AI-based rapid, accurate and efficient reinforcing steel bar acceptance method and system

Abstract

The invention relates to the field of image measurement, and discloses a rapid, accurate and efficient reinforcing steel bar acceptance method and system based on AI, and the method comprises the following steps: obtaining a reinforcing steel bar construction region image through a mobile terminal camera; sequentially carrying out geometric correction and image enhancement processing on the collected original image; inputting the optimized image into an improved deep learning model; based on pixel-level data output by AI identification, physical parameters of spacing, diameter and length of reinforcing steel bars are calculated; calling an acceptance rule base, comparing the calculation parameters with a preset standard, and judging single-item and overall eligibility; and generating a visual result graph containing annotation information and a structured acceptance report, and storing key acceptance data. Through a deep learning reinforcement recognition model based on multi-scale feature fusion and a self-adaptive image enhancement and perspective correction algorithm for a construction site scene, the effect of stably and accurately recognizing the number, the position and the specification of the reinforcement under the conditions of illumination variation, background disorder, reinforcement shielding and non-vertical shooting is achieved.

Description

Technical Field

[0001] This invention relates to the field of image measurement technology, specifically to an AI-based method and system for fast, accurate, and efficient steel reinforcement acceptance. Background Technology

[0002] In the quality control system of civil engineering construction, steel bars, as core load-bearing components, directly bear tensile and shear stresses. Their quantity, specifications, spacing, length, and connection methods directly determine the structure's load-bearing capacity, seismic performance, and durability. Insufficient steel bars, undersized diameters, or excessive spacing can lead to insufficient structural strength, easily causing cracking, deformation, or even collapse. If the protective layer thickness does not meet requirements, it will accelerate steel bar corrosion and shorten the project's service life. Therefore, acceptance testing is a crucial line of defense against potential structural safety hazards.

[0003] In traditional civil engineering practice, the acceptance, counting, and spacing inspection of reinforcing bars mainly rely on manual methods. Supervisors use tools such as tape measures to conduct on-site measurements and count the bars visually. The related technologies mostly rely on single-scale features or manual design features, which are not adaptable to complex site scenarios. Under conditions such as large changes in lighting, cluttered backgrounds, mutual obstruction of reinforcing bars, and non-perpendicular shooting, problems such as missed inspections, mis-inspections, or inaccurate identification of reinforcing bars specifications are prone to occur. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a fast, accurate, and efficient method and system for inspecting reinforcing bars based on AI. This system solves the problems of missed inspections, false inspections, or inaccurate specification identification of reinforcing bars that easily occur under conditions such as large changes in lighting, cluttered backgrounds, mutual obstruction of reinforcing bars, and non-perpendicular shooting.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A fast, accurate, and efficient method for inspecting reinforcing bars based on AI includes the following steps:

[0007] Images of the steel reinforcement construction area are acquired through the mobile terminal's camera, and data from the mobile terminal's inertial measurement unit, shooting distance, and ambient light intensity are recorded simultaneously.

[0008] The acquired raw images are sequentially subjected to geometric correction and image enhancement processing to obtain optimized images;

[0009] The optimized image is input into the improved deep learning model, which simultaneously outputs the bounding box coordinates, category confidence, diameter classification results, and center point coordinates of the reinforcing bars.

[0010] Based on the pixel-level data output by AI recognition, combined with the calibrated pixel precision, the physical parameters of rebar spacing, diameter and length are calculated;

[0011] The acceptance rule library is called up to compare the calculated parameters with the preset standards to determine the passability of individual items and the whole;

[0012] Generates visualized results charts with annotation information and structured acceptance reports, and stores key acceptance data.

[0013] The geometric correction employs a perspective correction method based on homography transformation, and the process includes the following steps:

[0014] Edge detection operators are used to extract the edges of the reinforcing mesh, and line detection algorithms are used to detect the intersections of lines. The four corner points of the reinforcing mesh are located as the corner point set of the source image.

[0015] Define the target corner point set under orthogonal viewpoints, and calculate the homography matrix based on the source image corner point set and the target corner point set;

[0016] A perspective transformation is performed on the original image based on the homography matrix to obtain a corrected image under orthogonal viewpoints.

[0017] By adopting the above technical solution, automatic perspective correction of images of the steel reinforcement construction area can be achieved without the need for additional physical calibration devices, reducing the impact of on-site shooting conditions on measurement results and solving the problem of unstable calculation accuracy of steel reinforcement parameters due to inconsistent shooting angles in the existing technology.

[0018] As a further description of the above technical solution: the image enhancement includes contrast-limited adaptive histogram equalization, Gaussian filtering, and gamma correction, and the processing includes the following steps:

[0019] Image processing employs contrast-limited adaptive histogram equalization for adjusting local shadows and lighting;

[0020] Noise suppression is achieved by using a Gaussian filter with a preset kernel size and standard deviation.

[0021] Adjust the gamma coefficient to correct the brightness of the image.

[0022] By adopting the above technical solution, the brightness and contrast of the image of the steel reinforcement construction area can be adaptively adjusted under different lighting conditions, reducing the impact of shadows, overexposure and uneven lighting on image quality. Through noise suppression and brightness correction, the edge features of the steel reinforcement are made clearer and more continuous, improving the recognizability of the steel reinforcement outline and texture information. This provides a stable input image for subsequent deep learning models for target detection, bounding box localization and key point extraction, thereby improving the accuracy and robustness of steel reinforcement recognition and parameter calculation.

[0023] As a further description of the above technical solution: the improved deep learning model is based on the YOLOv8 architecture and includes a backbone network, a neck network, and a multi-task detection head. The specific structure is as follows:

[0024] The backbone network replaces standard convolution with depthwise separable convolution and embeds a convolutional block attention mechanism, including a cross-stage local network and a spatial pyramid pooling fusion module.

[0025] The neck network adopts a fusion structure of feature pyramid and path aggregation network for multi-scale feature fusion;

[0026] The multi-task detection head includes a classification branch, a regression branch, a diameter classification branch, and a key point branch. The diameter classification branch is used for the classification and identification of rebar specifications.

[0027] By adopting the above technical solution, firstly, by introducing depthwise separable convolution and attention mechanisms into the YOLOv8 backbone network, the computational load of the model is reduced while ensuring feature extraction capabilities. Combined with a multi-scale feature fusion structure, the model can more stably identify steel bar targets of different sizes and dense distributions. At the same time, by setting up a multi-task detection head that includes localization, classification, diameter classification, and key point extraction, steel bar detection and specification recognition are integrated into the same model, thereby reducing errors caused by multi-model processing. Compared with traditional detection methods, this improves the accuracy and consistency of steel bar recognition and parameter calculation results.

[0028] As a further description of the above technical solution: the pixel precision is calibrated by one of the following two methods:

[0029] Pixel accuracy is calculated using the sensor width, shooting distance, image width, and focal length of the mobile terminal camera. The pixel accuracy is the ratio of the product of the sensor width and shooting distance to the product of the image width and focal length.

[0030] Pixel precision is calculated by using a reference object with a known actual length in the image and its corresponding image pixel length. Pixel precision is the ratio of the actual length of the reference object to the pixel length of the reference object's image.

[0031] By adopting the above technical solution, the pixel accuracy is first calibrated by using camera imaging parameters or reference objects of known size in the image, so that the image pixel size can establish a correspondence with the actual physical size. This allows for accurate conversion from pixel to actual length without the need for additional calibration equipment. Compared with methods that rely solely on experience or fixed proportions, this improves the accuracy and applicability of the calculation results for geometric parameters such as rebar length, spacing, and diameter.

[0032] As a further description of the above technical solution: the calculation process of the rebar spacing includes the following steps:

[0033] The coordinates of the center points of the reinforcing bars are sorted according to a preset direction to obtain the sorted set of center point coordinates;

[0034] Calculate the pixel distance between two adjacent points in the sorted center point coordinate set;

[0035] The product of pixel distance and pixel precision is the actual spacing of the reinforcing bars.

[0036] By adopting the above technical solution, the center points of the detected steel bars are first sorted according to a preset direction to ensure accurate correspondence between adjacent steel bars. Then, based on the pixel distance between adjacent center points and combined with the calibrated pixel precision, the distance in the image space is converted into the actual physical spacing, thereby realizing automatic and continuous calculation of steel bar spacing. Compared with manual measurement or experience estimation, this improves the accuracy and stability of steel bar spacing calculation results.

[0037] As a further description of the above technical solution: the calculation process of the steel bar diameter includes the following steps:

[0038] Obtain the width of the rebar boundary box output by AI intelligent analysis;

[0039] Calculate the angle between the steel bar axis and the horizontal direction of the image;

[0040] The product of the bounding box width, pixel precision, and the cosine of the included angle is the actual diameter of the reinforcing bar.

[0041] By adopting the above technical solution, the width of the bounding box of the reinforcing bar is directly obtained using AI intelligent analysis results. Combined with the angle between the reinforcing bar axis and the horizontal direction of the image, the width of the bounding box is corrected in direction. Then, the pixel size is converted into the actual physical size through the calibrated pixel precision. Thus, even when the reinforcing bar is tilted or the shooting angle changes, the actual diameter of the reinforcing bar can still be accurately calculated. Compared with the method of estimating based solely on pixel width, this improves the accuracy and reliability of the reinforcing bar diameter calculation results.

[0042] As a further description of the above technical solution: the calculation process of the steel bar length includes the following steps:

[0043] First, extract the pixel coordinates of the two ends of the rebar;

[0044] The endpoint pixel coordinates are converted to world coordinates using the inverse of the homography matrix;

[0045] Calculate the Euclidean distance between the two endpoints in world coordinates to obtain the actual length of the reinforcing bar.

[0046] By adopting the above technical solution, the pixel coordinates of the two ends of the steel bar are extracted and mapped to the real world coordinate system using the inverse transformation of the homography matrix. This establishes a correspondence between the image coordinates and the real spatial position. Then, the actual length of the steel bar is directly calculated based on the Euclidean distance between the two ends in the world coordinate system. Thus, even if there is a deviation in the shooting angle or perspective distortion in the image, a relatively accurate steel bar length calculation result can still be obtained, improving the accuracy and reliability of steel bar length measurement.

[0047] As a further description of the above technical solution: the acceptance rule base contains the construction quality acceptance specification for concrete structure engineering;

[0048] During the acceptance and judgment process, the acceptance tolerance threshold is dynamically adjusted according to the steel bar obstruction rate and binding status. When the confidence level of the test result is lower than the preset threshold, the test result is marked as requiring manual review.

[0049] During the acceptance and judgment process, quantitative indicators for the steel bar obstruction rate and binding status are set, and the acceptance tolerance threshold is dynamically adjusted according to the quantitative indicators at a preset ratio.

[0050] By adopting the above technical solution, the acceptance rules for concrete structure construction quality are built into the acceptance rule library, and custom configuration is supported according to actual project needs. This makes the acceptance judgment have a standardized basis and flexible adaptability. During the acceptance process, the acceptance tolerance threshold is dynamically adjusted according to the actual situation such as the degree of obstruction of the reinforcing bars and the binding status. When the confidence level of the test results is low, it is automatically marked as requiring manual review, thereby avoiding misjudgment caused by complex on-site conditions. Compared with the acceptance method with fixed thresholds, the rationality, reliability and practicality of the reinforcing bar acceptance results are improved.

[0051] A fast, accurate, and efficient rebar inspection system based on AI, comprising a hardware layer, a support layer, a core service layer, and an application layer, wherein:

[0052] The hardware layer also includes a camera module, a computing unit, a storage unit, and an inertial measurement unit. The camera module is used to acquire image data of the reinforcing bars, the computing unit is used to provide local computing capabilities, the storage unit is used to cache data and model files, and the inertial measurement unit is used to acquire motion state data.

[0053] The support layer also includes a mobile inference framework, a model management module, and a device adaptation module. The mobile inference framework is used to provide a model running environment, the model management module is used for model loading, version control, and dynamic updates, and the device adaptation module is used to detect hardware performance and allocate computing resources.

[0054] The core service layer also includes an image preprocessing module, an AI analysis engine, a geometric calculation module, and an acceptance rule engine. The image preprocessing module is used to perform image correction and enhancement operations, the AI ​​analysis engine is used to deploy improved deep learning models and rebar recognition, the geometric calculation module is used for parameter measurement, and the acceptance rule engine is used to perform acceptance judgment.

[0055] The application layer also includes a user interface module, a data management module, and a report generation module. The user interface module provides image acquisition and interaction functions, the data management module manages project data, and the report generation module outputs visualization results and acceptance reports.

[0056] By adopting the above technical solution, the mini-program is divided into a hardware layer, a support layer, a core service layer, and an application layer, achieving layered decoupling and collaborative work from data acquisition and model inference to result output. The hardware layer provides basic support for rebar image acquisition and local calculation. The support layer ensures the stable operation of the system on different terminals through model management and device adaptation. The core service layer completes key functions such as rebar recognition, parameter calculation, and acceptance judgment. The application layer realizes human-computer interaction and result display, thus forming a complete and clear business processing flow. Compared with the single-module implementation method, this improves the overall stability, scalability, and reliability of the system in actual engineering applications.

[0057] This invention provides a fast, accurate, and efficient method and system for steel reinforcement acceptance based on AI. It has the following beneficial effects:

[0058] 1. In this invention, a deep learning rebar recognition model based on multi-scale feature fusion and an adaptive image enhancement and perspective correction algorithm for construction site scenarios are used to achieve stable and accurate recognition of the quantity, location and specifications of rebars even under conditions of changing lighting, cluttered background, rebar occlusion and non-perpendicular shooting.

[0059] 2. In this invention, by combining monocular vision geometric modeling, camera imaging parameter derivation, and prior knowledge base of steel bar specifications, a high-precision measurement of steel bar spacing, diameter, and length is achieved without the need for physical rulers or dedicated measuring equipment.

[0060] 3. In this invention, through lightweight neural network structure design, model pruning and quantization optimization, and adaptive inference strategy, the near real-time output of steel reinforcement acceptance results is achieved in ordinary smartphones and mini-program environments.

[0061] 4. In this invention, by combining the steel bar identification and measurement results with the built-in acceptance specification rule engine, structured report generation and blockchain evidence storage mechanism, the results of the steel bar acceptance process are visualized, the judgment is automated and the data is tamper-proof and traceable. Attached Figure Description

[0062] Figure 1 This is a flowchart of the method of the present invention;

[0063] Figure 2 This is a system module architecture diagram of the present invention. Detailed Implementation

[0064] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] To better understand the above technical solutions, the technical solutions of the present invention will be clearly and completely described below in conjunction with embodiments.

[0066] Example:

[0067] Reference Figure 1 In an embodiment of the present invention, the present invention provides a fast, accurate and efficient method for steel reinforcement acceptance based on AI, comprising the following steps:

[0068] Images of the steel reinforcement construction area are acquired through the mobile terminal's camera, and data from the mobile terminal's inertial measurement unit, shooting distance, and ambient light intensity are recorded simultaneously.

[0069] The acquired raw images are sequentially subjected to geometric correction and image enhancement processing to obtain optimized images;

[0070] The optimized image is input into the improved deep learning model, which simultaneously outputs the bounding box coordinates, category confidence, diameter classification results, and center point coordinates of the reinforcing bars.

[0071] Based on the pixel-level data output by AI recognition, combined with the calibrated pixel precision, the physical parameters of rebar spacing, diameter and length are calculated;

[0072] The acceptance rule library is called up to compare the calculated parameters with the preset standards to determine the passability of individual items and the whole;

[0073] Generates visualized results charts with annotation information and structured acceptance reports, and stores key acceptance data.

[0074] Geometric correction employs a perspective correction method based on homography transformation, and the process includes the following steps:

[0075] Edge detection operators are used to extract the edges of the reinforcing mesh, and line detection algorithms are used to detect the intersections of lines. The four corner points of the reinforcing mesh are located as the corner point set of the source image.

[0076] Define the target corner point set under orthogonal viewpoints, and calculate the homography matrix based on the source image corner point set and the target corner point set;

[0077] A perspective transformation is performed on the original image based on the homography matrix to obtain a corrected image under orthogonal viewpoints.

[0078] Specifically, when performing geometric correction on the steel reinforcement image, the original image containing the steel reinforcement mesh structure is first used as input. The original image is then subjected to grayscale conversion and noise reduction to reduce the impact of lighting changes and background interference on the edge extraction results. Then, based on this, a preset edge detection operator is used to extract edge features from the image to obtain the edge contour information of the steel reinforcement mesh in the image.

[0079] After obtaining the edge contour information, a straight line detection algorithm is executed to extract multiple straight lines representing the direction of the rebar mesh. The lines are then classified and filtered by direction, and combinations of mutually perpendicular straight lines located in the image edge region are selected. By calculating the coordinates of the intersection points between the straight lines, the positions of the four corner points of the rebar mesh in the original image are determined, and the corner point coordinates are used to construct the source image corner point set. Under the orthogonal viewpoint, a rectangular target corner point set corresponding to the actual geometric shape of the rebar mesh is predefined, where each target corner point is located in the same plane coordinate system and adjacent sides are mutually perpendicular. Based on the correspondence between the source image corner point set and the target corner point set, a homography transformation model is used to solve the homography matrix that maps the original image coordinates to the orthogonal viewpoint coordinates.

[0080] Finally, perspective transformation processing is performed on the original rebar image based on the homography matrix to correct the perspective distortion in the original image into an image representation under orthogonal viewpoint, resulting in a geometrically corrected rebar image. The size ratio and spatial relationship of the rebar mesh in the corrected image are consistent with the actual engineering scene, and are used for subsequent rebar identification, geometric parameter measurement and acceptance judgment processes.

[0081] Image enhancement includes contrast-limited adaptive histogram equalization, Gaussian filtering, and gamma correction. The processing steps include:

[0082] Image processing employs contrast-limited adaptive histogram equalization for adjusting local shadows and lighting;

[0083] Noise suppression is achieved by using a Gaussian filter with a preset kernel size and standard deviation.

[0084] Adjust the gamma coefficient to correct the brightness of the image;

[0085] The gamma coefficient ranges from 0.8 to 1.2.

[0086] Specifically, when performing image enhancement processing on the steel bar image, the original image is used as input. First, the image is subjected to contrast-limited adaptive histogram equalization. By dividing the image into multiple local regions and performing histogram equalization on each region, while limiting the contrast gain, the image local details are enhanced while suppressing over-enhancement, thereby improving the effect of local shadow areas and uneven lighting.

[0087] After contrast enhancement, Gaussian filtering with preset kernel size and standard deviation is used to smooth the image. By weighted averaging of neighboring pixels, random noise interference introduced by the shooting environment, sensor noise, or compression process is reduced. While maintaining the overall structural continuity of the steel reinforcement edge, the overall smoothness of the image is improved. The gamma coefficient is set according to the overall brightness distribution of the image, and gamma correction is performed on the image under the condition that the gamma coefficient value ranges from 0.8 to 1.2. The image brightness is adjusted through non-linear brightness mapping, which enhances the dark areas and suppresses the bright areas, thereby making the brightness contrast between the steel reinforcement area and the background more obvious.

[0088] After the above image enhancement processing, the rebar image is improved in terms of contrast, noise level and brightness distribution, providing a more stable and clear image foundation for subsequent rebar edge extraction, target recognition and geometric parameter calculation.

[0089] The improved deep learning model is based on the YOLOv8 architecture and includes a backbone network, a neck network, and a multi-task detection head. The specific structure is as follows:

[0090] The backbone network replaces standard convolutions with depthwise separable convolutions and embeds a convolutional block attention mechanism, including a cross-stage local network and a spatial pyramid pooling fusion module;

[0091] The neck network employs a fusion structure of feature pyramid and path aggregation network for multi-scale feature fusion;

[0092] The multi-task detection head includes a classification branch, a regression branch, a diameter classification branch, and a key point branch. The diameter classification branch is used to classify and identify the specifications of the reinforcing bars.

[0093] Specifically, when processing rebar images, a deep learning model based on an improved YOLOv8 architecture is used as the recognition model. This model consists of a backbone network, a neck network, and a multi-task detection head, used for comprehensive detection and analysis of rebar targets and their related attributes. While maintaining the overall network hierarchy, the backbone network replaces the original standard convolutional structure with depthwise separable convolutions. By separating spatial convolution and channel convolution for computation, the number of model parameters and computational complexity are reduced while maintaining the ability to extract rebar texture and edge features. At the same time, a convolutional block attention mechanism is embedded in the backbone network. By weighting and adjusting the channel and spatial dimensions of features, the network pays more attention to the salient features of the rebar area and reduces the influence of background interference information. The backbone network also introduces a cross-stage local network structure and a spatial pyramid pooling fusion module. Through cross-layer feature connections and multi-scale receptive field fusion, the feature representation ability of rebar targets of different sizes is improved.

[0094] The neck network adopts a fusion structure that combines feature pyramid and path aggregation network, which performs bidirectional fusion of features from different levels from top to bottom and bottom to top, so as to fully combine high-level semantic information and low-level detailed information, thereby maintaining the stability of rebar target localization and recognition under multi-scale conditions. It is suitable for construction scenarios with dense rebar distribution or large scale differences.

[0095] The multi-task detection head is configured with multiple parallel branch structures, including a classification branch for judging the target category of rebar, a regression branch for rebar location regression, a diameter classification branch for rebar specification identification, and a key point branch for extracting key structural information of rebar. The diameter classification branch classifies and judges the appearance features of the rebar and outputs the corresponding rebar specification category. Through the collaborative work of the above multi-branch structures, the position, category, specification information and key point information of the rebar are output simultaneously in a single detection process, which achieves the effect of reducing redundant calculations and improving the overall detection efficiency.

[0096] Pixel precision is calibrated using one of the following two methods:

[0097] Pixel accuracy is calculated using the sensor width, shooting distance, image width, and focal length of the mobile terminal camera. Pixel accuracy is the ratio of the product of the sensor width and shooting distance to the product of the image width and focal length.

[0098] Pixel precision is calculated by using a reference object with a known actual length in the image and its corresponding image pixel length. Pixel precision is the ratio of the actual length of the reference object to the pixel length of the reference object's image.

[0099] Specifically, before measuring the geometric parameters of the steel reinforcement image, it is necessary to calibrate the pixel accuracy of the image in order to establish the correspondence between the image pixel size and the actual engineering size. Depending on the site conditions and equipment configuration, any of the following methods can be used to calibrate the pixel accuracy.

[0100] In the first calibration method, the pixel accuracy is calculated using the imaging parameters of the mobile terminal camera, including the physical width of the camera sensor, the distance between the camera and the steel bar during shooting, the pixel width of the captured image, and the focal length parameter of the camera. The pixel accuracy under the current shooting conditions is obtained by calculating the ratio of the product of the sensor width and the shooting distance to the product of the image width and the focal length, thereby completing the conversion between pixels and actual size without introducing additional reference objects.

[0101] In the second calibration method, a reference object with a known actual length is set or selected in the shooting image, including a standard ruler or the edge of a component with known specifications, and the pixel length corresponding to the reference object is measured in the image. The pixel accuracy is obtained by calculating the ratio between the actual length of the reference object and its pixel length in the image. This method can intuitively and reliably calibrate the pixel accuracy when the shooting parameters change or the on-site conditions are complex.

[0102] By using any of the above pixel accuracy calibration methods, a mapping relationship between image space and actual engineering dimensions is established, so that the subsequent calculation results of geometric parameters such as rebar spacing, diameter and length have clear physical meaning and meet the accuracy requirements of engineering measurement and acceptance judgment.

[0103] The calculation process for rebar spacing includes the following steps:

[0104] The coordinates of the center points of the reinforcing bars are sorted according to a preset direction to obtain the sorted set of center point coordinates;

[0105] Calculate the pixel distance between two adjacent points in the sorted center point coordinate set;

[0106] The product of pixel distance and pixel precision is the actual spacing of the reinforcing bars.

[0107] Specifically, the center point coordinates of each steel bar in the image are extracted based on the steel bar recognition results, and the sorting direction is pre-determined according to the actual laying direction of the steel bars. The center point coordinates of each steel bar are sorted according to the determined direction to obtain a set of center point coordinates that are consistent with the actual arrangement order of the steel bars, thereby avoiding the impact of the detection order difference on the spacing calculation results.

[0108] After sorting the center points, the pixel distance between any two adjacent center points in the sorted center point coordinate set is calculated sequentially. The pixel distance is obtained using Euclidean distance calculation to reflect the spacing between adjacent rebars in the image space. The calculated pixel distance is then multiplied by the calibrated pixel precision to obtain the corresponding actual rebar spacing value. Through this calculation process, the pixel distance in the image is accurately converted into actual engineering dimensions, ensuring that the rebar spacing measurement results meet the requirements for construction quality inspection and acceptance.

[0109] The calculation process for the diameter of reinforcing bars includes the following steps:

[0110] Obtain the width of the rebar boundary box output by AI intelligent analysis;

[0111] Calculate the angle between the steel bar axis and the horizontal direction of the image;

[0112] The product of the bounding box width, pixel precision, and the cosine of the included angle is the actual diameter of the reinforcing bar.

[0113] Specifically, the first step is to obtain the width of the rebar bounding box output by AI intelligent analysis. This width reflects the pixel scale of the rebar in the image. The angle between the rebar's axis and the horizontal direction of the image is then calculated. This angle can be determined through geometric analysis of the rebar's contour features and the horizontal line in the image, using the angle relationship between the line segment and the horizontal line to determine the actual direction of the rebar.

[0114] The obtained bounding box width is multiplied by the calibrated pixel precision to obtain the pixel size of the rebar. To correct the distortion caused by the rebar's tilt, it is further multiplied by the cosine of the included angle, thereby converting the width in the image into the actual diameter of the rebar. The actual diameter of the rebar can be accurately extracted from the image, meeting the requirements of construction drawings and actual engineering for the rebar diameter size, and providing reliable data support for subsequent quality control and acceptance judgment.

[0115] The calculation process for the length of reinforcing bars includes the following steps:

[0116] First, extract the pixel coordinates of the two ends of the rebar;

[0117] The endpoint pixel coordinates are converted to world coordinates using the inverse of the homography matrix;

[0118] Calculate the Euclidean distance between the two endpoints in world coordinates to obtain the actual length of the reinforcing bar.

[0119] Specifically, based on the results of rebar identification and key point extraction, the pixel coordinates of the two ends of each rebar in the image are obtained to represent the starting and ending positions of the rebar in the image space. According to the homography matrix obtained in the aforementioned geometric correction process, the homography matrix is ​​inverted, and its inverse matrix is ​​used to perform coordinate transformation on the pixel coordinates of the two ends of the rebar, transforming the endpoints from the image coordinate system to the corresponding world coordinate system, thereby eliminating the influence of shooting angle and perspective relationship on the length measurement results.

[0120] After coordinate transformation, the coordinates of the two ends of the rebar in the world coordinate system are obtained, and the Euclidean distance between the two ends is calculated to obtain the actual length of the rebar. Through the above calculation process, the pixel length in the image is accurately mapped to the actual engineering dimension, giving the rebar length measurement result a clear physical meaning and meeting the needs of engineering measurement and acceptance applications.

[0121] The acceptance rule library contains built-in standards for the acceptance of construction quality of concrete structures.

[0122] During the acceptance and judgment process, the acceptance tolerance threshold is dynamically adjusted according to the steel bar obstruction rate and binding status. When the confidence level of the test result is lower than the preset threshold, the test result is marked as requiring manual review.

[0123] During the acceptance and judgment process, quantitative indicators for rebar obstruction rate and binding status are set, and the acceptance tolerance threshold is dynamically adjusted according to the preset ratio based on the quantitative indicators.

[0124] Specifically, when conducting acceptance and judgment of the reinforcement construction quality, the system incorporates the relevant technical requirements and allowable deviation standards in the construction quality acceptance specifications for concrete structure engineering, and transforms the various specification clauses into acceptance rules that can be used for calculation and comparison, thus constructing an acceptance rule library for automatically judging the reinforcement parameter results obtained from the test.

[0125] During the acceptance assessment process, the occlusion and binding status of the reinforcing bars are first analyzed based on the results of rebar recognition and image analysis. The occlusion rate is quantified by the proportion of the area of ​​the obscured rebar in the image, while the binding status is determined by the integrity and morphological characteristics of the rebar intersections and binding points. Corresponding quantitative index values ​​are generated for each. Based on the quantitative indices of the occlusion rate and binding status, the acceptance tolerance threshold is dynamically adjusted according to a pre-set proportional relationship. When the occlusion rate is high or the binding status is unclear, the acceptance tolerance threshold is appropriately relaxed to reduce the risk of misjudgment due to incomplete image information. When the occlusion rate is low and the binding status is clear, a standard or tightened acceptance tolerance threshold is used to improve the accuracy and rigor of the acceptance assessment.

[0126] During the acceptance evaluation process, a comprehensive judgment is made based on the confidence level of the test results. When the confidence level of a certain test result is lower than a preset threshold, even if its calculated result is within the adjusted tolerance range, the result is marked as requiring manual review, and a prompt is included in the acceptance results. This approach enables the acceptance evaluation process to dynamically adjust while adhering to specifications, improving the reliability and applicability of acceptance results in complex construction scenarios.

[0127] Reference Figure 2 A fast, accurate, and efficient rebar acceptance system based on AI, comprising a hardware layer, a support layer, a core service layer, and an application layer, wherein...

[0128] The hardware layer also includes a camera module, a computing unit, a storage unit, and an inertial measurement unit. The camera module is used to acquire image data of the steel bars, the computing unit is used to provide local computing capabilities, the storage unit is used to cache data and model files, and the inertial measurement unit is used to acquire motion state data.

[0129] The support layer also includes a mobile inference framework, a model management module, and a device adaptation module. The mobile inference framework provides a model runtime environment, the model management module is used for model loading, version control, and dynamic updates, and the device adaptation module is used to detect hardware performance and allocate computing resources.

[0130] The core service layer also includes an image preprocessing module, an AI analysis engine, a geometric calculation module, and an acceptance rule engine. The image preprocessing module is used to perform image correction and enhancement operations, the AI ​​analysis engine is used to deploy improved deep learning models and identify steel bars, the geometric calculation module is used for parameter measurement, and the acceptance rule engine is used to perform acceptance judgments.

[0131] The application layer also includes a user interface module, a data management module, and a report generation module. The user interface module provides image acquisition and interaction functions, the data management module is used for project management and data storage, and the report generation module is used to output visualization results and acceptance reports.

[0132] Specifically, at the hardware layer, the camera module is used to acquire image data of steel bars at the construction site and transmit the image data to the computing unit in real time; the computing unit is a processor in a mobile terminal or embedded device, used to perform operations such as image preprocessing, model inference, and data calculation; the storage unit is used to cache the acquired image data, intermediate calculation results, and deep learning model files, supporting offline or breakpoint resume use; the inertial measurement unit is used to acquire motion state data such as the device's attitude, acceleration, and angular velocity, providing basic data support for subsequent image geometric correction and shooting posture compensation;

[0133] In the support layer, the mobile inference framework provides a unified inference environment for the operation of deep learning models, supporting the efficient execution of models on different mobile terminals; the model management module is responsible for loading, switching, version control and dynamic updates of models, enabling the system to flexibly adjust the recognition model according to actual needs; the device adaptation module is used to detect the hardware performance of the current terminal and allocate computing resources reasonably accordingly, so as to improve operating efficiency while ensuring recognition accuracy.

[0134] In the core service layer, the image preprocessing module receives raw image data from the hardware layer and performs image correction and enhancement to improve image quality; the AI ​​analysis engine deploys an improved deep learning model to identify rebar targets and extract features from the preprocessed image; the geometric calculation module completes the measurement calculation of parameters such as rebar spacing, diameter, and length based on the recognition results and pixel accuracy information; the acceptance rule engine combines the acceptance rule library to make a comprehensive judgment on the measurement results and outputs an acceptance conclusion that conforms to the engineering specifications.

[0135] In the application layer, the user interface module provides users with entry points for image acquisition, result viewing, and interactive operations; the data management module is used for unified management and storage of engineering projects, test records, and historical data; and the report generation module generates visual display content and standardized acceptance reports based on test and acceptance results, facilitating result archiving, sharing, and subsequent verification.

[0136] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fast, accurate, and efficient method for inspecting reinforcing bars based on AI, characterized in that, Includes the following steps: Images of the steel reinforcement construction area are acquired through the mobile terminal's camera, and data from the mobile terminal's inertial measurement unit, shooting distance, and ambient light intensity are recorded simultaneously. The acquired raw images are sequentially subjected to geometric correction and image enhancement processing to obtain optimized images; The optimized image is input into the improved deep learning model, which simultaneously outputs the bounding box coordinates, category confidence, diameter classification results, and center point coordinates of the reinforcing bars. Based on the pixel-level data output by AI recognition, combined with the calibrated pixel precision, the physical parameters of rebar spacing, diameter and length are calculated; The acceptance rule library is called up to compare the calculated parameters with the preset standards to determine the passability of individual items and the whole; Generates visualized results charts with annotation information and structured acceptance reports, and stores key acceptance data.

2. The AI-based method for rapid, accurate, and efficient steel reinforcement acceptance according to claim 1, characterized in that, The geometric correction employs a perspective correction method based on homography transformation, and the process includes the following steps: Edge detection operators are used to extract the edges of the reinforcing mesh, and line detection algorithms are used to detect the intersections of lines. The four corner points of the reinforcing mesh are located as the corner point set of the source image. Define the target corner point set under orthogonal viewpoints, and calculate the homography matrix based on the source image corner point set and the target corner point set; A perspective transformation is performed on the original image based on the homography matrix to obtain a corrected image under orthogonal viewpoints.

3. The AI-based method for rapid, accurate, and efficient steel reinforcement acceptance according to claim 1, characterized in that, The image enhancement includes contrast-limited adaptive histogram equalization, Gaussian filtering, and gamma correction. The processing steps include: Image processing employs contrast-limited adaptive histogram equalization for adjusting local shadows and lighting; Noise suppression is achieved by using a Gaussian filter with a preset kernel size and standard deviation. Adjust the gamma coefficient to correct the brightness of the image.

4. The AI-based method for rapid, accurate, and efficient steel reinforcement acceptance according to claim 1, characterized in that, The improved deep learning model is based on the YOLOv8 architecture and includes a backbone network, a neck network, and a multi-task detection head. Its specific structure is as follows: The backbone network replaces standard convolution with depthwise separable convolution and embeds a convolutional block attention mechanism, including a cross-stage local network and a spatial pyramid pooling fusion module. The neck network adopts a fusion structure of feature pyramid and path aggregation network for multi-scale feature fusion; The multi-task detection head includes a classification branch, a regression branch, a diameter classification branch, and a key point branch. The diameter classification branch is used for the classification and identification of rebar specifications.

5. The AI-based method for rapid, accurate, and efficient steel reinforcement acceptance according to claim 1, characterized in that, The pixel precision is determined by one of the following two methods: Pixel accuracy is calculated using the sensor width, shooting distance, image width, and focal length of the mobile terminal camera. The pixel accuracy is the ratio of the product of the sensor width and shooting distance to the product of the image width and focal length. Pixel precision is calculated by using a reference object with a known actual length in the image and its corresponding image pixel length. Pixel precision is the ratio of the actual length of the reference object to the pixel length of the reference object's image.

6. The AI-based method for rapid, accurate, and efficient steel reinforcement acceptance according to claim 1, characterized in that, The calculation process for the spacing of the reinforcing bars includes the following steps: The coordinates of the center points of the reinforcing bars are sorted according to a preset direction to obtain the sorted set of center point coordinates; Calculate the pixel distance between two adjacent points in the sorted center point coordinate set; The product of pixel distance and pixel precision is the actual spacing of the reinforcing bars.

7. The AI-based method for rapid, accurate, and efficient steel reinforcement acceptance according to claim 1, characterized in that, The calculation process for the diameter of the reinforcing bar includes the following steps: Obtain the width of the rebar boundary box output by AI intelligent analysis; Calculate the angle between the steel bar axis and the horizontal direction of the image; The product of the bounding box width, pixel precision, and the cosine of the included angle is the actual diameter of the reinforcing bar.

8. The AI-based method for rapid, accurate, and efficient steel reinforcement acceptance according to claim 1, characterized in that, The calculation process for the length of the reinforcing bar includes the following steps: First, extract the pixel coordinates of the two ends of the rebar; The endpoint pixel coordinates are converted to world coordinates using the inverse of the homography matrix; Calculate the Euclidean distance between the two endpoints in world coordinates to obtain the actual length of the reinforcing bar.

9. The AI-based method for rapid, accurate, and efficient steel reinforcement acceptance according to claim 1, characterized in that, The acceptance rule library contains the construction quality acceptance specifications for concrete structure engineering. During the acceptance judgment process, the acceptance tolerance threshold is dynamically adjusted according to the steel bar obstruction rate and binding status. When the confidence level of the test result is lower than the preset threshold, the test result is marked as requiring manual review. In the acceptance judgment process, quantitative indicators of rebar obstruction rate and binding status are set, and the acceptance tolerance threshold is dynamically adjusted according to the quantitative indicators at a preset ratio.

10. A fast, accurate, and efficient rebar acceptance system based on AI, employing the fast, accurate, and efficient rebar acceptance method based on AI as described in any one of claims 1-9, characterized in that, The system comprises a hardware layer, a support layer, a core service layer, and an application layer, wherein: The hardware layer also includes a camera module, a computing unit, a storage unit, and an inertial measurement unit. The camera module is used to acquire image data of the reinforcing bars, the computing unit is used to provide local computing capabilities, the storage unit is used to cache data and model files, and the inertial measurement unit is used to acquire motion state data. The support layer also includes a mobile inference framework, a model management module, and a device adaptation module. The mobile inference framework is used to provide a model running environment, the model management module is used for model loading, version control, and dynamic updates, and the device adaptation module is used to detect hardware performance and allocate computing resources. The core service layer also includes an image preprocessing module, an AI analysis engine, a geometric calculation module, and an acceptance rule engine. The image preprocessing module is used to perform image correction and enhancement operations, the AI ​​analysis engine is used to deploy improved deep learning models and identify rebars, the geometric calculation module is used for parameter measurement, and the acceptance rule engine is used to perform acceptance judgments. The application layer also includes a user interface module, a data management module, and a report generation module. The user interface module provides image acquisition and interaction functions, the data management module is used for project management and data storage, and the report generation module is used to output visualization results and acceptance reports.

Images