A deep learning-based concrete structure detection method

Abstract

This invention discloses a deep learning-based method for detecting concrete structures, comprising the following steps: acquiring images or video frames from the same area to form an observation dataset; establishing a set of defect hypotheses and constructing evidence items; constructing an observation action primitive library; calculating scores and performing quality gating using an observation quality scoring network based on MobileNetV3-Small; obtaining candidate defect instances through coarse detection; generating candidate regions and type distributions and forming multi-view packages; generating observation sequences and determining trigger states under time budget and reshoot count budget constraints; inputting an improved DUSt3R and outputting a structured evidence package upon triggering; fusing the type distribution and evidence package to determine if the criteria are met and outputting the results; updating the observation sequence if the criteria are not met; generating evidence data packages and updating the primitive selection strategy. This invention improves detection reliability, measurement consistency, and result traceability, and is suitable for intelligent detection scenarios of concrete structure defects.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for concrete structures, and in particular to a method for inspecting concrete structures based on deep learning. Background Technology

[0002] With the continuous growth of infrastructure stock and the increasing demand for refined structural safety management, technologies for the detection and quantitative assessment of apparent defects in concrete structures such as bridges, tunnels, and buildings have received widespread attention. Existing technologies mostly employ single-view image target detection or semantic segmentation networks to identify defects such as cracks and spalling, approximating indicators such as length and area using pixel statistics, or using simple thresholds and rules to make judgments and suggest re-images after detection. However, in practical engineering applications, the following problems are commonly encountered: The acquired images are significantly affected by blurring, overexposure, reflection, and occlusion, lacking effective observation quality quantification and gating mechanisms, resulting in unstable input quality and causing false positives and false negatives. Multi-view observation data lacks instance-level grouping and consistency constraints, making it difficult to stably associate the same defect across different views and frames, resulting in inconsistent results across views. Existing methods mostly rely on two-dimensional texture features and lack robust alignment and verification methods based on three-dimensional geometry, making it difficult to distinguish false defects such as template textures and water stain shadows. The measurement process often lacks scale anchoring and credibility modeling, resulting in insufficient reliability of millimeter-level measurement results such as crack width and peeling depth. Furthermore, it lacks a closed-loop process for determining evidence threshold compliance and triggering re-shooting, making it difficult to meet the requirements of traceable and re-inspectable detection and evaluation in engineering.

[0003] Therefore, how to provide a deep learning-based method for detecting concrete structures is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based method for detecting concrete structures. This invention fully utilizes observation quality scoring and quality gating based on MobileNetV3-Small, multi-view packetization, and reshoot sequence generation driven by the expected gain of observation action primitives, as well as 3D point map regression, confidence robust alignment, geometric evidence verification, and scale measurement output of the improved DUSt3R. This enables closed-loop detection of defect candidate generation, 3D structured evidence package construction, and evidence threshold compliance determination. When the threshold is not met, it triggers observation sequence updates and strategy self-updates. It has the advantages of high detection reliability, strong measurement consistency, low false alarms and false negatives, and traceable results.

[0005] A deep learning-based concrete structure detection method according to an embodiment of the present invention includes the following steps: Step 1: Collect images or video frames of the same concrete structure area and form an observation dataset, establish a set of defect hypotheses and set evidence threshold parameters and measurement requirements for each defect type, and construct evidence items; Step 2: Construct a primitive library of observation actions and set parameter ranges and safety boundaries. Use an observation quality scoring network based on MobileNetV3-Small to calculate the observation quality score of the observation dataset and perform quality gating. Step 3: Perform coarse detection on the observation dataset to obtain a set of candidate defect instances, generate candidate regions and type distributions for each candidate defect instance, and form a multi-view package; Step 4: Calculate the expected gain of the observation action primitive on the evidence item and generate the observation sequence under the constraints of time budget and retake number budget to determine the trigger state of the 3D evidence generation process; Step 5: In the triggered state, input the multi-view package into the improved DUSt3R model. The improved DUSt3R model includes a dot plot regression backbone module, a confidence robust alignment module, a geometric evidence verification module, and a scale measurement output module, and outputs a structured evidence package. Step Six: Combine the evidence type distribution with the evidence items in the structured evidence package to determine if the evidence threshold is met and output the detection results. If the threshold is not met, update the observation sequence based on the trigger signal. Step 7: Generate and save evidence data packages containing multi-view groups, structured evidence packages, and detection results, and update the selection strategy of observation action primitives based on the evidence data packages.

[0006] Optionally, step one specifically includes: Images or video frames of the same concrete structure area are collected and formed into an observation dataset. A uniform spatial coverage range is set for the same detection object. The acquisition time, shooting distance estimate and viewing angle information of each image or video frame are recorded. The images or video frames are grouped according to time order and viewing angle differences and written into the observation dataset. Establish a set of defect hypotheses, using the defect types on the concrete structure surface as the hypothesis categories, assign a defect type identifier to each defect type and generate a corresponding defect hypothesis record; Set corresponding evidence threshold parameters for each defect type, and set measurement requirements that match the defect type; An evidence item is constructed, which is defined as a structured evidence record established for the same defect hypothesis. The evidence item consists of texture evidence, geometric evidence, multi-view consistency evidence, and measurement credibility evidence.

[0007] Optionally, step two specifically includes: A library of observation action primitives is constructed, which are defined as acquisition action units that perform re-shooting and re-shooting on the same concrete structure area. The executable parameter range and safety boundary are set for each observation action primitive. An observation quality scoring network based on MobileNetV3-Small is constructed. MobileNetV3-Small is used as the feature extraction backbone network. A global average pooling layer, a first fully connected layer and a second fully connected layer are sequentially connected at the output of the backbone network to form a scoring head. Train the observation quality scoring network, establish a training sample set and label each training sample with blur label, overexposure label, reflection ratio label and occlusion ratio label, use the mean square error between the four-dimensional scoring vector and the corresponding label as the loss function, and use backpropagation to update the MobileNetV3-Small backbone network parameters and scoring head parameters. To calculate the observation quality score, each image or video frame in the observation dataset is subjected to input size alignment, pixel normalization, and channel normalization before being input into the observation quality score network. The output scores are then weighted and fused according to preset weights to obtain the observation quality score. Perform quality gating by comparing the observation quality score with a preset gating threshold parameter. If the observation quality score is lower than the gating threshold parameter, select a retake action from the observation action primitive library and re-acquire the corresponding image or video frame and write it into the observation dataset. If the observation quality score is not lower than the gating threshold parameter, retain the corresponding image or video frame.

[0008] Optionally, step three specifically includes: Perform size alignment, pixel normalization, and channel normalization on the images or video frames in the observation dataset; A coarse detection network is constructed, consisting of a backbone network, a neck network, and a detection head. The backbone network is composed of convolutional units formed by sequentially connecting convolutional layers, batch normalization layers, and ReLU activation layers. Multiple convolutional units are sequentially connected to form the backbone network. At the end of each stage, a residual connection layer is set to output a stage feature map, resulting in a three-layer stage feature map. The neck network uses a feature pyramid structure to perform top-down and bottom-up path fusion on the three-layer stage feature maps, aligning the channel dimensions during the fusion process to generate a fused feature map. The detection head includes a classification branch and a regression branch, outputting candidate region confidence and candidate region bounding box parameters. A coarse detection network is used to detect candidate regions in the processed image or video frame, and the candidate regions, candidate region confidence scores and candidate region bounding box parameters are obtained. The candidate region is instantiated, and its spatial location, scale information, confidence level, and bounding box parameters are written into the candidate defect instance set. A candidate defect instance identifier is assigned to each candidate defect instance. Input the candidate region corresponding to the candidate defect instance into the defect classification head, and output the category confidence vector of crack, peeling, exposed reinforcement, honeycomb surface, leakage and dampness marks. Write the category confidence vector as the type distribution into the candidate defect instance set. Based on the acquisition time and viewpoint information recorded in the observation dataset, candidate defect instances are correlated across frames, and candidate regions with overlapping spatial locations are grouped into the same package identifier to form a multi-view package.

[0009] Optionally, step four specifically includes: Set four-dimensional gain weights for each observation action primitive in the observation action primitive library; For each candidate defect instance in the candidate defect instance set, the coverage index and the difference index are calculated based on multi-view package, the measurement complementation index is obtained according to the measurement requirements, and the observation quality score is read. For each observation action primitive, the total expected gain is calculated. The total expected gain is obtained by weighting the coverage index and the difference index according to the gain weight corresponding to geometric evidence, the measurement complementation index according to the gain weight corresponding to measurement credibility evidence, and the observation quality score according to the gain weight corresponding to texture evidence and multi-view consistency evidence. The observation action score is obtained by multiplying the total expected gain by the observation quality score. Under the constraints of time budget and reshoot time budget, the observation action primitive with the highest observation action score is selected to generate the observation sequence, and the trigger state of the three-dimensional evidence generation process is determined based on the observation sequence.

[0010] Optionally, the improved DUSt3R model includes a dot plot regression backbone module, a confidence robust alignment module, a geometric evidence verification module, and a scale measurement output module. The dot plot regression backbone module performs input size alignment and pixel normalization on each view image within the multi-view group, and inputs the aligned image into the feature encoding network to obtain a feature map. The feature encoding network connects the dot plot regression head and the confidence regression head at the end. The dot plot regression head outputs three-dimensional coordinates pixel by pixel to form a dot plot, and the confidence regression head outputs confidence values ​​pixel by pixel to form a point-level confidence map. The confidence-robust alignment module receives point maps and point-level confidence maps from each viewpoint, linearly scales the point-level confidence maps to obtain an alignment weight map, normalizes the alignment weight map, and performs pixel-by-pixel weighted summation on the point maps based on the alignment weight map to obtain an initial global point map. The confidence-robust alignment module performs robust suppression processing on the residuals between the initial global point map and the point maps from each viewpoint to obtain the global point map and outputs the alignment quality index. The geometric evidence verification module receives the global point map and generates geometric evidence, which includes geometric discontinuity intensity, normal abrupt change and curvature anomaly. It performs multi-view consistency verification by back-projecting the global point map to each viewpoint and matching it pixel by pixel with the corresponding viewpoint point map. Based on the matching residual, the consistency weight is obtained by exponential decay and the consistency weight is averaged to obtain the multi-view consistency score. The drift amount is obtained by averaging the pixel displacement of the matching residual. When the multi-view consistency score is lower than the preset consistency threshold or the drift amount is higher than the preset drift threshold, a trigger signal is output. The scale measurement output module receives the global point map, geometric evidence, and measurement requirements, performs scale anchoring on the global point map, and outputs millimeter-level measurement results and measurement credibility evidence.

[0011] Optionally, step six specifically includes: Read the type distribution of candidate defect instances, and determine the corresponding evidence threshold parameters and measurement requirements according to the defect type with the highest category confidence in the type distribution; Read the evidence items and trigger signals in the structured evidence package, and convert the evidence items into texture evidence scores, geometric evidence scores, multi-view consistency evidence scores and measurement credibility evidence scores, respectively. The evidence threshold is determined by weighted fusion of texture evidence score, geometric evidence score, multi-view consistency evidence score and measurement confidence evidence score to obtain evidence compliance score. The evidence compliance score is then compared with the evidence threshold parameter. A detection result is generated when the evidence compliance score is not less than the evidence threshold parameter. When the evidence compliance score is less than the evidence threshold parameter, the observation sequence is updated according to the trigger signal, the multi-view group is updated according to step three, the structured evidence group is updated according to step five, and the evidence threshold compliance judgment is performed again.

[0012] Optionally, step seven specifically includes: The multi-view package, structured evidence package, and detection results are encapsulated to generate an evidence data package, and an evidence data package identifier is assigned to the evidence data package; Write the package identifier, the image or video frame index and view information contained in the package into the evidence data package, write the alignment quality index, multi-view consistency score, trigger signal and millimeter-level measurement results into the structured evidence package, and write the defect type and candidate region location into the detection results. The correspondence between the trigger signals in the statistical evidence data package and the actual observed action primitives is calculated. The number of times the standard is passed after each observed action primitive is triggered is counted, and the four-dimensional gain weight is updated based on the correspondence and the number of passes.

[0013] The beneficial effects of this invention are: This invention achieves end-to-end controllability and traceability from data acquisition to conclusion output through an overall process of "observation quality scoring—quality gating—coarse detection instantiation and multi-view packaging—observation sequence generation driven by expected gain of observation action primitives—improved DUSt3R 3D evidence generation—evidence threshold compliance determination—trigger reshooting closed loop": The observation quality scoring network based on MobileNetV3-Small suppresses false positives and false negatives caused by low-quality inputs such as blurring, overexposure, reflection, and occlusion at the source; multi-view packaging and trigger status determination prioritize observation resources to areas with insufficient evidence within the time and reshooting time budgets; the improved DUSt3R improves 2D texture discrimination to 3D geometric evidence and millimeter-level measurement through point map regression, confidence robust alignment, geometric evidence verification, and scale measurement output, and forms a structured evidence package by combining multi-view consistency and measurement credibility; evidence threshold compliance determination and trigger signal-driven observation sequence updates further ensure that the detection results stably meet the standards under complex field conditions. The method of this invention significantly improves the robustness and cross-perspective consistency of concrete defect identification, enhances the reliability and repeatability of millimeter-level measurements, and meets the key requirements of engineering inspection for reliable judgment, closed-loop evidence collection and standardized delivery. It has important engineering application value and promotion significance. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based concrete structure detection method proposed in this invention; Figure 2 This is a schematic diagram of a deep learning-based concrete structure detection method proposed in this invention. Figure 3 This is a framework diagram of the improved DUSt3R model in the deep learning-based concrete structure detection method proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figures 1-3A deep learning-based method for detecting concrete structures includes the following steps: Step 1: Collect images or video frames of the same concrete structure area and form an observation dataset, establish a set of defect hypotheses and set evidence threshold parameters and measurement requirements for each defect type, and construct evidence items; Step 2: Construct a primitive library of observation actions and set parameter ranges and safety boundaries. Use an observation quality scoring network based on MobileNetV3-Small to calculate the observation quality score of the observation dataset and perform quality gating. Step 3: Perform coarse detection on the observation dataset to obtain a set of candidate defect instances, generate candidate regions and type distributions for each candidate defect instance, and form a multi-view package; Step 4: Calculate the expected gain of the observation action primitive on the evidence item and generate the observation sequence under the constraints of time budget and retake number budget to determine the trigger state of the 3D evidence generation process; Step 5: In the triggered state, input the multi-view package into the improved DUSt3R model. The improved DUSt3R model includes a dot plot regression backbone module, a confidence robust alignment module, a geometric evidence verification module, and a scale measurement output module, and outputs a structured evidence package. Step Six: Combine the evidence type distribution with the evidence items in the structured evidence package to determine if the evidence threshold is met and output the detection results. If the threshold is not met, update the observation sequence based on the trigger signal. Step 7: Generate and save evidence data packages containing multi-view groups, structured evidence packages, and detection results, and update the selection strategy of observation action primitives based on the evidence data packages.

[0017] In this embodiment, step one specifically includes: Images or video frames of the same concrete structure area are collected and formed into an observation dataset. A uniform spatial coverage range is set for the same detection object. The acquisition time, shooting distance estimate and viewing angle information of each image or video frame are recorded. The images or video frames are grouped according to time order and viewing angle differences and written into the observation dataset. Establish a set of defect hypotheses, using the types of defects on the concrete structure surface as the hypothesis categories. Defect types include cracks, spalling, exposed reinforcement, honeycomb pitting, leakage and dampness marks. Assign a defect type identifier to each defect type and generate a corresponding defect hypothesis record. For each defect type, set corresponding evidence threshold parameters and set measurement requirements that match the defect type. The measurement requirements include the measurement requirements for the width and length of cracks, the measurement requirements for the area and depth of spalling, the measurement requirements for the exposed area of ​​exposed reinforcement, the measurement requirements for the area of ​​honeycomb pitting, and the measurement requirements for the area of ​​leakage and dampness marks. Evidence items are constructed and defined as structured evidence records established for the same defect hypothesis. An evidence item consists of texture evidence, geometric evidence, multi-view consistency evidence, and measurement credibility evidence. Texture evidence corresponds to image or video frames in the observation dataset, geometric evidence corresponds to the three-dimensional information output by the three-dimensional evidence generation process, multi-view consistency evidence corresponds to multi-view observation consistency information, and measurement credibility evidence corresponds to the execution credibility information of the measurement requirements.

[0018] In this embodiment, step two specifically includes: A primitive library of observation actions is constructed. The primitives of observation actions are defined as the acquisition action units that perform reshooting and reshooting on the same concrete structure area. The executable parameter range and safety boundary are set for each primitive of observation actions. The primitive library of observation actions includes close-range reshooting, changing the observation angle, adding scale anchoring and reshooting actions. An observation quality scoring network based on MobileNetV3-Small is constructed. MobileNetV3-Small is used as the feature extraction backbone network. At the output of the backbone network, a global average pooling layer, a first fully connected layer, and a second fully connected layer are connected in sequence to form a scoring head. The global average pooling layer aggregates the feature maps output by the backbone network into feature vectors. The first fully connected layer performs a linear transformation on the feature vectors and activates them with ReLU to obtain an intermediate representation. The second fully connected layer outputs a four-dimensional scoring vector and normalizes it with Sigmoid to obtain the blur score, overexposure score, reflection ratio score, and occlusion ratio score, with values ​​ranging from 0 to 1. Train the observation quality scoring network, establish a training sample set and label each training sample with blur label, overexposure label, reflection ratio label and occlusion ratio label, use the mean square error between the four-dimensional scoring vector and the corresponding label as the loss function, and use backpropagation to update the MobileNetV3-Small backbone network parameters and scoring head parameters. To calculate the observation quality score, each image or video frame in the observation dataset is subjected to input size alignment, pixel normalization, and channel normalization before being input into the observation quality score network. The output scores are then weighted and fused according to preset weights to obtain the observation quality score. Perform quality gating by comparing the observation quality score with a preset gating threshold parameter. If the observation quality score is lower than the gating threshold parameter, select a retake action from the observation action primitive library and re-acquire the corresponding image or video frame and write it into the observation dataset. If the observation quality score is not lower than the gating threshold parameter, retain the corresponding image or video frame.

[0019] In this embodiment, step three specifically includes: Perform size alignment, pixel normalization, and channel normalization on the images or video frames in the observation dataset; A coarse detection network is constructed, consisting of a backbone network, a neck network, and a detection head. The backbone network is composed of convolutional layers, batch normalization layers, and ReLU activation layers connected sequentially to form convolutional units. Multiple convolutional units are connected sequentially to form the backbone network. At the end of each stage, a residual connection layer is set to output the stage feature map, resulting in three-layer stage feature maps. The three-layer stage feature maps correspond to different downsampling ratios. The neck network uses a feature pyramid structure to perform top-down and bottom-up path fusion on the three-layer stage feature maps, aligning the channel dimensions during the fusion process to generate a fused feature map. The detection head includes a classification branch and a regression branch. The classification branch is composed of convolutional layers connected sequentially to a sigmoid output layer to output the confidence of candidate regions. The regression branch is composed of convolutional layers connected sequentially to a linear output layer to output the bounding box parameters of candidate regions. A coarse detection network is used to detect candidate regions in the processed image or video frame, and the candidate regions, candidate region confidence scores and candidate region bounding box parameters are obtained. The candidate region is instantiated, and its spatial location, scale information, confidence level, and bounding box parameters are written into the candidate defect instance set. A candidate defect instance identifier is assigned to each candidate defect instance. Input the candidate region corresponding to the candidate defect instance into the defect classification head, and output the category confidence vector of crack, peeling, exposed reinforcement, honeycomb surface, leakage and dampness marks. Write the category confidence vector as the type distribution into the candidate defect instance set. Based on the acquisition time and viewpoint information recorded in the observation dataset, candidate defect instances are correlated across frames. Candidate regions with overlapping spatial locations are grouped into the same package identifier to form a multi-view package, and the image or video frame index and viewpoint index contained in the package are recorded.

[0020] In this embodiment, step four specifically includes: Four-dimensional gain weights are set for each observation action primitive in the observation action primitive library. The four-dimensional gain weights correspond to the texture evidence, geometric evidence, multi-view consistency evidence and measurement credibility evidence in the evidence items, respectively. For each candidate defect instance in the candidate defect instance set, a coverage index and a difference index are calculated based on the multi-view package. A measurement complement index is obtained according to the measurement requirements. The coverage index is the ratio of the number of effective views in the multi-view package corresponding to the candidate defect instance to the number of preset target views. The difference index is the average value of the angle difference between the views in the multi-view package. The measurement complement index is the ratio of the number of measurement elements that have been satisfied in the measurement requirements to the total number of measurement elements included in the measurement requirements. The observation quality score is then read. For each observation action primitive, the total expected gain is calculated. The total expected gain is obtained by weighting the coverage index and the difference index according to the gain weight corresponding to geometric evidence, the measurement complementation index according to the gain weight corresponding to measurement credibility evidence, and the observation quality score according to the gain weight corresponding to texture evidence and multi-view consistency evidence. The observation action score is obtained by multiplying the total expected gain by the observation quality score. Under the constraints of time budget and reshoot time budget, the observation action primitive with the highest observation action score is selected to generate the observation sequence. The trigger state of the 3D evidence generation process is determined based on the observation sequence. The coverage index and difference index of the multi-view package after the observation sequence is executed are compared with the preset trigger threshold, and the measurement complementation index is compared with the preset measurement threshold. The trigger state is determined when the coverage index is not less than the preset trigger threshold and the difference index is not less than the preset trigger threshold. The trigger state is maintained when the measurement complementation index is not less than the preset measurement threshold. All other cases are determined as non-trigger state.

[0021] In this embodiment, the improved DUSt3R model includes a dot plot regression backbone module, a confidence robust alignment module, a geometric evidence verification module, and a scale measurement output module. The dot plot regression backbone module performs input size alignment and pixel normalization on each view image within the multi-view package. The aligned image is then input into the feature encoding network to obtain a feature map. The feature encoding network consists of convolutional layers, normalization layers, and activation functions connected sequentially. At the end, a dot plot regression head and a confidence regression head are connected. The dot plot regression head consists of convolutional layers and linear output layers connected sequentially, outputting three-dimensional coordinates pixel by pixel to form a dot plot. The confidence regression head consists of convolutional layers and sigmoid output layers connected sequentially, outputting confidence scores pixel by pixel to form a point-level confidence map. The dot plot is a set of three-dimensional point coordinates indexed by pixels, and the point-level confidence map is a set of confidence scores corresponding to each pixel of the dot plot. The confidence-robust alignment module receives point maps and point-level confidence maps from each viewpoint. It linearly scales the point-level confidence maps to obtain an alignment weight map and normalizes the alignment weight map. Based on the alignment weight map, it performs a pixel-by-pixel weighted summation on the point maps from each viewpoint to obtain an initial global point map. The confidence-robust alignment module performs robust suppression processing on the residuals between the initial global point map and the point maps from each viewpoint. The robust suppression processing includes truncating the residual magnitude and assigning attenuation weights to residuals that exceed the threshold. Then, it multiplies the attenuation weights with the alignment weight map to update the alignment weight map. The updated alignment weight map is used to repeatedly perform weighted summation on the point maps from each viewpoint to obtain the global point map, and the alignment quality index is output. The alignment quality index includes the statistical value of the attenuated residuals and the effective coverage rate. The effective coverage rate is the percentage of pixels in the global point map whose confidence weights meet a preset threshold. The geometric evidence verification module receives the global point map and generates geometric evidence, including geometric discontinuity intensity, normal abrupt change, and curvature anomaly. Geometric discontinuity intensity is obtained by performing neighborhood difference on the depth component of the global point map, taking the absolute value, and then thresholding. Normal abrupt change is obtained by calculating the local plane normal on the global point map and taking the absolute value and thresholding the angle between adjacent normals. Curvature anomaly is obtained by calculating the local second-order difference on the global point map and thresholding the difference magnitude. Multi-view consistency verification is performed by back-projecting the global point map to each viewpoint and performing pixel-by-pixel matching with the corresponding viewpoint point map. Based on the matching residual, a consistency weight is obtained by exponential decay, and the consistency weight is averaged to obtain the multi-view consistency score. The drift amount is obtained by averaging the pixel displacement based on the matching residual. When the multi-view consistency score is lower than the preset consistency threshold or the drift amount is higher than the preset drift threshold, a trigger signal is output. The scale measurement output module receives the global point map, geometric evidence, and measurement requirements. It performs scale anchoring on the global point map and outputs millimeter-level measurement results and measurement credibility evidence. Scale anchoring maps the three-dimensional coordinate scale of the global point map to the millimeter scale by reading scale reference information. The millimeter-level measurement results include crack width, crack length, spalling area, spalling depth, exposed reinforcement area, honeycomb surface area, and leakage and moisture trace area. The measurement credibility evidence is obtained by weighted multiplication of the alignment quality index and the multi-view consistency score and then normalization. Write the point maps of each viewpoint, the confidence maps of each viewpoint, the global point map, the alignment quality index, the geometric evidence, the multi-view consistency score, the drift amount, the trigger signal, the millimeter-level measurement results, and the measurement confidence evidence into a structured evidence package and output it.

[0022] The improved DUSt3R model maintains the same overall framework as the original DUSt3R model, both using multi-view images as input, extracting features through a backbone network and regressing point map representations, then aligning the multi-view results to the same reference frame. This implementation also retains the core idea of ​​point map regression and the multi-view geometric fusion path, enabling the model to form a stable three-dimensional representation of the same concrete structure region in complex scenes. Compared to the original DUSt3R model, this invention improves the subsequent processing links for concrete defects without changing the core of the point map regression: In the confidence-robust alignment module, point-level confidence weighting and robust residual suppression are introduced to suppress the impact of anomalous matching caused by reflections, water stains, and repetitive textures on the global point map; in the geometric evidence verification module, geometric evidence extraction such as geometric discontinuity strength, normal abrupt changes, and curvature anomalies are introduced and combined with multi-view consistency verification output trigger signals; in the scale measurement output module, scale-anchored millimeter-level measurements and measurement credibility evidence generation are introduced, so that the structured evidence package simultaneously includes alignment quality indicators, multi-view consistency scores, and measurement results. Through improvements, the model output has been upgraded from "general 3D reconstruction results" to structured evidence that is "determinable, measurable, and loop-closed": confidence weighting and robust suppression reduce alignment drift caused by false matching and improve the stability of the global point map; geometric evidence verification and multi-view consistency constraints enhance the ability to eliminate false defects such as template textures, shadows, and water stains; scale anchoring and measurement credibility evidence improve the reliability of millimeter-level measurements such as crack width and peeling depth, and drive reshooting and replanning through trigger signals to make the evidence threshold easier to meet, thereby improving the robustness, repeatability, and delivery credibility of engineering inspection as a whole.

[0023] In this embodiment, step six specifically includes: Read the type distribution of candidate defect instances, and determine the corresponding evidence threshold parameters and measurement requirements according to the defect type with the highest category confidence in the type distribution; Read the evidence items and trigger signals in the structured evidence package, and convert the evidence items into texture evidence scores, geometric evidence scores, multi-view consistency evidence scores, and measurement confidence evidence scores, respectively. The texture evidence score is generated by statistically analyzing the candidate region confidence and texture features within the candidate region, the geometric evidence score is generated by statistically analyzing the amplitude of geometric evidence, the multi-view consistency evidence score is generated by multi-view consistency scoring, and the measurement confidence evidence score is generated by measurement confidence evidence. The evidence threshold is determined by weighted fusion of texture evidence score, geometric evidence score, multi-view consistency evidence score and measurement confidence evidence score to obtain evidence compliance score. The evidence compliance score is then compared with the evidence threshold parameter. If the evidence compliance score is not less than the evidence threshold parameter, the detection result is generated. The detection result includes defect type, candidate region location and measurement result corresponding to measurement requirements. When the evidence pass score is less than the evidence threshold parameter, the observation sequence is updated according to the trigger signal. Updating the observation sequence includes selecting the observation action primitive corresponding to the trigger signal from the observation action primitive library and appending it to the observation sequence. The appended observation sequence is executed to collect new image or video frames and write them into the observation dataset. The multi-view group package is updated according to step three, the structured evidence package is updated according to step five, and the evidence threshold pass judgment is executed again.

[0024] In this embodiment, step seven specifically includes: The multi-view package, structured evidence package, and detection results are encapsulated to generate an evidence data package, and an evidence data package identifier is assigned to the evidence data package; Write the package identifier, the image or video frame index and view information contained in the package into the evidence data package, write the alignment quality index, multi-view consistency score, trigger signal and millimeter-level measurement results into the structured evidence package, and write the defect type and candidate region location into the detection results. The correspondence between the trigger signals in the statistical evidence data package and the actual observed action primitives is calculated. The number of times the standard is passed after each observed action primitive is triggered is counted, and the four-dimensional gain weight is updated based on the correspondence and the number of passes.

[0025] Example 1: To verify the feasibility of this invention in practice, it was applied to the annual inspection scenario of the piers and box girder bottom plates of an expressway viaduct in a certain city. This scenario commonly involves strong backlighting, slight vibrations caused by traffic flow, and the coexistence of textured concrete surfaces and water stains reflecting off the surface. Traditional single-view inspection easily misjudges textures as cracks or misses detections when the image is blurry or overexposed. On-site, a handheld gimbal camera was used to capture RGB video at a resolution of 3840×2160 and a frame rate of 30fps. The estimated acquisition time for the same suspected area of ​​the same component was 45 seconds, with a budget of no more than 3 retakes. The number of target viewpoints was set to 6, with an average angle difference of no less than 15°. 30mm ruler stickers were used for dimensional anchoring, required to be placed near the same plane as the suspected measurement area. Evidence threshold parameters were preset as follows: cracks 0.70, spalling 0.75, and leakage / dampness marks 0.65. Measurement requirements were preset as follows: crack output width and length, spalling output area and depth, exposed rebar output area, honeycomb surface output area, and leakage and dampness mark output area. The observation quality scoring network uses the MobileNetV3-Small backbone and outputs quality scores. The gating threshold is set to 0.60; if the score is below the threshold, a retake is triggered. The observation action primitive library is configured according to close-range retake (distance 2.0m → 1.2m), changing the observation angle (adding 20° oblique view to the left and right), adding scale anchoring (adding or shifting the scale), and retake (stabilizing the gimbal and resetting the exposure). The coarse detection network outputs candidate regions and type distributions. Multi-view packages are formed based on the association rule of "candidate region center point projection overlap rate ≥ 0.5 and time interval ≤ 8s". When the coverage reaches 6 / 6, the difference meets the target, and the measurement complement is 1.0, the 3D evidence generation process is triggered. The package is sent to the improved DUSt3R to output a structured evidence package, and the detection result is output according to the evidence threshold. If the threshold is not met, the observation sequence is updated and supplemented by the trigger signal.

[0026] Two common methods were selected for comparison: Comparison A involved single-view target detection and segmentation, with width / area estimated by pixel conversion, relying solely on a single-frame confidence threshold of 0.50 for judgment; Comparison B involved simple multi-view fusion after single-view detection (without point-level confidence robust alignment and geometric verification), with measurements using only a small amount of viewpoint triangulation, lacking quality gating and re-enhancing loop. The method of this invention, along with the two comparison methods, was used to detect and measure 120 component units (40 piers, 40 spans of box girder bottom plates, and 40 connection areas) within the same inspection window. Manual verification was used as the reference true value. The results showed 312 cracks, 74 spalling points, 58 leaks / dampness marks, 41 honeycomb pitting areas, and 19 exposed rebar areas. The comprehensive indicators are summarized in Table 1. Table 1. Comparison of Comprehensive Indicators of the Three Methods

[0027] As shown in Table 1, the crack F1 of Comparison A is 0.79, mainly due to the failure to detect fine cracks under vibration and overexposure. Comparison B improved the crack F1 to 0.85 through simple multi-view fusion, but the false defect false alarm rate was still 8.6%, indicating that "multi-view fusion" alone is insufficient to stably distinguish between template texture, water stain reflection, and real defects. This invention reduces the false defect false alarm rate to 4.1%, directly improving precision and driving an overall increase in F1. This improvement is closely related to the "observation quality scoring + gating" to eliminate low-quality frames at the source and the "improved DUSt3R geometric evidence verification" to eliminate false textures at the three-dimensional level. In terms of measurement, the crack width MAE of this invention decreased from 0.18mm to 0.07mm, and the P90 error decreased from 0.32mm to 0.12mm, indicating that the error tail was significantly compressed. For the peeling depth, the MAE decreased from 2.6mm to 0.9mm, meaning that the most controversial issue in maintenance grading, "depth estimation being too large / too small," was significantly suppressed. The reason is that scale anchoring unifies the scale of the dot map to the millimeter dimension, while confidence-robust alignment reduces alignment drift, and geometric evidence verification makes it difficult for areas with "texture-like peeling but geometrically invalid" to pass the standard judgment.

[0028] It is important to emphasize that the average time per component of this invention is 36.8 seconds, which is higher than the 28.4 seconds of Comparison A, but still lower than the budget of 45 seconds, resulting in a stable delivery with a final compliance rate of 99.2%. Comparisons A and B, under complex lighting and reflective conditions, still had a certain percentage of components that failed to meet the standards or had unstable measurements, with final compliance rates of 92.5% and 95.0%, respectively. This invention, through "expected gain," directs the re-shooting action to the area with the most lacking evidence, and uses a trigger signal to drive the update of the observation sequence, transforming the time consumed by re-shooting into a certainty of evidence compliance. In other words, this invention does not pursue "the fastest results," but rather pursues "results that are verifiable, measurements that are usable, and false alarms that are controllable" within a controllable budget.

[0029] In summary, this invention, in a real bridge inspection environment, achieves input quality gating through the MobileNetV3-Small observation quality scoring network, generates observation sequences through multi-view grouping of candidate defect instances and expected gain strategies, and completes three-dimensional geometric verification and scale-anchored millimeter measurement by outputting structured evidence packages with an improved DUSt3R. Ultimately, under the evidence threshold compliance judgment and trigger re-shooting closed loop, it significantly improves defect detection performance, reduces false defect alarms, and enhances measurement accuracy and compliance rate, thus meeting the key requirements of engineering inspection for reliable judgment, traceable evidence, and standardized delivery.

[0030] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for detecting concrete structures, characterized in that, Includes the following steps: Step 1: Collect images or video frames of the same concrete structure area and form an observation dataset, establish a set of defect hypotheses and set evidence threshold parameters and measurement requirements for each defect type, and construct evidence items; Step 2: Construct a primitive library of observation actions and set parameter ranges and safety boundaries. Use an observation quality scoring network based on MobileNetV3-Small to calculate the observation quality score of the observation dataset and perform quality gating. Step 3: Perform coarse detection on the observation dataset to obtain a set of candidate defect instances, generate candidate regions and type distributions for each candidate defect instance, and form a multi-view package; Step 4: Calculate the expected gain of the observation action primitive on the evidence item and generate the observation sequence under the constraints of time budget and retake number budget to determine the trigger state of the 3D evidence generation process; Step 5: In the triggered state, input the multi-view package into the improved DUSt3R model. The improved DUSt3R model includes a dot plot regression backbone module, a confidence robust alignment module, a geometric evidence verification module, and a scale measurement output module, and outputs a structured evidence package. Step Six: Combine the evidence type distribution with the evidence items in the structured evidence package to determine if the evidence threshold is met and output the detection results. If the threshold is not met, update the observation sequence based on the trigger signal. Step 7: Generate and save evidence data packages containing multi-view groups, structured evidence packages, and detection results, and update the selection strategy of observation action primitives based on the evidence data packages.

2. The method for detecting concrete structures based on deep learning according to claim 1, characterized in that, Step one specifically includes: Images or video frames of the same concrete structure area are collected and formed into an observation dataset. A uniform spatial coverage range is set for the same detection object. The acquisition time, shooting distance estimate and viewing angle information of each image or video frame are recorded. The images or video frames are grouped according to time order and viewing angle differences and written into the observation dataset. Establish a set of defect hypotheses, using the defect types on the concrete structure surface as the hypothesis categories, assign a defect type identifier to each defect type and generate a corresponding defect hypothesis record; Set corresponding evidence threshold parameters for each defect type, and set measurement requirements that match the defect type; An evidence item is constructed, which is defined as a structured evidence record established for the same defect hypothesis. The evidence item consists of texture evidence, geometric evidence, multi-view consistency evidence, and measurement credibility evidence.

3. The method for detecting concrete structures based on deep learning according to claim 1, characterized in that, Step two specifically includes: A library of observation action primitives is constructed, which are defined as acquisition action units that perform re-shooting and re-shooting on the same concrete structure area. The executable parameter range and safety boundary are set for each observation action primitive. An observation quality scoring network based on MobileNetV3-Small is constructed. MobileNetV3-Small is used as the feature extraction backbone network. A global average pooling layer, a first fully connected layer and a second fully connected layer are sequentially connected at the output of the backbone network to form a scoring head. Train the observation quality scoring network, establish a training sample set and label each training sample with blur label, overexposure label, reflection ratio label and occlusion ratio label, use the mean square error between the four-dimensional scoring vector and the corresponding label as the loss function, and use backpropagation to update the MobileNetV3-Small backbone network parameters and scoring head parameters. To calculate the observation quality score, each image or video frame in the observation dataset is subjected to input size alignment, pixel normalization, and channel normalization before being input into the observation quality score network. The output scores are then weighted and fused according to preset weights to obtain the observation quality score. Perform quality gating by comparing the observation quality score with a preset gating threshold parameter. If the observation quality score is lower than the gating threshold parameter, select a retake action from the observation action primitive library and re-acquire the corresponding image or video frame and write it into the observation dataset. If the observation quality score is not lower than the gating threshold parameter, retain the corresponding image or video frame.

4. The method for detecting concrete structures based on deep learning according to claim 1, characterized in that, Step three specifically includes: Perform size alignment, pixel normalization, and channel normalization on the images or video frames in the observation dataset; A coarse detection network is constructed, consisting of a backbone network, a neck network, and a detection head. The backbone network is composed of convolutional units formed by sequentially connecting convolutional layers, batch normalization layers, and ReLU activation layers. Multiple convolutional units are sequentially connected to form the backbone network. At the end of each stage, a residual connection layer is set to output a stage feature map, resulting in a three-layer stage feature map. The neck network uses a feature pyramid structure to perform top-down and bottom-up path fusion on the three-layer stage feature maps, aligning the channel dimensions during the fusion process to generate a fused feature map. The detection head includes a classification branch and a regression branch, outputting candidate region confidence and candidate region bounding box parameters. A coarse detection network is used to detect candidate regions in the processed image or video frame, and the candidate regions, candidate region confidence scores and candidate region bounding box parameters are obtained. The candidate region is instantiated, and its spatial location, scale information, confidence level, and bounding box parameters are written into the candidate defect instance set. A candidate defect instance identifier is assigned to each candidate defect instance. Input the candidate region corresponding to the candidate defect instance into the defect classification head, and output the category confidence vector of crack, peeling, exposed reinforcement, honeycomb surface, leakage and dampness marks. Write the category confidence vector as the type distribution into the candidate defect instance set. Based on the acquisition time and viewpoint information recorded in the observation dataset, candidate defect instances are correlated across frames, and candidate regions with overlapping spatial locations are grouped into the same package identifier to form a multi-view package.

5. The method for detecting concrete structures based on deep learning according to claim 1, characterized in that, Step four specifically includes: Set four-dimensional gain weights for each observation action primitive in the observation action primitive library; For each candidate defect instance in the candidate defect instance set, the coverage index and the difference index are calculated based on multi-view package, the measurement complementation index is obtained according to the measurement requirements, and the observation quality score is read. For each observation action primitive, the total expected gain is calculated. The total expected gain is obtained by weighting the coverage index and the difference index according to the gain weight corresponding to geometric evidence, the measurement complementation index according to the gain weight corresponding to measurement credibility evidence, and the observation quality score according to the gain weight corresponding to texture evidence and multi-view consistency evidence. The observation action score is obtained by multiplying the total expected gain by the observation quality score. Under the constraints of time budget and reshoot time budget, the observation action primitive with the highest observation action score is selected to generate the observation sequence, and the trigger state of the three-dimensional evidence generation process is determined based on the observation sequence.

6. The method for detecting concrete structures based on deep learning according to claim 1, characterized in that, The improved DUSt3R model includes a dot plot regression backbone module, a confidence robust alignment module, a geometric evidence verification module, and a scale measurement output module. The dot plot regression backbone module performs input size alignment and pixel normalization on each view image within the multi-view group, and inputs the aligned image into the feature encoding network to obtain a feature map. The feature encoding network connects the dot plot regression head and the confidence regression head at the end. The dot plot regression head outputs three-dimensional coordinates pixel by pixel to form a dot plot, and the confidence regression head outputs confidence values ​​pixel by pixel to form a point-level confidence map. The confidence-robust alignment module receives point maps and point-level confidence maps from each viewpoint, linearly scales the point-level confidence maps to obtain an alignment weight map, normalizes the alignment weight map, and performs pixel-by-pixel weighted summation on the point maps based on the alignment weight map to obtain an initial global point map. The confidence-robust alignment module performs robust suppression processing on the residuals between the initial global point map and the point maps from each viewpoint to obtain the global point map and outputs the alignment quality index. The geometric evidence verification module receives the global point map and generates geometric evidence, which includes geometric discontinuity intensity, normal abrupt change and curvature anomaly. It performs multi-view consistency verification by back-projecting the global point map to each viewpoint and matching it pixel by pixel with the corresponding viewpoint point map. Based on the matching residual, the consistency weight is obtained by exponential decay and the consistency weight is averaged to obtain the multi-view consistency score. The drift amount is obtained by averaging the pixel displacement of the matching residual. When the multi-view consistency score is lower than the preset consistency threshold or the drift amount is higher than the preset drift threshold, a trigger signal is output. The scale measurement output module receives the global point map, geometric evidence, and measurement requirements, performs scale anchoring on the global point map, and outputs millimeter-level measurement results and measurement credibility evidence.

7. The method for detecting concrete structures based on deep learning according to claim 1, characterized in that, Step six specifically includes: Read the type distribution of candidate defect instances, and determine the corresponding evidence threshold parameters and measurement requirements according to the defect type with the highest category confidence in the type distribution; Read the evidence items and trigger signals in the structured evidence package, and convert the evidence items into texture evidence scores, geometric evidence scores, multi-view consistency evidence scores and measurement credibility evidence scores, respectively. The evidence threshold is determined by weighted fusion of texture evidence score, geometric evidence score, multi-view consistency evidence score and measurement confidence evidence score to obtain evidence compliance score. The evidence compliance score is then compared with the evidence threshold parameter. A detection result is generated when the evidence compliance score is not less than the evidence threshold parameter. When the evidence compliance score is less than the evidence threshold parameter, the observation sequence is updated according to the trigger signal, the multi-view group is updated according to step three, the structured evidence group is updated according to step five, and the evidence threshold compliance judgment is performed again.

8. The method for detecting concrete structures based on deep learning according to claim 1, characterized in that, Step seven specifically includes: The multi-view package, structured evidence package, and detection results are encapsulated to generate an evidence data package, and an evidence data package identifier is assigned to the evidence data package; Write the package identifier, the image or video frame index and view information contained in the package into the evidence data package, write the alignment quality index, multi-view consistency score, trigger signal and millimeter-level measurement results into the structured evidence package, and write the defect type and candidate region location into the detection results. The correspondence between the trigger signals in the statistical evidence data package and the actual observed action primitives is calculated. The number of times the standard is passed after each observed action primitive is triggered is counted, and the four-dimensional gain weight is updated based on the correspondence and the number of passes.

Images