussian weighted fusion, and outputs the result after segmentation according to the threshold of each working point. Results of multi-rigidity crack detection.

[0137] The maintenance decision interface module corresponds to step S6-3, using the cutting mask at the selected working point and... Using the connected skeleton length, average width, and region area density as input, the region crack index is obtained by weighted summation. It outputs maintenance priority information for road sections in three levels: high, medium, and low, according to preset threshold values.

[0138] In another implementation, the multi-scale visual encoder can be replaced by other segmentation encoders that can output multi-level, asynchronous, long feature maps, including convolution-based hierarchical encoders and self-attention-based hierarchical encoders.

[0139] In another implementation, the visual language encoder can be replaced with other pre-trained models that can simultaneously provide image embeddings and text embeddings in the same embedding space, or the image encoder can be subjected to low-rank adaptation while keeping the text encoder frozen before use.

[0140] In another implementation, the concept library can be expanded according to road surface type: for cement concrete pavement, it can add joint board edges, joint sealing tape and misalignment; for pavement inside tunnels, it can add water seepage marks and light strip reflections; for pavement in winter, it can add snow accumulation edges and ice marks.

[0141] In another implementation, when the number of available annotators... When insufficient, it can be compensated for by extending the annotation and review process; it can also be... In the case of single-person annotation results being processed by the same structure regularization operator, Construct the support field At this point, the centerline distance attenuation and fixed quantile normalization still hold true, and the centerline to boundary sorting and crack relative crack-like interference sorting can still be applied. Only the consensus strength information on which the response hierarchy sorting is based is reduced accordingly.

[0142] In another implementation, the number of working points can be more than three, and a working point threshold can be added between the sensitive working point threshold and the strict working point threshold and incorporated synchronously in step S5-4 to maintain a one-to-one correspondence between training and inference at the working points.

[0143] In another implementation, the structure regularization operator The exponential distance decay can be replaced by other functions that monotonically decrease with respect to the centerline distance, including linear decay functions and truncated quadratic decay functions; the thinning process can be replaced by other skeleton extraction methods that can obtain a single-pixel wide centerline. The various implementation methods described above can be implemented individually or in combination.

[0144] The following is a specific embodiment. The values ​​in this embodiment are from experiments implemented according to the method described in this specification, and are only used to illustrate the feasibility and effectiveness of the method, and do not constitute a limitation on the scope of protection.

[0145] This embodiment uses A series of road images from a vehicle's perspective, with a resolution of [resolution value missing]. According to the agreement , and Divided into training set Amplitude, Verification Set Aspect Ratio and Test Set Each image is composed of The designated annotators independently annotate according to the annotation protocol in step S1. The structure regularization operator... Pick , , structural elements , , The normalized quantile boundary is labeled on the positive pixels of the crack in the training set as follows: and And fixed. A separate, independent audit subset is established, which includes... The test images obtained by stratified sampling are only used for human-machine consistency verification and are not used for training and selection.

[0146] The multi-scale visual encoder uses SegFormer-B2, the visual language encoder uses CLIP ViT-L / 14 with frozen parameters, and the calibration head is two-layered. Convolution. Training clipping size is... The optimizer is AdamW, and the learning rate is... ,train Round, global batch size is Working point threshold is taken as follows , , , Loss weighting The sorted samples were cut for each total And each of the four categories Yes; concept selection results are cached as a whole graph, and sliding windows overlap during inference. Pixel.

[0147] The five evaluation indicators used in this embodiment are defined as follows. Soft Dice index and The overlap is calculated in fuzzy set form. During the training phase, a product-based soft Dice is used, while during the evaluation phase, a smaller or larger soft Dice is used.

[0148]

[0149] in, For evaluation purposes, use the soft Dice; This is a pixel-by-pixel subtraction operation; summation is performed on all pixels; higher values ​​indicate a greater degree of overlap.

[0150] Alpha mDice refers to the threshold at each working point. Each place is respectively for and The Dice coefficient is calculated after segmentation, and then the average value is obtained. Spearman refers to the Spearman rank correlation coefficient, i.e. and The rank correlation coefficient on sampled pixels is used to characterize the degree of consistency between the two in terms of relative pixel order. clDice refers to the overlap based on the centerline, which is the harmonic mean of the proportions of the skeletons of the two cut results covered by the mask of the other, used to characterize the connectivity of the curve structure. HN-FR refers to the hard negative sample false response rate, that is, the mean of the predicted support response within the hard negative sample region of the test set.

[0151]

[0152] in, This refers to the hard negative sample region on the test set that has been verified. This represents the number of pixels in the area; a lower value indicates more effective suppression of crack-like interference.

[0153] The comparison methods include hard-label general segmentation baselines, soft-label general segmentation baselines, gap-specific networks, probabilistic segmentation baselines, and cue-based segmentation and visual-language segmentation methods. Table 1 presents the main results on the test set, where the values ​​for trainable methods are the mean values ​​under multiple random seeds.

[0154] Table 1 Comparison of key metrics for the test set

[0155]

[0156] As shown in Table 1, the method of this invention outperforms the other methods listed in Table 1 in all four indicators: soft Dice, Alpha mDice, Spearman, and HN-FR. Its soft Dice is... Alpha mDice is Spearman is HN-FR is .

[0157] Regarding the selection of working points, the output under strict working points focuses on high-confidence crack evidence, the output under sensitive working points focuses on weak support and hairline crack evidence, and the output under standard working points strikes a balance between the two. During implementation, the appropriate working point can be selected from large-scale screening, routine inspection and high-priority confirmation according to the inspection target, without having to use the result under a single threshold as the only criterion.

[0158] Table 2 shows the results after removing a constraint item by item from the complete configuration.

[0159] Table 2 Leave-one-out ablation in complete configuration

[0160]

[0161] As shown in Table 2: After removing the structure-aware ranking constraint, the soft Dice... Down to Spearman Down to After removing the visual language calibration branch, HN-FR is... Rise to And the soft Dice remains at This indicates that the function of this branch is to suppress crack-like interference responses, rather than to increase overall overlap; after removing the topological constraints of the curve structure, clDice is... Down to After removing the multi-working-point cutting constraint, Alpha mDice is... Down to It is evident that the roles of each constraint term in the same continuous crack support prediction field are divided among themselves. Furthermore, in a comparison of reconstructing the support field with different numbers of annotators, the number of annotators increased from... The number of names has increased to When naming, in contrast to The ranking relevance and weak consensus region performance improved as a result of the annotation results of renowned annotators were used as a reference.

[0162] In this embodiment , , Structural element dimensions , , , , , , to as well as , , The values ​​are all examples and do not constitute a limitation on the scope of protection. The multi-scale visual encoder is not limited to SegFormer-B2, and the visual language encoder is not limited to a specific contrastive learning model, as long as it can provide multi-level image features and conceptual similarity between images and text respectively.

[0163] Those skilled in the art will understand that all or part of the above process can be completed by computer program instructions to related hardware. The computer program can be stored in a non-volatile computer-readable storage medium such as a read-only memory, flash memory, disk, or optical disk. When the computer program is executed by a processor, it implements the steps of the above method.

[0164] The above embodiments are merely several implementations of the present invention and should not be construed as limiting the scope of patent protection of the present invention. Any changes and improvements made without departing from the concept of the present invention are within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A road crack detection method based on multi-person soft consensus and visual language calibration, characterized in that, include: Obtain road images and binary masks independently annotated by multiple annotators, calculate the original voting consensus, and obtain a continuous crack support field through structural regularization; The road image is input into the crack detection network. The basic segmentation response field is output by multi-scale visual encoding and soft decoding. Then, the pre-trained visual language encoder dynamically selects the positive crack concept and the negative crack-like concept according to the image, generates the calibration correction field and fuses it to obtain the continuous crack support prediction field. During training, soft reconstruction, structure-aware ranking, and multi-operating-point cutting constraints are jointly applied on the same prediction field. During inference, the results are cut according to at least two preset working point thresholds to obtain multi-rigidity detection results.

2. The method according to claim 1, characterized in that, acquire a road image, and acquire a set of binary crack masks independently labeled by two annotators for the same road image wherein, the first binary crack label of the first annotator at a pixel ​ The road images are captured by a vehicle-mounted camera device from a vehicle-mounted perspective. The captured road images cover perspective distortion, lighting changes, road surface shadows, lane markings, repair areas, road surface joints, asphalt textures, vehicle occlusion, tire tracks, and road edge boundaries. Each annotator independently completes annotations on a blank canvas without automatic proposals or preset masks at full resolution, allowing for local zoom-in. Pixels showing evidence of cracks are annotated as 1, and the remaining pixels are annotated as 0. Annotations that are incomplete or whose annotation positions are significantly off-center are rejected and re-annotated.

3. The method according to claim 1, characterized in that, The raw voting consensus score is calculated as follows : ; wherein, is a pixel position, is an annotator index; for applying a structure regularizer , resulting in a continuous crack support field , ; the structure regularizer at least comprises a monotonically decreasing shaping of the consensus response based on the distance of a pixel to a crack centerline, and a normalization of the shaped response by a boundary that is calibrated and fixed on the training set; The structure regularization operator is implemented as follows: to , a Gaussian smoothing is performed to obtain a Gaussian smoothed consensus response ; a support mask extraction threshold is determined according to the following formula ; a support mask is extracted according to the support mask extraction threshold , is an indicator function, which takes when the condition is true, and takes otherwise, and to , a skeleton thinning is performed to obtain a centerline set ; to , a gray-scale closing operation and a gray-scale opening operation are sequentially performed to obtain a morphological refinement response ; a centerline distance decay factor is calculated according to the following formula : ; in, For pixels to Euclidean distance, The attenuation scale is based on the centerline distance; according to the fusion weight. Obtain the fused intermediate support response ;Statistically count and fix the normalized quantile boundaries on all positive pixels of the crack in the training set. and The crack positive pixel is satisfied The pixels; then obtain the following formula : ; When is zero field; the scale of the Gaussian smoothing is zero field; the scale of the Gaussian smoothing , the structure element size of the gray scale morphological operation, , , and the quantile boundary are all fixed after being calibrated on the training set once.

4. The method according to claim 1, characterized in that, The road image is input into a crack detection network, multi-level features are extracted by a multi-scale visual encoder, and a basic segmentation response field is output after the multi-level features are fused by a soft decoder , is the unnormalized segmentation output.

5. The method according to claim 1, characterized in that, By the visual language calibration branch of the crack detection network, a pre-trained visual language image encoder and a text encoder are used to dynamically select a positive concept subset and a negative concept subset from a fixed visual language concept library, only according to the appearance of the road image, calculate the dense concept similarity of each selected concept on the road image, and map it to a single-channel visual language calibration correction field through a calibration head ; and fused with to obtain a continuous crack support prediction field , ; The positive concept library consists of text descriptions of seven positive concepts: road surface cracks, hairline cracks, longitudinal cracks, transverse cracks, branching cracks, network cracks, and repaired cracks. The negative concept library consists of text descriptions of eleven negative concepts: lane markings, road surface shadows, tire tracks, road surface joints, repair marks, asphalt texture, vehicle edges, curb edges, road edges, oil stains, and water stains. Each negative concept refers to a type of crack interference. The parameters of the visual language image encoder and the text encoder are frozen during the training of the crack detection network. The positive concept subset and the negative concept subset are selected in the following manner: global similarities between image embeddings of the entire road image and text embeddings of all concepts in the concept library are calculated, and the top positive concepts and the top negative concepts are selected from high to low according to the global similarities , and ; the selection is only based on the road image and the concept library; the selection results are cached according to the entire image, and are inherited and shared by each sliding window cropped from the road image. For each concept selected with its text embedded do normalize, and compute the dense concept similarity as : ; in, For the soft decoder at the pixel The decoder projection features at the output; the selected positive concept set and the selected negative concept set are respectively calculated using the concept pooling temperature coefficient. Logarithmic and exponential pooling are performed to obtain the positive concept pooling response field. With negative concept pooling response field , Calibration head by , and The tensor obtained by concatenating the channels is used as input, and the output is a single-channel tensor. ; according to the following formula and Fusion: ; in, It is a sigmoid activation function. This refers to the calibration scaling factor trained along with the crack detection network.

6. The method according to claim 1, characterized in that, Support field with continuous cracks The crack detection network is jointly trained with a soft-supervised objective, and the training objective includes at least soft reconstruction constraints, structure-aware ranking constraints, and multi-operating-point cutting constraints; wherein, the soft reconstruction constraints... and The structure-aware sorting constraint constrains the differences in values ​​at each pixel. and The relative order on the same pixel pair is consistent, and the multi-working-point cutting constraint is constrained. and The cutting results are consistent at at least two preset working point thresholds; The structure-aware sorting constraints include at least two of the following: response hierarchy sorting, centerline-to-boundary sorting, and crack-to-crack interference sorting; the response hierarchy sorting is based on... The value of divides the pixel into high-response partitions. Response partition With low response partition ,in, , and The first, second, and third preset partition thresholds are, in order, the first, second, and third. ,make and A buffer zone that does not participate in pairing is left between them; and Between and and Sample pairs are constructed between these pairs, and a hinge loss with spacing is used to constrain the predicted values ​​of pixels in high-response partitions to be higher than those in low-response partitions. The spacing between the sample pairs is... With weight All follow The increase is due to the increase in the centerline to the boundary sorting constrains the predicted value of pixels closer to the centerline of the crack within the local crack neighborhood to be higher; the crack relative crack interference sorting constrains the predicted value of candidate crack pixels to be higher than the predicted value of interference-side pixels. The interval and the weight are determined by the following formula: ; in, and For two pixels in the same sample pair, Based on the interval, This is the upper limit of the interval. and The adjustment coefficient is positive, and ; The interfering object side pixels are taken from at least one of the hard negative sample region and the negative concept high response region; the hard negative sample region is the non-crack region on the road image of the training set after the region where the crack-like interfering object is located is marked and verified, and the negative concept high response region is the region determined according to the dense concept similarity and whose response to the selected negative concept is higher than a preset response threshold; each training clipping is sampled in layers according to a fixed quota. right , right There are four types of sample pairs: center line to boundary, crack to crack-like interference. When there are no usable pixels on the side of the interference during training cropping, the quota of the sample pair of the type is replenished to the other three types of sample pairs. The at least two preset operating point thresholds include a sensitive operating point threshold. Standard operating point threshold With strict operating point threshold , The three correspond to three levels of stringency, from low to high; the soft reconstruction constraint adopts... and Between item, At least one of the terms and product-form soft Dice terms; the multi-operating-point cutting constraint is applied according to the following formula. The structure is micro-divable: ; in, It is a sigmoid activation function. The cutting temperature coefficient is denoted by . For target segmentation, the threshold values ​​for each working point are calculated separately. and The classification loss and Dice loss are averaged and used as the loss term of the multi-operating-point cutting constraint. The operating point thresholds used when applying the multi-operating-point cutting constraint correspond one-to-one with the operating point thresholds used when cutting the continuous crack support prediction field during the inference stage, and have the same value. This allows the crack detection network to use the operating point thresholds actually used in the inference stage as the cutting target during the training stage, and to derive crack detection results of each strictness from the same continuous crack support prediction field, without the need to train a segmentation head separately for each strictness.

7. The method according to claim 6, characterized in that, The training objective also includes curve structure topology constraints and gradient consistency constraints; the curve structure topology constraints are... to and A soft centerline Dice loss is applied to constrain the consistency of the two in terms of centerline connectivity; the gradient consistency constraint is taken as... The total loss of the joint training includes at least the following weighted combination: ; in, , , , and The loss terms, in order, are the soft reconstruction constraints, structure-aware sorting constraints, multi-operating-point cutting constraints, curve structure topology constraints, and gradient consistency constraints. to These are the weights of each loss term, and all are positive numbers.

8. A method for detecting road cracks, characterized in that, include: Acquire the image of the road to be detected; The road image is input into a trained crack detection network, whose multi-scale visual encoder and soft decoder output a basic segmentation response field. Its visual language calibration branch calculates the similarity of dense concepts based on positive and negative concepts selected from a fixed visual language concept library, and generates a single-channel visual language calibration correction field via the calibration head. Then and Fusion yields a continuous crack support prediction field. ; According to at least two preset working point thresholds Cut the cracks separately and output the multi-strictness crack detection results that correspond one-to-one with the threshold values ​​of each working point; The trained crack detection network is a continuous crack support field obtained according to the method of claim 1. It is trained as a soft-supervised target, according to the training target defined by the method of claim 1.

9. A road crack detection system based on multi-user soft consensus and visual language calibration, characterized in that, The system for implementing the method according to any one of claims 1 to 8, the system comprising: The multi-annotation acquisition module is used to acquire road images and obtain the annotations from... The annotation module obtains a set of binary crack masks by annotators independently annotating the same road image; the multi-annotation acquisition module returns and re-annotates annotation results that are incomplete or whose annotation positions are significantly off-target, and stores and outputs the road image that has passed the verification and the set of binary crack masks according to the image index; A continuous crack support field construction module is used to receive the binary crack mask set, calculate the original voting consensus score per pixel, and then... After applying the structure regularization operator, the output is a continuous crack support field. The structure regularization operator includes at least a monotonically decreasing shaping of the consensus response based on the distance from the pixel to the center line of the crack, and normalization based on the boundary calibrated and fixed on the training set. The basic segmentation response prediction module receives the road image, extracts at least four levels of feature maps with different strides using a multi-scale visual encoder of the crack detection network, and then merges these feature maps by a soft decoder of the crack detection network after mapping them to a uniform channel width, outputting an unnormalized basic segmentation response field. With decoder projection features and will Send to the visual language calibration module; The visual language calibration module is used by a pre-trained visual language image encoder and text encoder to dynamically select positive and negative concept subsets from a fixed visual language concept base based solely on the appearance of the road image. Calculate the dense concept similarity with the selected concept, and output a single-channel visual language calibration correction field via the calibration head. Then combine it with Fusion, outputting continuous cracks to support the predicted field ; Joint training module for continuous crack support field The crack detection network is trained for a soft-supervised objective, the training objective of which includes at least constraints. and Soft reconstruction constraints and constraints for value differences and Structure-aware sorting constraints based on the relative order of pixel pairs, and constraints and The network parameters after training are output by multi-operating-point cutting constraints where the cutting results are consistent at at least two preset operating-point thresholds. The multi-working-point output module is used to load the trained network parameters, predict the road image window by window in a sliding-window manner, and perform Gaussian weighted fusion on the overlapping areas of adjacent windows to obtain the full image. Then, according to at least two preset working point thresholds, The cracks are cut separately, and the multi-strictness crack detection results are output, which correspond one-to-one with the threshold values ​​of each working point.

10. The system according to claim 9, characterized in that, The visual language calibration module caches the selection results of the positive concept subset and the negative concept subset on the entire road image, and makes the cached selection results shared by each sliding window crop of the same road image; the system also includes a maintenance decision interface module, which performs connected component analysis on the mask cut from the multi-strictness crack detection results under the selected working point threshold, obtains the regional crack index by weighted summation of road segment-level statistics including connected skeleton length, average crack width and regional area density, and outputs the results after dividing each road segment into high, medium and low maintenance priorities according to the preset grading threshold.