A reference-free visual assessment method and system for the spacing of stirrups in a pile foundation reinforcement cage

Abstract

The present application relates to the technical field of pile foundation engineering quality detection, and particularly relates to a pile foundation reinforcement cage stirrup spacing non-reference visual evaluation method and system. The method comprises the following steps: acquiring a single frame image of a pile foundation reinforcement cage; constructing an intersection recognition model, and recognizing effective intersections of main reinforcement and stirrups in the single frame image by using the intersection recognition model; calculating a stirrup pixel spacing representative value, a proportionality coefficient and a standard uncertainty based on the effective intersections; calculating a true spacing and an expanded uncertainty based on the proportionality coefficient, the standard uncertainty and the stirrup pixel spacing representative value; and generating an evaluation result of the pile foundation reinforcement cage stirrup spacing based on the true spacing and the expanded uncertainty. The present application solves the problem of insufficient intelligence and accuracy of pile foundation reinforcement cage stirrup spacing evaluation in the prior art, and improves the intelligence and accuracy of evaluation by improving the construction of a high-precision intersection recognition model, realizing accurate recognition of intersections, and deducing the true spacing and the expanded uncertainty in combination with parameters and uncertainty.

Description

Technical Field

[0001] This invention relates to the field of pile foundation engineering quality testing technology, specifically a referenceless visual evaluation method and system for the spacing of stirrups in pile foundation reinforcement cages. Background Technology

[0002] As the core load-bearing component of bored cast-in-place piles, the processing accuracy of the stirrup spacing in the reinforcing cage directly affects the shear bearing capacity and seismic performance of the pile foundation. Both the current "Code for Acceptance of Construction Quality of Concrete Structures" and the "Technical Code for Building Pile Foundations" clearly stipulate that the allowable deviation of the stirrup spacing shall not exceed ±20mm. This indicator is a mandatory inspection item for the factory acceptance of reinforcing cages.

[0003] Currently, processing plants commonly use manual steel ruler sampling for inspection. Quality inspectors must bend over to measure each section and record the data by hand, with each cage inspection taking approximately 15-20 minutes. This method is inefficient and prone to human error. In recent years, computer vision-based inspection methods have begun to be applied to rebar cage measurement. These methods can be categorized into two types: the first involves attaching rulers or targets to the surface of the rebar cage, using a known reference object to convert pixels to actual dimensions. However, carrying targets on-site is inconvenient, and targets are easily damaged or detached during processing. The second method relies on manually inputting the rebar's design diameter as a calibration benchmark. However, the actual diameter of rebar after cold drawing and rolling has manufacturing tolerances, and on-site input of design data increases operational complexity and the risk of errors. Both methods fail to eliminate reliance on external references or manual parameters, limiting their practicality in the high-frequency, fast-paced acceptance scenarios of processing plants. Furthermore, the complex environment of rebar processing plants, with its strong arc welding, direct sunlight, and dust obstruction, leads to problems such as high dynamic range, local overexposure or underexposure, and lens distortion in the acquired images. In this environment, general object detection algorithms are prone to missing or falsely detecting intersections, which affects the accuracy of distance calculation. At the same time, existing vision methods usually directly output all detected distance values ​​without considering the systematic errors caused by lens edge distortion, and have not established a measurement uncertainty assessment model corresponding to the acceptance specifications, making it difficult to meet the legal metrological requirements for conformity judgment.

[0004] In summary, the industry urgently needs an automated detection method for the spacing of steel cage stirrups that requires no physical reference, no manual input of design parameters, is resistant to light interference in the processing area, and can output acceptance-level accuracy and uncertainty assessment results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a referenceless visual evaluation method and system for the spacing of stirrups in pile foundation reinforcement cages, solving the problem that the evaluation of stirrup spacing in pile foundation reinforcement cages is not intelligent and accurate enough in existing technologies.

[0006] To achieve the above objectives, the present invention provides a referenceless visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage. The method includes: acquiring a single-frame image of the pile foundation reinforcement cage; constructing an intersection point recognition model and using the intersection point recognition model to identify the effective intersection points of the main reinforcement and stirrups in the single-frame image; calculating a representative value of the stirrup pixel spacing, a scaling factor, and a standard uncertainty based on the effective intersection points; calculating the true spacing and expanded uncertainty based on the scaling factor, the standard uncertainty, and the representative value of the stirrup pixel spacing; and generating an evaluation result of the stirrup spacing of the pile foundation reinforcement cage based on the true spacing and the expanded uncertainty.

[0007] This invention acquires a single-frame image of the pile foundation reinforcement cage and accurately identifies the effective intersection points of the main reinforcement and stirrups using an intersection point recognition model, avoiding errors from manual measurement. Based on the effective intersection points, it calculates core parameters and uncertainties to ensure data rigor. Combining parameters and uncertainties, it derives the true spacing and expanded uncertainties to improve measurement accuracy. Finally, it generates evaluation results, achieving automated judgment and improving the intelligence and accuracy of pile foundation reinforcement cage stirrup spacing detection and evaluation.

[0008] Optionally, obtaining a single-frame image of the pile foundation reinforcement cage includes: obtaining an original single-frame image of the pile foundation reinforcement cage and lens distortion parameters; using the lens distortion parameters to perform distortion correction on the original single-frame image to obtain a corrected single-frame image; and performing grayscale conversion and Gaussian smoothing on the corrected single-frame image to obtain a single-frame image of the pile foundation reinforcement cage.

[0009] This invention obtains the original single-frame image and lens distortion parameters to provide a precise basis for image correction. Using these parameters, distortion correction is performed to eliminate the morphological distortion of the main reinforcement and stirrups caused by lens optical deviation. The corrected image is then grayscaled to remove redundant color information to reduce computational complexity. Finally, Gaussian smoothing is applied to suppress environmental noise and arc light interference, thereby optimizing image quality and improving the data quality of the single-frame image.

[0010] Optionally, constructing the intersection recognition model and using the intersection recognition model to identify valid intersections of main bars and stirrups in the single-frame image includes: introducing a YOLOv10 model, improving the YOLOv10 model using a pre-set deformable convolution module and a lighting perception module to obtain an improved YOLOv10 model; training the improved YOLOv10 model based on pre-acquired training data to obtain an intersection recognition model; inputting the single-frame image into the intersection recognition model to obtain initial intersections and confidence scores; and filtering the initial intersections based on the confidence scores to obtain valid intersections.

[0011] This invention introduces the YOLOv10 model, combines it with a deformable convolution module to adapt to the geometric deformation of oblique intersection points, and utilizes a lighting sensing module to resist complex lighting interference in the processing field, resulting in an improved YOLOv10 model. After training, an intersection point recognition model is obtained. A single-frame image is input into the model to output the initial intersection points and confidence scores, achieving accurate recognition. Intersection points are filtered according to confidence scores, eliminating low-confidence and duplicate results, thereby improving the accuracy of effective intersection points.

[0012] Optionally, the step of calculating the representative value of the stirrup pixel spacing, the proportionality coefficient, and the standard uncertainty based on the effective intersection points includes: performing linear fitting on the effective intersection points using a random sampling consensus algorithm to obtain a set of horizontal lines for the main reinforcement and a set of vertical lines for the stirrups; determining the central main reinforcement and the upper and lower main reinforcements based on the set of horizontal lines for the main reinforcement; determining the intersection points of the main reinforcements based on the central main reinforcement, the upper and lower main reinforcements, and the set of vertical lines for the stirrups, and calculating the Euclidean distance between adjacent intersection points of the main reinforcements to obtain a pixel spacing sequence for the central main reinforcement and a pixel spacing sequence for the upper and lower main reinforcements; performing median filtering on the pixel spacing sequence for the central main reinforcement and the pixel spacing sequence for the upper and lower main reinforcements respectively to obtain an optimized central pixel spacing sequence and an optimized upper and lower pixel spacing sequence; performing a weighted average on the optimized central pixel spacing sequence and the optimized upper and lower pixel spacing sequence to obtain a representative value of the stirrup pixel spacing; obtaining a prior distribution, and calculating the proportionality coefficient and the standard uncertainty based on the prior distribution and the optimized central pixel spacing sequence.

[0013] This invention uses a random sampling consensus algorithm to fit effective intersection points, accurately separates the straight sets of main reinforcement and stirrups, filters the central and upper and lower main reinforcements, focuses on high-quality imaging areas, calculates the Euclidean distance between adjacent intersection points, obtains the original spacing data, removes outliers by median filtering, optimizes data reliability, generates representative values ​​by weighted averaging, highlights the weight of core areas, and combines prior distribution and optimized sequence to estimate key parameters, ensuring measurement rigor and improving the accuracy of stirrup spacing-related parameter calculations.

[0014] Optionally, the pixel spacing of the stirrups represents the value of the following formula: , in, This represents the pixel spacing of the stirrups. For median operations, To optimize the pixel spacing sequence, To optimize the central pixel spacing sequence, To optimize the pixel spacing sequence.

[0015] This invention optimizes the lower, central, and upper pixel spacing sequences by taking the median values, and then weights them according to a weight ratio of 0.25:0.5:0.25 to calculate the representative value of the stirrup pixel spacing. The median calculation eliminates abnormal spacing caused by interference such as local occlusion and arcing, ensuring the robustness of single-sequence data. The differentiated weights highlight the dominant role of the clearest imaging region, while incorporating upper and lower sequence data to take into account the overall distribution characteristics, thus improving the stability and scientific validity of the representative value of the stirrup pixel spacing.

[0016] Optionally, the calculation of the proportionality coefficient and standard uncertainty based on the prior distribution and the optimized central pixel spacing sequence includes: obtaining a sample of the main reinforcement pixel diameter based on the central main reinforcement; obtaining a sample of the main reinforcement spacing based on the set of vertical lines of the stirrups and the upper and lower main reinforcements; obtaining a sample of the stirrup spacing based on the optimized central pixel spacing sequence; and using the expectation-maximization algorithm with the prior distribution as a constraint, jointly estimating the sample of the main reinforcement pixel diameter, the sample of the main reinforcement spacing, and the sample of the stirrup spacing to obtain the proportionality coefficient and standard uncertainty.

[0017] This invention extracts three types of samples from the central main reinforcement, stirrups, and upper and lower main reinforcement respectively, to achieve multi-source data complementarity and enrich the basis for parameter estimation. By using prior distribution as a constraint, it avoids the parameter drift problem without reference calibration, ensuring that the results meet engineering specifications. The expectation-maximization algorithm is used to jointly estimate the three types of samples, integrate multi-dimensional information to reduce noise interference from single data, and improve the estimation accuracy of the proportional coefficient and standard uncertainty.

[0018] Optionally, the step of using the expectation-maximization algorithm to jointly estimate the proportionality coefficient and standard uncertainty of the main reinforcement pixel diameter samples, the main reinforcement spacing samples, and the stirrup spacing samples, constrained by the prior distribution, includes: calculating the median diameter based on the main reinforcement pixel diameter samples; obtaining a national standard diameter library and determining an initial proportionality coefficient based on the median diameter and the national standard diameter library; obtaining true size samples of the main reinforcement pixel diameter, the main reinforcement spacing, and the stirrup spacing based on the initial proportionality coefficient, the main reinforcement pixel diameter samples, the main reinforcement spacing samples, and the stirrup spacing samples; constructing a three-channel joint likelihood function using the true size samples of the main reinforcement pixel diameter, the main reinforcement spacing, and the stirrup spacing, constrained by the prior distribution; obtaining an updated proportionality coefficient by maximizing the three-channel joint likelihood function; iterating based on the updated proportionality coefficient until a pre-set convergence condition is met to obtain the proportionality coefficient; and calculating the standard uncertainty based on the proportionality coefficient and the three-channel joint likelihood function.

[0019] This invention calculates the median diameter and determines the initial proportionality coefficient using a national standard diameter database, ensuring that the initial value is compliant and close to the true value. Based on the initial value, it transforms three types of real-size samples to provide high-quality data for subsequent estimation. A three-channel joint likelihood function is constructed using prior distribution constraints, and multi-dimensional information is integrated to avoid parameter drift. The likelihood function is iteratively maximized using the expectation-maximization algorithm, and the optimal solution is approximated by the monotonically convergent nature of the algorithm. Finally, the standard uncertainty is calculated by combining the proportionality coefficient and the likelihood function, ensuring measurement rigor and improving the estimation accuracy and reliability of the proportionality coefficient and standard uncertainty.

[0020] Optionally, calculating the true spacing and expanded uncertainty based on the scaling factor, the standard uncertainty, and the representative value of the stirrup pixel spacing includes: multiplying the scaling factor by the representative value of the stirrup pixel spacing to obtain the true spacing; and calculating the expanded uncertainty based on the standard uncertainty and the representative value of the stirrup pixel spacing.

[0021] This invention achieves a precise mapping from pixel size to actual physical size by multiplying a proportionality coefficient by a representative value of stirrup pixel spacing, directly obtaining the true spacing that reflects the average distribution density of stirrups. Then, based on the standard uncertainty and the representative value of stirrup pixel spacing, the expanded uncertainty is calculated to quantify the reliable interval of the true spacing, thereby improving the accuracy and reliability of the stirrup spacing measurement results of pile foundation steel cages.

[0022] Optionally, the evaluation result of the pile foundation reinforcement cage stirrup spacing based on the actual spacing and the expanded uncertainty includes: obtaining the design spacing and tolerance threshold of the pile foundation reinforcement cage stirrups; determining that when the expanded uncertainty is less than the tolerance threshold, subtracting the expanded uncertainty from the tolerance threshold to obtain a corrected tolerance threshold; calculating the absolute difference between the actual spacing and the design spacing, and comparing the absolute difference with the corrected tolerance threshold to obtain the pile foundation reinforcement cage stirrup spacing evaluation result.

[0023] This invention obtains the tolerance threshold for the design spacing to match the specification, clarifies the qualification judgment criteria, and obtains the corrected tolerance threshold by subtracting the expanded uncertainty from the tolerance threshold when the expanded uncertainty is less than the tolerance threshold. This avoids the risk of misjudgment caused by measurement error. The absolute difference between the actual spacing and the design spacing is calculated and compared with the corrected tolerance threshold, making the judgment logic more rigorous and improving the scientificity and accuracy of the evaluation results of the pile foundation reinforcement cage stirrup spacing.

[0024] Another aspect of the present invention provides a referenceless visual evaluation system for the spacing of stirrups in a pile foundation reinforcement cage, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute a referenceless visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage as described in any of the preceding aspects of the present invention.

[0025] The present invention provides a referenceless visual evaluation system for the spacing of stirrups in a pile foundation reinforcement cage. It has a compact structure, stable performance, high integration, and simple configuration. It can stably perform the referenceless visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage provided in the preceding aspect of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0026] Figure 1 This is a flowchart of a non-reference visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a referenceless visual evaluation system for the spacing of stirrups in a pile foundation reinforcement cage, according to an embodiment of the present invention. Detailed Implementation

[0027] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0028] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0029] Please see Figure 1 In order to solve the problems in the prior art, in an alternative embodiment, such as Figure 1The method for non-reference visual evaluation of stirrup spacing in pile foundation reinforcement cages, as shown, includes the following steps: Step S1: Obtain a single-frame image of the pile foundation reinforcement cage.

[0030] The process of acquiring a single-frame image of the pile foundation reinforcement cage includes the following sub-steps: Step S101: Obtain the original single-frame image and lens distortion parameters of the pile foundation reinforcement cage.

[0031] In this embodiment, the quality inspection is conducted in the rebar processing yard after the finished rebar cage is fabricated and before it is hoisted into the borehole. The rebar cage is laid horizontally in the rebar processing yard, with its main bars parallel to the support axis in a horizontal straight line, and the stirrups surrounding the main bars in a vertical straight line, forming a grid structure. A static image containing the complete rebar cage grid and whose imaging quality meets the requirements of sub-pixel detection is captured, i.e., the original single-frame image, providing a unique data source for subsequent intersection point positioning, geometric screening, and referenceless calibration.

[0032] Using a regular tablet or smartphone as the image acquisition device, no external supplementary lighting, targets, rulers, or other auxiliary measuring devices are required. Testing has shown that the device's camera has an optical focal length of at least 24mm (35mm equivalent focal length) to ensure sufficient imaging field of view and distortion control. The device's built-in accelerometer and gyroscope assist in determining the shooting posture, ensuring the necessary vertical shooting conditions for subsequent calculations. Before shooting, the device's lens surface must be cleaned to avoid dust or oil contamination that could cause image blurring.

[0033] Shooting distance It should be controlled within the nominal diameter of the reinforcing cage. Within the range of 1.5 to 2.5 times, that is This distance range ensures that the main reinforcing bars occupy at least 200 pixels of vertical height in the image, meeting the sampling requirements for subpixel edge detection, while avoiding trapezoidal distortion and insufficient depth of field caused by excessively close shooting distance. When the diameter of the reinforcing cage is 600mm-1200mm, the corresponding shooting distance is 900mm-3000mm. During shooting, the device screen displays a reference frame in real time. When the entire reinforcing cage is within the frame and the edges have a margin of at least 50 pixels, the software interface displays a green ready signal.

[0034] The camera's optical axis must be perpendicular to the central axis of the steel cage, and the angle between the optical axis and the horizontal plane must be [missing information]. Should be kept to Between, that is The attitude constraint is achieved through real-time feedback from the device's built-in gyroscope. When the attitude angle exceeds this range, the acquisition software interface displays a red warning signal, and the shutter can only be triggered after the attitude is adjusted to comply with the requirements. A vertical shooting attitude minimizes the measurement error of the stirrup spacing caused by perspective projection, ensuring that the main reinforcement bars appear as horizontal straight lines in the original single-frame image. Preferably, the device is fixed to a tripod or handheld gimbal to maintain attitude stability.

[0035] The original single-frame image resolution should be no less than 1920×1080 pixels, preferably 4K resolution (3840×2160 pixels) to obtain a higher spatial sampling rate. The acquisition mode uses still single-frame shooting, without recording video sequences or performing multi-frame fusion. The image storage format is uncompressed RAW or lossless compressed PNG to avoid blocky artifacts introduced by JPEG compression interfering with edge detection. The shutter speed is automatically matched to the ambient lighting to ensure clear grayscale gradients at the edges of the steel bars. Through testing and comparison, the exposure time is controlled between 1 / 60s and 1 / 500s, and the ISO sensitivity is no higher than 800 to reduce noise.

[0036] This method is applicable to various lighting environments in steel rebar processing plants, including natural light, artificial lighting, and arc welding light, allowing for localized overexposure (brightness value greater than 250) or underexposure (brightness value less than 20) in the image. The acquisition software employs automatic exposure and automatic white balance strategies, with a dynamic range covering 0-120,000 lux. When the ambient light level is below 500 lux, the device's built-in LED supplementary lighting is automatically activated, but the supplementary lighting intensity is limited to within 200 lux to avoid specular reflection on the steel rebar surface. In strong backlight scenes, the software automatically triggers HDR mode, capturing two images with different exposures and performing pixel-level fusion to ensure that the intersection of dark and bright areas is simultaneously discernible. The fusion weight is automatically calculated based on the pixel brightness gradient to avoid ghosting artifacts.

[0037] Lens distortion parameters can be directly obtained through image acquisition equipment. These parameters are pre-determined and stored by the equipment manufacturer through a standard calibration process before leaving the factory, without the need for additional on-site calibration.

[0038] Step S102: Use the lens distortion parameters to perform distortion correction on the original single-frame image to obtain a corrected single-frame image.

[0039] In this embodiment, the acquired raw single-frame image is first subjected to distortion correction using the lens distortion parameters (radial distortion coefficient) calibrated at the device's factory settings. With tangential distortion coefficient The entire image is inversely mapped and corrected using the Brown-Conrady model, achieving a pixel mapping accuracy of 0.1 pixels, resulting in a corrected single-frame image.

[0040] Step S103: Perform grayscale conversion and Gaussian smoothing on the corrected single-frame image to obtain a single-frame image of the pile foundation reinforcement cage.

[0041] In this embodiment, the rectified single-frame image after distortion correction is first converted to grayscale, transforming the RGB color image into a single-channel grayscale image. The grayscale gradient between the rebar and the background is preserved to support subsequent edge detection. Then, a 3×3 Gaussian smoothing kernel is used to filter the grayscale image, and the standard deviation of the Gaussian function is set. By performing a weighted average operation on the pixel neighborhood, high-frequency noise such as dust and reflection in the processing environment is effectively suppressed, while the gray-scale gradient information of the steel bar edge is preserved to the maximum extent, and finally a single frame image of the pile foundation steel cage that meets the requirements of sub-pixel detection is obtained.

[0042] Step S2: Construct an intersection recognition model and use the intersection recognition model to identify the valid intersections of the main reinforcement and stirrups in the single frame image.

[0043] The specific steps of constructing an intersection recognition model and using the intersection recognition model to identify the valid intersections of the main reinforcement and stirrups in the single frame image include the following sub-steps: Step S201: Introduce the YOLOv10 model and improve it using a pre-set deformable convolution module and a light-sensing module to obtain an improved YOLOv10 model.

[0044] The YOLOv10 baseline network uses C2f modules as feature extraction units. Its convolutional kernels sample the feature map with a fixed grid, making it difficult to adapt to the oblique geometric deformation of steel reinforcement intersections. In the last two stages of the backbone network (corresponding to output strides of 16 and 32), the regular convolutions of the C2f modules are replaced with DCNv4 (DeformableConvolutionv4), forming deformable convolutional modules. DCNv4 introduces a normalized modulation mechanism on top of traditional deformable convolution, reducing the sampling offset. With modulation weights Parallel prediction allows the convolutional kernel shape to adapt to the local geometry of the intersection points. Specifically, for the input feature map X, the output feature Y is represented as: in, This represents the current output position coordinates (two-dimensional vector). The total number of sampling points. For the first Each sampling location For learnable two-dimensional offset, This is for normalized modulation weights. This mechanism enables the network to... In scenarios with large-angle oblique intersections, the sampling points are always focused on the intersection area rather than the background, effectively improving the recall rate of the oblique intersection points.

[0045] The dynamic range of illumination in a steel rebar processing plant reaches up to 120dB, and the static weights of traditional convolution cannot adapt to local overexposure and strong shadows. An IA-LAYER (Illumination-Aware Layer) is inserted between the backbone network and the FPN to form an illumination-aware module, enabling dynamic brightness compensation. The IA-LAYER structure consists of two steps: global illumination vector extraction and channel multiplicative modulation.

[0046] Global illumination vector extraction: Extracting the last layer feature map (size: ...) from the backbone network output. / 32× (32×512), using a Net VLAD structure to aggregate global illumination statistics. Specifically, the feature map spatial dimension is expanded to... 100 pixels (of which) ), using global cluster centers The feature space is centered. A 128-dimensional illumination vector is obtained through weighted aggregation. : in, This represents the total number of pixels after the final layer feature map is flattened. For the first Feature vectors of spatial locations Total number of visual words, As learnable cluster centers, For projection weights ( (This indicates transpose). This vector encodes the overall image brightness distribution, contrast, and highlight region locations, decoupled from the specific intersection content.

[0047] Channel multiplicative modulation: the extracted illumination vector The channel modulation coefficient vector is obtained by mapping to 512 dimensions through a learnable fully connected layer. The calculation process for a fully connected layer is as follows: The weight matrix With bias vector Optimization during model training ensures that the ReLU activation function... Each component is non-negative.

[0048] Then, the modulation coefficient With backbone network output feature map Perform channel-wise multiplicative modulation (Hadamard product) to generate illumination-compensated feature maps. Each of its spatial locations The calculation is as follows: in , These are the vertical and horizontal coordinate indices of the feature map, respectively. For channel index ( ), For the first The modulation coefficients of each channel are shared across all spatial locations in the feature map.

[0049] In the overexposed area of ​​the arc light, the network automatically learns... Suppress highlight clipping; in underexposed shadow areas, This mechanism enhances the network's robustness to dynamic lighting ranges of 0-120000 lux, significantly reduces the false negative rate at intersections, and eliminates the need for multi-frame HDR image synthesis.

[0050] The EMA-FPN+BiFPN bidirectional feature fusion addresses the issues of small rebar intersection size, insufficient resolution of high-level semantic feature maps, and weak semantics in low-level texture feature maps. It upgrades the conventional FPN of YOLOv10 to EMA-FPN (Exponential Moving AverageFPN) and introduces BiFPN bidirectional weighted fusion between multi-scale features.

[0051] The EMA-FPN structure introduces an exponential moving average mechanism in the top-down feature fusion path, causing high-level features to decay with a lower coefficient. The semantic aliasing is gradually propagated down to lower layers to avoid semantic aliasing caused by direct addition. The fusion formula is: in For the first Layer original features, Features after fusion This is an upsampling operation (2x bilinear interpolation). This mechanism allows shallow features to retain texture details while gaining stable high-level semantic support.

[0052] Furthermore, BiFPN employs cross-scale weighting in the FPN output. A bidirectional weighted connection is established between the three scales. For any two scales... and Its fusion weight Determined by learnable parameters, and the weights are guaranteed to sum to 1 through fast normalization: in, This represents the number of input feature maps (i.e., the total number of scales involved in the fusion). Index of input features ( ). For the first Layer receives the first The learnable raw weights are calculated for each input, and the summation term in the denominator represents the normalization process for all input weights of that layer. This structure allows small intersections of 15×15 pixels to be... Layer acquisition Layer context information, and the large intersection of 50×50 pixels is in Layer acquisition Layer detail information significantly improves the average accuracy (mAP) of multi-scale detection.

[0053] The detection head employs a decoupled structure, with the classification and regression branches operating independently. The regression branch outputs the offset of the intersection point's center coordinates. , The width and height are determined by 1 pixel, and the classification branch outputs three types of confidence scores.

[0054] The improved YOLOv10 model is obtained by using the above method.

[0055] Step S202: Train the improved YOLOv10 model based on the pre-acquired training data to obtain the intersection recognition model.

[0056] In this embodiment, training data is first constructed, covering images of various specifications of steel cages with pile diameters of 600mm-1500mm, main reinforcement diameters of 15mm-28mm, and stirrup spacing of 100mm-200mm. Image acquisition covers all lighting conditions in the processing plant: direct sunlight on sunny days (illuminance 80,000-120,000 lux), diffused light on cloudy days (5,000-10,000 lux), artificial lighting at night (500-2,000 lux), and instantaneous intense light from arc welding (peak value 150,000 lux). The annotation categories are divided into three types: orthogonal intersection points of main reinforcement and stirrups (intersection angles). ), the intersection point of the main reinforcement and the spiral reinforcement (intersection angle) ), and non-intersecting, easily confused points (such as rust spots on steel reinforcement surfaces or concrete drip marks). The annotation method is point coordinate marking, with each intersection point annotated with its geometric center sub-pixel coordinates. The accuracy was refined to 0.1 pixels after manual annotation and cubic spline interpolation. The dataset consisted of no fewer than 8000 images, divided into training, validation, and test sets in a 7:2:1 ratio to ensure model generalization ability. Training employed a two-stage strategy: freezing the backbone network for the first 50 rounds and then fine-tuning the entire network for 150 rounds after unfreezing. The optimizer used was AdamW with a weight decay coefficient of 0.05. Cosine annealing was applied after 10 warm-up rounds, with an initial learning rate of 0.001 and a final learning rate of 0.0001. Data augmentation included random brightness adjustment (±30%), random contrast adjustment (±20%), random rotation (±5°), and Mosaic stitching to simulate actual imaging changes in a processing area. After training, the model achieved a mean intersection point localization error of 0.3 pixels and a standard deviation of 0.15 pixels on the test set, meeting sub-pixel measurement requirements, thus obtaining the final intersection point recognition model.

[0057] Step S203: Input the single-frame image into the intersection recognition model to obtain the initial intersection and confidence level.

[0058] The single-frame image of the pile foundation reinforcement cage, after grayscale and Gaussian smoothing, is input into the training intersection recognition model. The intersection recognition model completes inference through multi-scale feature fusion and dynamic illumination compensation, and identifies three types of targets in the image: orthogonal intersection points of main reinforcement and stirrups, oblique intersection points, and non-intersection easily confused points. It outputs the sub-pixel coordinates of each initial intersection point and the corresponding confidence score (obtained by multiplying the classification score and the IoU score).

[0059] Step S204: Based on the confidence level, the initial intersection points are filtered to obtain valid intersection points.

[0060] In this embodiment, when filtering the initial intersection points based on confidence, the final confidence level obtained by multiplying the classification score and the IoU score is first used as the basis, and 0.85 is set as the filtering threshold to remove low-confidence intersection points with confidence levels below this threshold. Then, the Soft-NMS algorithm is used for post-processing, and an 8-pixel suppression radius is set to avoid the same intersection point being detected repeatedly, thus obtaining an accurate and non-redundant set of effective intersection points.

[0061] Step S3: Calculate the representative value of the stirrup pixel spacing, the proportionality coefficient, and the standard uncertainty based on the effective intersection points.

[0062] The calculation of the representative value of the stirrup pixel spacing, the scaling factor, and the standard uncertainty based on the effective intersection points specifically includes the following sub-steps: Step S301: Based on the random sampling consensus algorithm, perform line fitting on the effective intersection points to obtain the set of horizontal lines of the main reinforcement and the set of vertical lines of the stirrups.

[0063] In this embodiment, the set of coordinates of valid intersection points ,in To determine the effective number of intersection points, RANSAC (Random Sample Consensus Algorithm) linear fitting was performed. The main reinforcement linear model was set as a horizontal line with a slope tolerance of ±3°, 500 iterations, and an interior point threshold of 2. The stirrup linear model was set as a vertical line with a slope tolerance of... to After fitting, the set of main reinforcement horizontal lines is obtained. Each straight line is expressed as , The vertical coordinates in the image pixel coordinate system. For the first The slope of a straight line, The horizontal coordinates in the image pixel coordinate system. For the first The intercept of a straight line, where the absolute value of the slope is... For each main reinforcement line, those with fewer than 5 internal points are discarded to avoid interference from isolated points, ultimately resulting in the set of horizontal main reinforcement lines and the set of vertical stirrup lines.

[0064] Step S302: Determine the central main reinforcement and the upper and lower main reinforcement based on the set of main reinforcement horizontal lines.

[0065] In this embodiment, the vertical center line of a single frame image is used. ( Using the image height as a reference, calculate the average distance from each main reinforcement line to the centerline. : Indicates the first The total number of intersections on the main reinforcement bars (i.e., the number of main reinforcement-stirrup intersections detected on the straight line). Indicates the index of the valid intersection point on the line. For the first The vertical coordinates of the effective intersection points, then the main reinforcement is... Sort the reinforcement bars from smallest to largest, and select the three main reinforcement bars closest to the centerline, denoted as follows: : Top main reinforcement (v coordinate is smaller); Central main reinforcement (center of v coordinate); : Bottom main reinforcement (larger v coordinate), where the v coordinate is the longitudinal (vertical) coordinate in the image coordinate system, that is, the top and bottom main reinforcements include the top main reinforcement and the bottom main reinforcement.

[0066] This selection method does not rely on pixel coordinate cropping, but is achieved through geometric distance sorting, ensuring that the selected main ribs are always located in the optical center area of ​​the lens, while retaining the upper and lower ribs for redundancy verification.

[0067] Step S303: Based on the set of vertical lines of the central main reinforcement, the upper and lower main reinforcements and the stirrups, determine the intersection points of the main reinforcements, and calculate the Euclidean distance between adjacent intersection points of the main reinforcements to obtain the pixel spacing sequence of the central main reinforcement and the pixel spacing sequence of the upper and lower main reinforcements.

[0068] In this embodiment, based on the three horizontal lines selected—the central main reinforcement, the upper main reinforcement, and the lower main reinforcement—geometric intersection calculations are performed one by one with each vertical line in the set of stirrup vertical lines to accurately locate the sub-pixel level intersection points (i.e., main reinforcement intersection points) between each main reinforcement and the stirrups. Then, the main reinforcement intersection points on each main reinforcement are arranged according to the horizontal coordinates. The nodes are arranged in ascending order to ensure that the order of the intersection points is consistent with the actual distribution of the stirrups on the steel cage. Then, the spatial distance between two adjacent intersection points on each main bar is calculated using the Euclidean distance formula. Finally, the pixel spacing sequence of the central main bar and the corresponding pixel spacing sequence of the upper and lower main bars are obtained. The two together constitute the pixel spacing sequence of the upper and lower main bars.

[0069] Euclidean distance : Three sets of pixel spacing sequences were obtained: the pixel spacing sequences corresponding to each of the lower main ribs. Central main rib pixel spacing sequence Upper main rib pixel spacing sequence , This is a sequence index, representing the main reinforcement. The number of the valid intersection points. For the first The horizontal pixel coordinates of the valid intersection points For the first The horizontal pixel coordinates of the valid intersection points For the first Vertical pixel coordinates of the valid intersection points For the first Vertical pixel coordinates of the three valid intersection points As the central main reinforcement, The main tendon, It is the lower main tendon.

[0070] Step S304: Perform median filtering on the central main rib pixel spacing sequence and the upper and lower main rib pixel spacing sequence respectively to obtain the optimized central pixel spacing sequence and the optimized upper and lower pixel spacing sequence.

[0071] In this embodiment, median filtering is performed on the central main rib pixel spacing sequence and the upper and lower main rib pixel spacing sequence to remove outliers that deviate from the median by more than 1.5 times the interquartile range. These outliers are mostly caused by local occlusion or arc reflection, resulting in an optimized central pixel spacing sequence and an optimized upper and lower pixel spacing sequence.

[0072] Step S305: Perform a weighted average of the optimized central pixel spacing sequence and the optimized upper and lower pixel spacing sequence to obtain the representative value of the stirrup pixel spacing.

[0073] In this embodiment, the central main reinforcement Being closest to the optical axis results in minimal imaging distortion and optimal depth of field, thus it is given weight. Main reinforcement bars (top and bottom) and It provides spatial redundancy, but the image quality is slightly inferior; each is assigned a weight. This weighting is based on an imaging physics model; the closer to the optical axis, the greater the projection distortion. The smaller it is, the more likely it is to approximately satisfy the condition. ,in, This refers to the distance to the optical axis. The central main reinforcement... Typically, it's 1 / 3 of the weight of the edge main reinforcement, hence the weight is doubled. Ultimately... This serves as the sole representative value for pixel spacing in the image, ensuring that all subsequent calculations are based on the optimal subset of data.

[0074] The pixel spacing of the stirrups represents the value that satisfies the following formula: in, This represents the pixel spacing of the stirrups. For median operations, To optimize the pixel spacing sequence, To optimize the central pixel spacing sequence, To optimize the pixel spacing sequence.

[0075] Step S306: Obtain the prior distribution, and calculate the scaling factor and standard uncertainty based on the prior distribution and the optimized central pixel spacing sequence.

[0076] In this embodiment, the prior distribution includes the circular cross-section prior, the equidistant main reinforcement prior, and the equidistant stirrup prior.

[0077] Circular cross-section prior: The cross-section of the steel bar is an ideal circle, but the edge of the image is an ellipse, and the major axis of the ellipse is the pixel diameter. The national standard specifies a diameter tolerance of ±0.3mm for reinforcing bars; therefore, the actual diameter... Obey the nominal value Centered, standard deviation Cut-off normal distribution: .

[0078] in, Let be the probability density function of the true diameter of the reinforcing bar. To truncate the normal distribution (domain is ).

[0079] Priority for equidistant main reinforcement: Design spacing of main reinforcement Although the exact distance is unknown, the acceptance standard allows a deviation of ±10mm, therefore the actual spacing between adjacent main reinforcement bars is... Obey Centered, standard deviation Cut-off normal distribution: .

[0080] Priority for equidistant stirrups: design spacing of stirrups The allowable deviation is ±20mm, therefore the actual spacing between adjacent stirrups Obey Centered, standard deviation Cut-off normal distribution: .

[0081] The role of the aforementioned prior knowledge is to extract distribution parameters from the acceptance specifications rather than relying on specific design values, so that the prior knowledge is supported by legally mandated measurement documents and conforms to the physical reality of the "allowable deviation" at the construction site.

[0082] The calculation of the scaling factor and standard uncertainty based on the prior distribution and the optimized central pixel spacing sequence specifically includes the following sub-steps: Step S30601: Obtain the pixel diameter sample of the main rib based on the central main rib.

[0083] In this embodiment, in the central main reinforcement Each of the left and right ends is selected as a 128×128 pixel Region of Interest (ROI). The center of the ROI is offset by 30 pixels along the main rib direction to avoid local distortion at the intersection. Subpixel edge detection is performed on each ROI using Zernike moments, and ellipse fitting is performed based on the circular cross-section prior to obtain a set of pixel diameter samples. ,in The Zernike moment is set to order 4, achieving an edge positioning accuracy of 0.2 pixels.

[0084] Step S30602: Obtain the main reinforcement spacing sample based on the set of vertical lines of the stirrups and the upper and lower main reinforcement bars.

[0085] In this embodiment, the spacing samples of the main reinforcement bars are obtained based on the set of vertical lines of the stirrups and the upper and lower main reinforcement bars. The sampling range is limited to the effective grid area formed by the three central main reinforcement bars and the vertical lines of the stirrups to ensure that the data is from the same source as the calibration scene. Then, multiple vertical lines are selected from the set of vertical lines of the stirrups, and the vertical distance between the two intersection points of each vertical line and the upper and lower main reinforcement bars (i.e., the pixel value of the spacing of the main reinforcement bars) is calculated to form the spacing samples of the main reinforcement bars. , .

[0086] Step S30603: Obtain stirrup spacing samples based on the optimized central pixel spacing sequence.

[0087] In this embodiment, the optimized central pixel spacing sequence (i.e., the central main rib) after median filtering is directly extracted. corresponding The sequence is used as a sample of stirrup spacing. This sequence has eliminated outliers caused by local occlusion and arc reflection, and naturally satisfies the prior of equidistant stirrups.

[0088] Step S30604: Using the prior distribution as a constraint, the expectation-maximization algorithm is used to jointly estimate the main reinforcement pixel diameter sample, the main reinforcement spacing sample, and the stirrup spacing sample to obtain the proportionality coefficient and standard uncertainty.

[0089] Specifically, using the prior distribution as a constraint, the expectation-maximization algorithm is employed to jointly estimate the proportionality coefficient and standard uncertainty of the main reinforcement pixel diameter samples, the main reinforcement spacing samples, and the stirrup spacing samples, including the following sub-steps: Step S3060401: Calculate the median diameter based on the main rib pixel diameter sample.

[0090] In this embodiment, all main rib pixel diameter samples are sorted in ascending order of numerical value. If the number of samples is odd, the value in the middle position after sorting is selected as the median diameter. If the number of samples is even, the arithmetic mean of the two middle values ​​after sorting is taken as the median diameter. This calculation method can effectively suppress outlier interference caused by arc reflection, local occlusion or edge detection errors in individual ROIs.

[0091] Step S3060402: Obtain the national standard diameter library, and determine the initialization ratio coefficient based on the median diameter and the national standard diameter library.

[0092] In this embodiment, the national standard diameter library is a dataset that conforms to national standards and includes commonly used nominal steel bar diameter specifications such as 12mm, 14mm, and 16mm. This dataset is used to match the median pixel diameter of the main reinforcement bars to initialize the scaling factor. initial value (Initialized scaling factor) is roughly estimated from the diameter channel: for the main rib pixel diameter sample Take the median Assuming it corresponds to the closest nominal value in the national standard diameter database. ,but The initial value error is usually less than 15%, ensuring that the EM algorithm converges quickly.

[0093] Step S3060403: Based on the initial scaling factor, the main reinforcement pixel diameter sample, the main reinforcement spacing sample, and the stirrup spacing sample, obtain the actual size sample of the main reinforcement pixel diameter, the actual size sample of the main reinforcement spacing, and the actual size sample of the stirrup spacing.

[0094] In this embodiment, for each data point in the main rib pixel diameter sample, the true size is obtained by multiplying the initial scaling factor and the pixel diameter value, thus obtaining the true size sample of the main rib pixel diameter. Similarly, by multiplying each pixel spacing value in the main reinforcement spacing sample by the initial scaling factor, a sample of the true size of the main reinforcement spacing is generated. For the stirrup spacing samples, the initial scaling factor was used as the conversion benchmark to calculate the actual stirrup spacing samples point by point. The three types of real-size samples correspond one-to-one with the original pixel samples, fully preserving the statistical distribution characteristics of each sample. This lays the data foundation for constructing a three-channel joint likelihood function with prior distribution as a constraint and realizing the iterative optimization of the EM algorithm.

[0095] Step S3060404: Using the prior distribution as a constraint, construct a three-channel joint likelihood function using the actual size samples of the main reinforcement pixel diameter, the actual size samples of the main reinforcement spacing, and the actual size samples of the stirrup spacing.

[0096] In this embodiment, when constructing the three-channel joint likelihood function with three prior distributions as constraints, the corresponding sub-channel likelihood terms are first calculated for the three types of real-size samples: for the real-size samples of the main rib pixel diameter Based on the prior knowledge of circular cross-sections, a one-dimensional registration is performed with the national standard diameter database, and the diameter channel likelihood term is calculated using the Gaussian likelihood function. The samples are constrained to converge towards the nominal diameter of the national standard; for samples with the actual size of the main reinforcement spacing... Based on the prior of equidistant main reinforcement bars, their histograms are matched with a ±10mm truncated normal distribution, and the channel likelihood term of the main reinforcement bar spacing is obtained by cross-entropy likelihood calculation. Ensure that the fluctuation range meets the acceptance specifications; for the actual size samples of stirrup spacing Based on the prior knowledge of equidistant stirrups, the histogram of the stirrups is matched with the convolution of a ±20mm truncated normal distribution, and the channel likelihood term of the stirrup spacing is obtained by cross-entropy likelihood calculation. Then, based on the effective number of samples in the three categories, a normalized weighted average is calculated, and finally, the three weighted averages are summed to obtain a unified three-channel joint likelihood function.

[0097] Diameter channel likelihood: This involves sampling the actual diameter of the main rib pixels. One-dimensional registration was performed with the national standard diameter library D={12,14,16,18,20,22,25,28}mm, and the Gaussian likelihood was calculated. : in, The index of the actual size sample of the main rib pixel diameter; The total number of samples for the diameter of the main reinforcement bars used in the calculation; Indicates from the national standard diameter database Iterate through each nominal diameter; For the first During the nth iteration, the 1st The actual size values ​​of each sample after conversion using a scaling factor; This corresponds to the center value of the nominal diameter; This represents the standard deviation of the diameter distribution; Let represent the probability density function of the truncated normal distribution. This likelihood term utilizes the prior of the circular cross-section, making... Converge towards the closest national standard diameter to avoid random wandering.

[0098] Stirrup spacing channel likelihood: This involves sampling the actual dimensions of the stirrup spacing. histogram and Perform convolution matching on a normal distribution truncated by ±20mm, and calculate the histogram cross-entropy: in, For continuous physical spacing variables, used for interval integration of the probability density function; This is the index of the center position of the histogram bins. Total number of blocks; If the width of the histogram block is half, then Indicates the first The statistical interval of each block; For the t-th iteration, the true size sample falls within the t-th iteration. Normalized frequency (or count) within each block; and These are the a priori central value and variance of the stirrup design spacing, respectively; Let be an integral infinitesimal element, representing that in An infinitesimal increment in the direction. This likelihood term utilizes the equidistant prior of the stirrups to make... The statistical distribution conforms to the fluctuation range allowed by the acceptance specifications.

[0099] Main reinforcement spacing channel likelihood: Similarly, the actual size sample of the main reinforcement spacing... histogram and ±10 mm cutoff normal distribution matching: Parallel main reinforcement prior to pass The distribution should be narrower than This relative constraint is implicitly realized, namely < This prevents the algorithm from misinterpreting the spacing of the main reinforcement bars as the spacing of the stirrups.

[0100] Step S3060405: The updated scaling factor is obtained by maximizing the joint likelihood function of the three channels.

[0101] The update ratio coefficient satisfies the following formula: in, , , These are the normalized weights for the sample counts of the main reinforcement pixel diameter samples, the stirrup spacing samples, and the main reinforcement spacing samples, respectively. For the number of iterations, the three-channel joint optimization makes It is estimated to have the ability to resist outliers.

[0102] The normalized weights for the sample size are the ratios of the three samples to the total sample.

[0103] Step S3060406: Iterate based on the updated scaling factor until the pre-set convergence condition is met to obtain the scaling factor.

[0104] In this embodiment, the iteration termination condition is: or After convergence, the proportionality constant is obtained. .

[0105] Step S3060407: Calculate the standard uncertainty based on the proportionality coefficient and the three-channel joint likelihood function.

[0106] Standard uncertainty Approximate calculation using the Fisher information matrix: in These are the summation indices for the diameter channel, stirrup channel, and main reinforcement channel samples, respectively. These represent the total number of valid samples participating in the calculation across the three channels. This uncertainty simultaneously quantifies the confidence level of the geometric prior, rather than considering only pixel noise.

[0107] Step S4: Calculate the true spacing and expanded uncertainty based on the proportionality coefficient, the standard uncertainty, and the representative value of the stirrup pixel spacing.

[0108] The calculation of the true spacing and expanded uncertainty based on the proportionality coefficient, the standard uncertainty, and the representative value of the stirrup pixel spacing specifically includes the following sub-steps: Step S401: Multiply the proportional coefficient by the representative value of the stirrup pixel spacing to obtain the actual spacing.

[0109] In this embodiment, the formula is used Mapping statistical values ​​from pixel space back to physical space. This is the proportionality coefficient. This represents the pixel spacing of the stirrups. This represents the actual spacing.

[0110] Step S402: Calculate the expanded uncertainty based on the standard uncertainty and the representative value of the stirrup pixel spacing.

[0111] In this embodiment, the expanded uncertainty is calculated using a coverage factor of 2 at a 95% confidence level. ,in, For standard uncertainty, This represents the pixel spacing of the stirrups. This is the expanded uncertainty. This expanded uncertainty includes not only the random error from the pixel detection stage, but also... It integrates the model uncertainty caused by the inconsistency between the prior distribution and the observed data.

[0112] Step S5: Generate the evaluation result of the spacing of the pile foundation reinforcement cage stirrups based on the actual spacing and the expanded uncertainty.

[0113] The evaluation result of generating the stirrup spacing of the pile foundation reinforcement cage based on the actual spacing and the expanded uncertainty specifically includes the following sub-steps: Step S501: Obtain the design spacing and tolerance threshold of the stirrups in the pile foundation reinforcement cage.

[0114] In this embodiment, the design spacing is obtained by automatically retrieving the design parameters of the corresponding steel cage model from the preset engineering parameter library. The tolerance threshold is set to ±20mm by default according to the "Code for Acceptance of Construction Quality of Concrete Structures". At the same time, it supports custom adjustment according to the special acceptance requirements of the project or the special provisions of the design documents to ensure that the threshold meets the statutory acceptance standards and actual construction needs.

[0115] Step S502: When it is determined that the expanded uncertainty is less than the tolerance threshold, the corrected tolerance threshold is obtained by subtracting the expanded uncertainty from the tolerance threshold.

[0116] In this embodiment, when the expanded uncertainty is less than the tolerance threshold, in order to eliminate the interference of measurement uncertainty on the conformity determination and ensure that the conclusion has legal metrological traceability, the corrected tolerance threshold is calculated by subtracting the expanded uncertainty from the tolerance threshold. If the expanded uncertainty is greater than or equal to the tolerance threshold, the system directly outputs a prompt that the uncertainty is too large and the image needs to be re-captured for detection, so as to avoid misjudgment due to insufficient measurement accuracy.

[0117] Step S503: Calculate the absolute difference between the actual spacing and the designed spacing, and compare the absolute difference with the corrected tolerance threshold to obtain the spacing evaluation result of the pile foundation reinforcement cage stirrups.

[0118] In this embodiment, the absolute difference between the actual spacing and the designed spacing is calculated. This absolute difference is then directly compared to a corrected tolerance threshold. If the absolute difference is less than or equal to the corrected tolerance threshold, the spacing of the pile foundation reinforcement cage stirrups is determined to meet the requirements of the "Code for Acceptance of Construction Quality of Concrete Structures," and the evaluation result is "qualified." If the absolute difference is greater than the corrected tolerance threshold, the evaluation result is "unqualified." Simultaneously, the spacing of the central main reinforcement can be adjusted. Each paragraph Calculate residuals : in, This is the proportionality coefficient. To optimize the central pixel spacing sequence, This is the a priori center value for the design spacing of the stirrups. If , To expand the uncertainty, the section is identified as a construction anomaly (such as missing or deformed stirrups), and its specific location is highlighted in red in the report to facilitate rework by workers.

[0119] like Figure 2 As shown, in another aspect, the present invention also provides a referenceless visual evaluation system for the spacing of stirrups in a pile foundation reinforcement cage, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to call the program instructions to execute the relevant steps of the relevant embodiments of the referenceless visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage of the present invention.

[0120] This invention provides a referenceless visual evaluation system for the spacing of stirrups in a pile foundation reinforcement cage. The functional components can be integrated into a single processing unit, exist as separate physical entities, or be integrated into a single unit. These integrated components can be implemented in hardware or software.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A referenceless visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage, characterized in that, The method includes: Acquire a single frame image of the pile foundation reinforcement cage; Construct an intersection recognition model and use the intersection recognition model to identify the valid intersections of the main reinforcement and stirrups in the single frame image; Based on the effective intersection points, the representative value of the stirrup pixel spacing, the proportionality coefficient, and the standard uncertainty are calculated, including: Based on the random sampling consensus algorithm, the effective intersection points are fitted with straight lines to obtain the set of horizontal lines for main reinforcement and the set of vertical lines for stirrups; The central main reinforcement and the upper and lower main reinforcements are determined based on the set of horizontal lines of the main reinforcement. The intersection points of the main reinforcement bars are determined based on the set of vertical lines of the central main reinforcement bars, the upper and lower main reinforcement bars, and the stirrups. The Euclidean distance between the intersection points of the adjacent main reinforcement bars is calculated to obtain the pixel spacing sequence of the central main reinforcement bars and the pixel spacing sequence of the upper and lower main reinforcement bars. Median filtering is performed on the central main rib pixel spacing sequence and the upper and lower main rib pixel spacing sequence to obtain the optimized central pixel spacing sequence and the optimized upper and lower pixel spacing sequence. The optimized central pixel spacing sequence and the optimized upper and lower pixel spacing sequence are weighted and averaged to obtain the representative value of the stirrup pixel spacing; Obtain the prior distribution, which includes the prior distribution of circular cross-section, the prior distribution of main reinforcement equidistant bars, and the prior distribution of stirrups equidistant bars. Calculate the scaling factor and standard uncertainty based on the prior distribution and the optimized central pixel spacing sequence, including: Based on the central main rib, obtain the pixel diameter sample of the main rib; Based on the set of vertical lines of the stirrups and the upper and lower main bars, a sample of the spacing between the main bars is obtained; Based on the optimized central pixel spacing sequence, stirrup spacing samples are obtained; Using the prior distribution as a constraint, the expectation-maximization algorithm is employed to jointly estimate the proportionality coefficient and standard uncertainty of the main reinforcement pixel diameter samples, the main reinforcement spacing samples, and the stirrup spacing samples, including: Calculate the median diameter based on the main rib pixel diameter sample; Obtain the national standard diameter database, and determine the initial scaling factor based on the median diameter and the national standard diameter database; Based on the initial scaling factor, the main reinforcement pixel diameter sample, the main reinforcement spacing sample, and the stirrup spacing sample, the actual size samples of the main reinforcement pixel diameter, the main reinforcement spacing, and the stirrup spacing are obtained. Using the prior distribution as a constraint, a three-channel joint likelihood function is constructed using the actual size samples of the main reinforcement pixel diameter, the actual size samples of the main reinforcement spacing, and the actual size samples of the stirrup spacing. The updated scaling factor is obtained by maximizing the joint likelihood function of the three channels; The iteration is performed based on the updated scaling factor until the pre-set convergence condition is met, and the scaling factor is obtained. The standard uncertainty is calculated based on the proportionality coefficient and the three-channel joint likelihood function. The true spacing and expanded uncertainty are calculated based on the proportionality coefficient, the standard uncertainty, and the representative value of the stirrup pixel spacing. The evaluation results of the pile foundation reinforcement cage stirrup spacing are generated based on the actual spacing and the expanded uncertainty.

2. The non-reference visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage according to claim 1, characterized in that, The acquisition of a single-frame image of the pile foundation reinforcement cage includes: Obtain the original single-frame image and lens distortion parameters of the pile foundation reinforcement cage; The original single-frame image is distorted using the lens distortion parameters to obtain a corrected single-frame image; The single-frame image of the pile foundation reinforcement cage is obtained by performing grayscale conversion and Gaussian smoothing on the corrected single-frame image.

3. The non-reference visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage according to claim 1, characterized in that, The construction of the intersection recognition model and the use of the intersection recognition model to identify the valid intersections of the main reinforcement and stirrups in the single frame image include: The YOLOv10 model is introduced, and the YOLOv10 model is improved by using a pre-set deformable convolution module and a light-sensing module to obtain an improved YOLOv10 model; The improved YOLOv10 model is trained based on pre-acquired training data to obtain an intersection recognition model; The single-frame image is input into the intersection recognition model to obtain the initial intersection points and confidence levels; The initial intersection points are filtered based on the confidence level to obtain valid intersection points.

4. The non-reference visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage according to claim 1, characterized in that, The pixel spacing of the stirrups represents the value that satisfies the following formula: , in, This represents the pixel spacing of the stirrups. For median operations, To optimize the pixel spacing sequence, To optimize the central pixel spacing sequence, To optimize the pixel spacing sequence.

5. The non-reference visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage according to claim 1, characterized in that, The calculation of the true spacing and expanded uncertainty based on the scaling factor, the standard uncertainty, and the representative value of the stirrup pixel spacing includes: The actual spacing is obtained by multiplying the proportional coefficient by the representative value of the stirrup pixel spacing; The expanded uncertainty is calculated based on the standard uncertainty and the representative value of the stirrup pixel spacing.

6. The non-reference visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage according to claim 1, characterized in that, The evaluation results of generating the pile foundation reinforcement cage stirrup spacing based on the actual spacing and the expanded uncertainty include: Obtain the design spacing and tolerance threshold of the stirrups in the pile foundation reinforcement cage; When the expanded uncertainty is determined to be less than the tolerance threshold, the corrected tolerance threshold is obtained by subtracting the expanded uncertainty from the tolerance threshold. Calculate the absolute difference between the actual spacing and the designed spacing, and compare the absolute difference with the corrected tolerance threshold to obtain the spacing evaluation result of the pile foundation reinforcement cage stirrups.

7. A referenceless visual evaluation system for the spacing of stirrups in a pile foundation reinforcement cage, characterized in that, include: The system includes a processor, an input device, an output device, and a memory, all interconnected, wherein the memory stores a computer program comprising program instructions, and the processor is configured to invoke the program instructions to execute a non-reference visual evaluation method for the spacing of stirrups in a pile foundation reinforcement cage as described in any one of claims 1 to 6.

Images