Grout masonry quality detection system based on machine vision

Abstract

The invention belongs to the technical field of computer vision and image processing, particularly relates to a grouted rubble quality detection system based on machine vision, and aims to solve the problems that manual detection is low in efficiency and high in subjectivity, and internal defects are difficult to recognize. The system comprises an image acquisition unit, a three-dimensional reconstruction unit, a defect analysis unit and a decision output unit. A high-precision three-dimensional point cloud is generated through a multi-view high-definition image, and automatic identification and quantitative evaluation of defects such as cracks, cavities and dislocation are realized in combination with an improved PointNet model and a rule engine. And the decision-making unit executes grading judgment according to defect types and combinations, generates a structured report containing a thermodynamic diagram, a parameter table and repair suggestions, and supports docking with a quality management system. The detection precision of the system reaches 98%, 1200 pieces of grouted rubble can be processed per hour, and the detection efficiency and the judgment objectivity are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, and specifically relates to a machine vision-based quality inspection system for masonry. Background Technology

[0002] Computer vision technology, as an important branch of artificial intelligence, plays an increasingly crucial role in the field of industrial automation inspection. It achieves objective assessment of product quality through image acquisition, processing, and analysis, and is widely used in manufacturing, construction, and other scenarios. It effectively improves inspection efficiency and accuracy, and provides important support for ensuring engineering safety and quality control.

[0003] Among them, the machine vision-based masonry quality inspection system focuses on solving the problem of defect identification in masonry structure construction. Its core principle is to use a high-resolution camera to acquire images of the masonry surface and combine them with image processing algorithms to automatically analyze potential quality hazards such as cracks, voids and uneven strength, so as to replace traditional manual visual inspection and achieve standardization and efficiency in the inspection process.

[0004] Existing technologies face significant challenges in the quality inspection of masonry. Changes in ambient lighting lead to insufficient stability in image acquisition, complex surface textures of masonry interfere with the accuracy of feature extraction, and real-time processing capabilities are limited by algorithm efficiency, making it difficult to meet the needs of rapid on-site inspection. Furthermore, existing systems have poor robustness in identifying multi-scale defects. These problems can easily lead to safety hazards in building quality control, necessitating a highly adaptable and reliable inspection solution. Summary of the Invention

[0005] The purpose of this invention is to provide a machine vision-based quality inspection system for masonry, in order to solve the problems of low efficiency, strong subjectivity, difficulty in accurately identifying internal defects, and inability to simultaneously quantify multi-dimensional quality parameters in existing technologies.

[0006] The technical solution of this invention includes an image acquisition unit, a 3D reconstruction unit, a defect analysis unit, and a decision output unit. The image acquisition unit is used to acquire multi-view high-definition image data of the masonry structure surface; the 3D reconstruction unit is used to generate a high-precision 3D point cloud model based on the multi-view images; the defect analysis unit is used to extract and quantitatively evaluate defect features from the 3D point cloud model; and the decision output unit is used to generate a structured quality assessment report based on the quantitative evaluation results.

[0007] Furthermore, the image acquisition unit consists of a ring array of six industrial-grade CMOS cameras, evenly distributed around the masonry inspection area. Each camera's optical sensor resolution is set to 2048×2048 pixels, with a fixed frame rate of 30 frames per second. The camera array employs a synchronous triggering mechanism, establishing a real-time data transmission channel with the central processing server via a gigabit Ethernet interface to ensure temporal consistency and spatial alignment accuracy of multi-view images. The light source system uses a ring-shaped LED array with a stable color temperature of 5600K and illuminance uniformity error controlled within 5%, effectively eliminating surface reflections and shadow interference. The output of the image acquisition unit is directly connected to the input of the 3D reconstruction unit, transmitting data including the original image sequence and corresponding camera pose parameters.

[0008] In one embodiment of the present invention, the 3D reconstruction unit employs a motion recovery structure-based algorithm to process multi-view image sequences. This algorithm first extracts key points from the images using a scale-invariant feature transform operator, then uses a random sampling consensus algorithm to eliminate mismatched point pairs, establishing precise geometric constraints between images. Subsequently, a bundle adjustment optimization process is performed, simultaneously solving for the camera extrinsic matrix and the 3D point cloud coordinates to generate a dense point cloud model with a density of no less than 10,000 points per square meter. The coordinate system of the point cloud model has its origin at the geometric center of the masonry, with the X-axis along the length direction, the Y-axis along the width direction, and the Z-axis along the height direction, using millimeters as the coordinate unit. The 3D reconstruction unit incorporates a point cloud filtering module, employing a statistical outlier removal algorithm to eliminate noise points, achieving a point cloud smoothness error of less than 0.1 millimeters, ensuring the fidelity of the model's geometric features.

[0009] The defect analysis unit comprises a dual-core processing module: a deep learning model and a rule engine. The deep learning model employs an improved PointNet architecture. The input layer receives 3D point cloud data, which is then processed through a 5-level shared multilayer perceptron layer for feature extraction. The number of neurons in each layer is 64, 128, 256, 512, and 1024, respectively. The feature aggregation layer captures global contextual information through max pooling. The output layer generates identification results for three types of defects: cracks, cavities, and misalignments, including defect type labels, spatial coordinates, width, length, and depth quantization parameters. The rule engine module pre-configures a quality threshold matrix based on the water conservancy industry standard SL274-2020, classifying the quantization parameters output by the deep learning model: cracks wider than 2 mm and longer than 50 mm are marked as severe defects; cavities larger than 1000 cubic millimeters are classified as structural hazards; and misalignments exceeding 15 mm are categorized as construction deviations. The defect analysis unit operates on a serial structure, with the output of the deep learning model connected to the input of the rule engine, ensuring data transmission latency of less than 100 milliseconds.

[0010] The decision output unit receives defect analysis results and performs a comprehensive quality level assessment. This unit has a built-in dynamic threshold matrix that sets three quality standards based on defect type combinations: when more than three severe defects are detected, the system automatically classifies it as unacceptable; if one or two severe defects are present, it is marked as requiring repair; otherwise, it is classified as acceptable. The quality assessment report includes a defect distribution heatmap, a table of quantitative parameters, and repair recommendations, with report generation time controlled within 5 seconds. The decision output unit interfaces with the quality management system via a standard OPC UA protocol interface, supporting real-time data upload and historical record traceability. Data transmission encryption uses the AES-256 algorithm.

[0011] Furthermore, the camera array layout of the image acquisition unit follows the principle of minimum coverage angle, with the overlap rate of adjacent camera field of view set at 30% to ensure no blind spots on the masonry surface. The camera's optical lens focal length is fixed at 12 mm, and the aperture value is adjusted to f / 2.8, covering a depth of field range of 0.5 meters to 3 meters for detection distances. The image preprocessing module performs white balance correction and gamma correction in real time at the acquisition end, expanding the dynamic range to 12 bits and improving the visibility of details in low-contrast areas. The beam adjustment optimization process of the 3D reconstruction unit introduces a robust kernel function to reduce the impact of outliers on pose solving, controlling the reprojection error within 0.5 pixels. After the point cloud model is generated, coordinate normalization is automatically performed, scaling the size to a standard reference frame to eliminate evaluation bias caused by scale differences.

[0012] In one embodiment of the present invention, the deep learning model training of the defect analysis unit employs a transfer learning strategy. The base model is pre-trained on the publicly available dataset ShapeNet, and then fine-tuned using 10,000 labeled masonry defect samples. The training samples cover different lighting conditions, surface contamination, and structural morphologies. Data augmentation strategies include random rotation, translation, and Gaussian noise injection. The loss function is designed as a combination of weighted cross-entropy and IoU loss, with weight coefficients set to 0.7 and 0.3, respectively. The optimization objective focuses on improving the recognition accuracy of small-sized defects. During the model inference stage, non-maximum suppression is performed, and the overlap region suppression threshold is set to 0.3 to ensure accurate defect boundary localization. The threshold matrix of the rule engine supports a dynamic update mechanism, iteratively optimizing parameters using actual engineering data fed back from the quality management system, with an update cycle set to 7 days.

[0013] The quality judgment logic of the decision output unit adopts a multi-condition priority strategy. When cracks and voids exist simultaneously, the system prioritizes evaluating the volume ratio of voids. If the void volume ratio exceeds 5% of the total volume of the masonry, it is directly judged as unqualified, ignoring other defect parameters. The repair suggestion text generation module, based on defect type and location information, calls a pre-set knowledge base to output standardized treatment solutions: for cracks requiring repair, epoxy resin grouting is recommended, with a grouting pressure set at 0.5 MPa; for structural voids, loose materials should be removed before filling with C30 fine aggregate concrete. The report template conforms to the GB / T1.1-2020 standard format and includes a testing timestamp, equipment serial number, and digital signature to ensure the legal validity and traceability of the report.

[0014] Furthermore, the system integrates an environmental sensing submodule to monitor temperature, humidity, and vibration parameters at the testing site in real time. The temperature sensor has an accuracy of ±0.5 degrees Celsius, the humidity sensor covers a range of 30% to 90% relative humidity, and the vibration acceleration meter has a range of 0 to 10 meters per second squared. When environmental parameters exceed preset thresholds, the system automatically triggers a compensation mechanism: when the temperature change rate is greater than 2 degrees Celsius per minute, the image color correction algorithm is activated; when the vibration amplitude exceeds 0.2 meters per second squared, the camera exposure time is extended to 1 / 60 of a second. Environmental data and test results are stored synchronously to form a complete quality assessment context. The central processing server is equipped with dual Intel Xeon Silver4310 processors with a clock speed of 2.1 GHz, 64 gigabytes of memory, and an NVIDIA Tesla T4 GPU with 16 gigabytes of video memory, ensuring the processing of 1200 masonry stone inspection tasks per hour.

[0015] In one embodiment of the present invention, a data closed-loop verification mechanism is established between the 3D reconstruction unit and the defect analysis unit. The system randomly selects 5% of the detection samples, projects the 3D point cloud model onto the original image plane, and calculates the matching error between the projected points and image feature points. When the average error exceeds 0.8 pixels, the reconstruction parameter calibration process is automatically triggered, adjusting the convergence threshold of the bundle adjustment and the point cloud filtering intensity. The defect analysis unit periodically performs model health diagnosis, analyzing the false positive rate and false negative rate using a confusion matrix. When the false negative rate rises above 3%, the system prompts that additional training samples are needed. The data storage module adopts a distributed architecture, retaining raw detection data for 30 days, permanently archiving structured reports, and setting the automatic storage space cleanup strategy to a capacity threshold of 90%.

[0016] Compared with existing technologies, the advantages and positive effects of this invention are as follows: This system achieves full-domain detection of surface and internal defects in masonry through the deep coupling of multi-view image acquisition and 3D reconstruction technologies, with a detection accuracy of 98%, which is 40 percentage points higher than the manual tapping detection method; the collaborative processing mechanism of deep learning models and rule engines reduces the identification speed of key defects such as cracks and voids to 5 seconds per piece, increasing detection efficiency by 8 times compared to traditional photogrammetry; the application of dynamic threshold matrices and multi-condition priority judgment logic ensures that the quality grade judgment results match the actual engineering situation at a rate of 99.2%, completely eliminating subjective bias caused by human experience; the integrated environmental perception submodule effectively compensates for on-site interference factors, ensuring the stability of detection results under complex working conditions; the structured reports generated by the system directly connect to the quality management system, supporting statistical analysis and trend prediction of defect data, providing quantitative decision-making basis for engineering quality control, and significantly reducing the safety risks caused by masonry quality problems in water conservancy projects. Attached Figure Description

[0017] Figure 1 This is a flowchart of the overall technical solution of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the collaborative processing mechanism between the deep learning model and the rule engine in this invention. Detailed Implementation

[0018] Please refer to Figure 1 and Figure 2 This invention provides a machine vision-based quality inspection system for masonry, aiming to solve the problems of low efficiency, strong subjectivity, difficulty in accurately identifying internal defects, and inability to synchronously quantify multi-dimensional quality parameters in manual inspection. The system achieves full-domain detection of surface and internal defects of masonry through deep coupling of multi-view image acquisition and 3D reconstruction technology. In this embodiment, the system is deployed at a quality inspection workstation in a water conservancy project site, and consists of an image acquisition unit, a 3D reconstruction unit, a defect analysis unit, and a decision output unit forming a closed-loop processing link. The overall technical process of the system begins with the image acquisition unit capturing multi-view high-definition image data of the masonry structure surface, which is then transmitted to the 3D reconstruction unit to generate a high-precision 3D point cloud model. The defect analysis unit performs defect feature extraction and quantitative evaluation based on the point cloud model, and finally, the decision output unit generates a structured quality judgment report. Each unit strictly follows the principle of unidirectional data flow, with the output of the front-end unit directly connected to the input of the back-end unit, ensuring the real-time performance and data consistency of the inspection process.

[0019] The core function of the image acquisition unit is to provide high-fidelity visual input for subsequent processing; the unit employs six industrial-grade CMOS cameras arranged in a circular array, each camera using a Sony optical sensor. The IMX253 camera system features an effective pixel size of 5.86 μm × 5.86 μm, a precisely set resolution of 2048 × 2048 pixels, and a dynamic range of 74 dB. The camera array is physically installed around the masonry inspection area, with an array diameter of 3 meters. The overlap of the field of view between adjacent cameras is strictly controlled to 30%, ensuring no blind spots on the masonry surface. Each camera has a fixed 12 mm focal length and a precisely adjusted aperture of f / 2.8, covering a depth of field from 0.5 meters to 3 meters, with a minimum working distance of 0.8 meters. The camera synchronization triggering mechanism uses the IEEE 1588 precision time protocol. The trigger signal is distributed by the central processing server via a gigabit Ethernet interface, with a signal transmission delay of less than 100 nanoseconds, ensuring that the image acquisition time difference between the six cameras is less than 0.1 milliseconds. The data transmission channel uses UDP protocol encapsulation, with the data packet header containing a timestamp, camera ID, and checksum, and the payload being the original image sequence. The system includes corresponding camera pose parameters, described by pre-calibrated rotation matrices and translation vectors. The calibration process utilizes the Zhang Zhengyou calibration method, achieving a corner detection accuracy of 0.05 pixels. The light source system features a ring array of 64 high-power LEDs, maintaining a stable color temperature of 5600K. Illumination uniformity error is controlled within 5% by a feedback control system. The LED driver circuit employs a constant current source design, with current fluctuations less than 0.5%, effectively eliminating surface reflections and shadow interference. The image preprocessing module performs real-time white balance correction at the acquisition end. The algorithm is based on the gray-scale world assumption, dynamically calculating the gain coefficients of the red, green, and blue channels, with the gain range limited to 0.8 to 1.2 to avoid over-correction. The gamma correction parameter is fixed at 2.2, improving detail visibility in low-contrast areas. The dynamic range is extended to 12 bits, converting the 8-bit original image into a 12-bit high dynamic range image through a non-linear mapping function, with the mapping function slope parameter set to 0.5. Ensures enhanced detail in dark areas while preventing overexposure in highlight areas; the unit has a built-in data integrity verification mechanism, using a 32-bit cyclic redundancy check algorithm to verify image data. When verification fails, a retransmission request is automatically triggered, with a retransmission interval of 10 milliseconds and a maximum of 3 retransmissions; anomaly handling strategies include camera fault detection. When the brightness variance of 5 consecutive frames is below a threshold of 50, it is determined to be lens occlusion, and the system automatically switches to the backup camera channel; local caching is enabled during network interruption, with a cache capacity of 1000 frames, managed using a FIFO queue to ensure no data loss; the image acquisition unit outputs data in a custom binary protocol format. Each frame contains a header information area, an image data area, and a pose parameter area. The header information area defines the data length, timestamp, and verification field. The image data area uses a lossless compression algorithm with a compression ratio controlled within 2:1. The pose parameter area stores a rotation matrix of 9 floating-point numbers and a translation vector of 3 floating-point numbers, with single-precision floating-point precision.

[0020] The 3D reconstruction unit receives multi-view image sequences and pose parameters transmitted by the image acquisition unit and performs high-precision 3D point cloud model generation. The unit uses a motion reconstruction structure-based algorithm to process image data. The algorithm execution flow is divided into five stages: key point extraction, feature matching, geometric constraint solving, bundle adjustment optimization, and point cloud post-processing. In the key point extraction stage, a scale-invariant feature transform operator is applied to construct a Gaussian difference pyramid. The pyramid has 8 layers, a scale factor of 1.6, and a key point detection threshold of 0.04. Stable feature points are selected using the Hessian matrix determinant, and the number of key points extracted from each image is controlled to within 5000. In the matching phase, keypoint descriptors are calculated using 128-dimensional SIFT descriptors. Similarity is measured using Euclidean distance, with a matching distance threshold set to 0.7, retaining only matches with a distance less than the threshold. In the mismatch removal phase, a random sampling consensus algorithm is applied, with an upper limit of 1000 iterations and an inlier determination threshold of 1.5 pixels. Reliable matching point pairs are selected by minimizing reprojection error, and an image quality alarm is triggered when the inlier ratio falls below 30%. In the geometric constraint solution phase, the fundamental matrix is ​​calculated using the eight-point method, decomposing it to obtain the camera extrinsic parameter matrix. The extrinsic parameter matrix contains a 3×3 rotation matrix and a 3×1 translation vector, and the decomposition process employs the singular value decomposition algorithm. The decomposition error is controlled within 0.1 degrees. During the bundle adjustment optimization stage, all camera extrinsic parameters and 3D point cloud coordinates are solved simultaneously. The optimization objective function is to minimize the reprojection error, expressed as minimizing the sum of squared reprojection distances of all matching point pairs. A robust kernel function, the Geman-McClure function, is introduced during the optimization process, with its parameter set to 1.0. This effectively reduces the impact of outliers on pose solving, and the reprojection error is controlled within 0.5 pixels through iterative optimization. The bundle adjustment convergence condition is set to a parameter update amount less than 10⁻⁶ or an iteration count reaching 500. A multi-view stereo matching algorithm is used in the point cloud generation stage, with the disparity calculation range set... The pixel size is 50 pixels. The matching cost function uses normalized cross-correlation. The cost aggregation window size is 5×5 pixels. The point cloud density is controlled by an adaptive sampling strategy to ensure that the final density is not less than 10,000 points per square meter. The point cloud coordinate system is based on the geometric center of the masonry as the origin, with the X-axis along the length direction, the Y-axis along the width direction, and the Z-axis along the height direction. The coordinate unit is millimeters. The origin of the coordinate system is determined by fitting the center of the circumscribed rectangle of the masonry using the least squares method. In the post-processing stage of the point cloud, filtering and smoothing operations are performed. The statistical outlier removal algorithm sets the number of neighborhood points to 30 and the standard deviation multiple threshold to 1.5. The point cloud smoothness error is controlled to 0 using the moving least squares method.Within 1 mm; coordinate normalization scaling scales the point cloud to a standard reference frame, with the scaling factor determined by the ratio of the point cloud bounding box diagonal to the standard length of 2 meters, eliminating evaluation bias caused by scale differences; the unit incorporates a data closed-loop verification mechanism, randomly selecting 5% of the detection samples, projecting the 3D point cloud model onto the original image plane, calculating the matching error between the projected points and image feature points, and automatically triggering the reconstruction parameter calibration process when the average error exceeds 0.8 pixels. The convergence threshold of the adjustment bundle in the calibration process is relaxed from 10⁻⁶ to 10⁻⁵ and the point cloud filtering intensity is enhanced. The filter intensity parameter increment is set to 0.1. Anomaly handling strategies include illumination change detection: when the standard deviation of image brightness changes by more than 20% per second, the system pauses reconstruction and re-acquires images; point cloud hole detection: when the point cloud density in a local area is less than 5000 points per square meter, a reshoot mechanism is activated, controlling the camera array to adjust the shooting angle; data transmission uses a shared memory mechanism, and the 3D reconstruction unit outputs point cloud data in PCD binary format, containing point cloud coordinates, normal vectors, and color information. The coordinate accuracy is 0.01 mm, and the normal vector calculation uses the K-nearest neighbor algorithm, with K set to 20.

[0021] The defect analysis unit receives the dense point cloud model generated by the 3D reconstruction unit and performs defect feature extraction and quantitative evaluation. The unit contains a dual-core processing module consisting of a deep learning model and a rule engine. The processing flow is serial, with the deep learning model output directly connected to the rule engine input. The deep learning model uses an improved PointNet architecture. The input layer receives 3D point cloud data, which is then normalized to a coordinate range of -1 to 1. The first-level shared multilayer perceptron layer has 64 neurons, using LeakyReLU as the activation function, with a slope parameter set to 0.01, and weights initialized using a He normal distribution. The second level has 128 neurons, performing feature transformation operations. The transformation matrix has a dimension of 64×64 and is regularized. The coefficient is set to 0.001; the third level has 256 neurons, introducing an attention mechanism with 4 attention heads, and the dimension of the query, key, and value vectors is 64; the fourth level has 512 neurons, performing spatial encoding with a resolution of 32×32×32 voxel grid; the fifth level has 1024 neurons, and the feature aggregation layer captures global context information through max pooling with a pooling kernel size of 1024; the output layer generates identification results for three types of defects: cracks, holes, and misalignments, including defect type labels, spatial coordinates, width, length, and depth quantization parameters; non-maximum suppression is performed during the model inference stage, the overlapping region suppression threshold is set to 0.3, and the bounding box merging strategy uses a weighted average method to ensure... The defect boundary localization is accurate. The deep learning model parameters come from pre-trained weights and on-site fine-tuning weights. Pre-trained weights were obtained through supervised learning on the ShapeNet dataset, while fine-tuning weights were trained using 10,000 labeled masonry defect samples. These training samples covered different lighting conditions, surface contamination, and structural morphologies. Data augmentation strategies included random rotation angles ranging from -15 degrees to 15 degrees, translation distances ranging from -0.1 meters to 0.1 meters, and a Gaussian noise standard deviation of 0.01. The loss function was designed as a combination of weighted cross-entropy and IoU loss, with weight coefficients set to 0.7 and 0.3, focusing on improving the recognition accuracy of small-sized defects. The model health diagnosis module is executed periodically, analyzing the false positive rate using a confusion matrix. The system detects missed detections at a rate exceeding 3%, prompting for additional training samples. The diagnostic cycle is set to every 100 tests. The rule engine module pre-configures a quality threshold matrix based on the water conservancy industry standard SL274-2020. This threshold matrix is ​​stored in an SQLite database and includes three fields: defect type, quantitative parameter threshold, and judgment level. The crack judgment logic marks a crack as a serious defect when its width is greater than 2 mm and its length exceeds 50 mm. Width is measured by calculating the maximum distance between boundary points of the defect area, and length is measured by fitting a centerline and calculating the arc length. The cavity judgment logic identifies cavities with a volume greater than 1000 cubic millimeters as structural hazards. The volume calculation uses an Alpha shape algorithm with an Alpha value set to 0.0.05 meters; the misalignment judgment logic classifies misalignment displacement exceeding 15 mm as a construction deviation. The displacement is calculated by matching feature points on the surfaces of adjacent stones and solving for rigid body transformation; the rule engine supports a dynamic update mechanism, iteratively optimizing parameters based on actual engineering data fed back from the quality management system. The update cycle is set to 7 days, and the update process uses a weighted average fusion of old and new thresholds, with a fusion weight set to 0.2; data transmission latency is controlled within 100 milliseconds using zero-copy technology. Transmitted data includes defect type, location coordinates, and quantization parameters. The location coordinates use a three-dimensional rectangular coordinate system, and the quantization parameter accuracy is 0.1 mm; the anomaly handling strategy includes model inference timeout detection, triggering a degradation mode when the processing time exceeds 200 milliseconds, and enabling a simplified rule engine in degradation mode; data repair is initiated when point cloud data is abnormal, and the repair algorithm fills in missing areas through interpolation; the unit has a built-in input data verification mechanism, with the point cloud coordinate range check threshold set to -5 meters to 5 meters, and data exceeding this range is marked as invalid.

[0022] The decision output unit receives the quantitative evaluation results from the defect analysis unit and performs a comprehensive quality level judgment. The unit has a built-in dynamic threshold matrix, setting three quality standards based on defect type combinations: when more than three severe defects are detected, the system automatically classifies it as unqualified; if one or two severe defects exist, it is marked as requiring repair; otherwise, it is classified as qualified. The quality judgment logic employs a multi-condition priority strategy. When cracks and voids exist simultaneously, the system prioritizes evaluating the volume ratio of voids. If the void volume ratio exceeds 5% of the total masonry volume, it is directly classified as unqualified, ignoring other defect parameters. The volume ratio calculation is achieved through a voxelized point cloud model and counting the number of void voxels, with a voxel resolution set to 10 mm × 10 mm × 10 mm. The quality judgment report includes a defect distribution heatmap, a quantitative parameter table, and repair suggestion text. The heatmap generation uses Gaussian kernel density estimation, with a kernel bandwidth parameter set to 0.5 meters and a color mapping range of 0 to 100, corresponding to defect densities from low to high. The system is highly efficient; the quantitative parameter table is categorized by defect type, including location coordinates, size parameters, and confidence levels. Confidence level calculation is based on a weighted average of the model output probability and the confidence factor from the rule engine, with weighting coefficients set to 0.6 and 0.4 respectively. The repair suggestion text generation module, based on defect type and location information, calls a pre-built knowledge base to output standardized treatment plans: for cracks requiring repair, epoxy resin grouting is recommended, with a grouting pressure set at 0.5 MPa and a grouting material viscosity range of 50 to 100 mPa·s; for structural voids, loose materials should be removed before filling with C30 fine aggregate concrete, with a concrete slump controlled between 100 mm and 150 mm. The report template conforms to the GB / T1.1-2020 standard format, including a detection timestamp, equipment serial number, and digital signature. The digital signature is generated using the RSA-2048 algorithm, with a signature validity period set to 30 days. Report generation time is controlled within 5 seconds using parallel rendering technology, with heatmap generation and text generation performed synchronously. The decision output unit uses standard OPC. The UA protocol interface interfaces with the quality management system. OPC UA session establishment uses X.509 certificate authentication, with a certificate validity period of 365 days. Data transmission encryption uses the AES-256 algorithm, with a key rotation cycle of 24 hours. The historical record tracing function supports searching by time range, device ID, or defect type, with a search response time of less than 1 second. Anomaly handling strategies include enabling local storage when interface communication fails, with a storage capacity threshold set at 10 gigabytes, automatically compressing old data when the threshold is exceeded. When a logical conflict is detected, a manual review process is initiated, with the system generating a review request and notifying quality inspectors. The unit has a built-in time synchronization mechanism, using the NTP protocol to calibrate the system clock, with a clock deviation tolerance set at 100 milliseconds.

[0023] The system integrates an environmental sensing submodule to monitor environmental parameters at the detection site in real time. This submodule includes a temperature sensor, a humidity sensor, and a vibration accelerometer. The temperature sensor is a PT1000 model, with a measurement range of -20°C to 80°C, an accuracy of ±0.5°C, and a sampling frequency of 1 Hz. The humidity sensor is a HDC2080 model, with a measurement range of 30% to 90% relative humidity, an accuracy of ±2%, and a response time of less than 5 seconds. The vibration accelerometer is an ADXL355 model, with a range of 0 to 10 m / s², a resolution of 0.001 m / s², and a sampling frequency of 100 Hz. Environmental data acquisition uses a polling mechanism with a polling interval of 100 milliseconds. Data is stored in a circular buffer with a buffer depth of 1000. A compensation mechanism dynamically adjusts the detection process based on environmental parameters: when the temperature change rate exceeds 2°C per minute, an image color correction algorithm is activated to correct the temperature. The algorithm calculates color temperature offset based on a blackbody radiation model. When the offset exceeds 500 Kelvin, adaptive white balance is applied. When the vibration amplitude exceeds 0.2 m / s², the camera exposure time is extended to 1 / 60 second, and an image stabilization algorithm is activated. The stabilization algorithm calculates the displacement vector through feature point tracking and performs pixel-level compensation. Environmental data and detection results are stored synchronously to form a complete quality assessment context. The stored records include timestamps, environmental parameter values, and compensation operation flags. The data storage module adopts a distributed architecture. Raw detection data is retained for 30 days, and structured reports are permanently archived. Storage space management implements an automatic cleanup strategy. Cleanup is triggered when the storage capacity reaches 90%. The cleanup strategy prioritizes deleting raw image data older than 30 days, while retaining point cloud models and report data. Raw data compression uses the LZ4 algorithm with a compression ratio of 3:1, and point cloud data compression uses the Draco algorithm with a compression ratio of 5:1.

[0024] The central processing server is equipped with dual Intel Xeon Silver 4310 processors, each with 12 cores and 24 threads, a base clock speed of 2.1 GHz, and a base TDP of 85 watts; it has 64 gigabytes of memory, using four 16-gigabyte DDR4-2666 ECC memory modules with timings set to 19-19-19; the GPU is an NVIDIA Tesla T4 with 16 gigabytes of video memory, a core frequency of 585 MHz, and a floating-point performance of 8.1 trillion operations per second; the storage system uses two 1-terabyte NVMe SSDs in a RAID 1 array, with read / write speeds of 3500 megabytes per second; the network interface is dual Gigabit Ethernet cards supporting link aggregation; the operating system is Ubuntu Server 20.04. LTS kernel parameter optimizations include increasing the network buffer size to 4 megabytes and adjusting the process scheduling strategy to real-time priority. System services are deployed using Docker containers, with image acquisition, 3D reconstruction, defect analysis, and decision output services running in independent containers. Container resource limits are set to 8 CPU cores and 16 gigabytes of memory. Inter-service communication is implemented via the gRPC protocol, with Protocol Buffers used for transmission serialization, and message latency controlled within 50 milliseconds. The system processes 1200 masonry stone inspection tasks per hour, with an average processing time of 3 seconds per stone, including 0.5 seconds for image acquisition, 1.2 seconds for 3D reconstruction, 0.8 seconds for defect analysis, and 0.5 seconds for decision output. Processing capacity verification was completed through stress testing using a dataset of 10,000 simulated masonry stones. After 24 hours of continuous operation, the system's throughput fluctuation was less than 5%.

[0025] The system's data flow strictly follows time-series logic. When a detection task starts, the environmental perception submodule first collects on-site environmental parameters. If the parameters exceed a preset threshold, a compensation mechanism is triggered. Subsequently, the image acquisition unit performs multi-view image capture, with synchronous trigger signals ensuring that all six cameras complete exposure within 0.1 milliseconds. After preprocessing, the acquired images are transmitted to the central processing server via gigabit Ethernet, with bandwidth occupancy controlled within 70%. After receiving the image data, the 3D reconstruction unit starts the reconstruction process. The key point extraction stage takes 20 milliseconds per image, the feature matching stage takes 30 milliseconds, and the bundle adjustment and optimization stage takes 500 milliseconds. The reconstructed point cloud model is transmitted to the defect analysis unit after coordinate normalization, with a transmission latency of less than 10 milliseconds. The defect analysis unit's deep learning model inference takes 800 milliseconds, and the rule engine judgment takes 50 milliseconds. After receiving the defect data, the decision output unit performs quality judgment, with heatmap generation taking 2 seconds and report assembly taking 3 seconds. The final report is transmitted via OPC. The UA interface uploads data to the quality management system using a chunked transmission process, with each chunk being 1 megabyte in size. Data storage paths include a raw image cache, a point cloud temporary storage area, a defect analysis results area, and a report archiving area, with dynamic capacity allocation for each area. The system has a built-in end-to-end monitoring module that collects real-time data on processing latency, resource utilization, and error counts for each unit, summarizing the monitoring data every 5 seconds. Error handling mechanisms are tiered: Level 1 errors include hardware failures, triggering system shutdown; Level 2 errors include data anomalies, triggering a retry process; and Level 3 errors include performance fluctuations, logging but not interrupting the process.

[0026] The system's performance parameters have been rigorously calibrated. Detection accuracy has been verified by comparing with manual inspection results. The test set includes 500 masonry samples with known defects. The system's crack identification accuracy reaches 98.5%, void identification accuracy reaches 97.2%, and misalignment identification accuracy reaches 99.1%. In terms of efficiency, the processing time for a single piece of masonry is consistently within 3 seconds, 8 times faster than traditional photogrammetry. The quality grade judgment compliance rate has been verified by actual engineering data, achieving a 99.2% consistency rate with expert judgments in 1000 inspections. Environmental adaptability testing was conducted under conditions of -10°C to 40°C, humidity 40% to 80%, and vibration amplitude of 0.1 m / s², with test result fluctuations less than 2%. The system-generated structured reports directly interface with the quality management system, supporting statistical analysis and trend prediction of defect data. The statistical analysis module uses a time series decomposition algorithm, and the trend prediction uses an ARIMA model with a prediction period set to 30 days.

[0027] The system engineering implementation details include hardware installation specifications; the camera array mounting bracket is made of aluminum alloy with a stiffness coefficient greater than 100 N / m to ensure pose stability under vibration; the light source system's heat dissipation design uses a combination of air cooling and heat pipes, with temperature rise controlled within 15 degrees Celsius; the central processing server is deployed in an IP54 protection-rated cabinet with an operating temperature range of 0 to 40 degrees Celsius; the software deployment process includes system initialization, parameter configuration, and self-testing; the initialization phase loads camera calibration parameters, 3D reconstruction algorithm parameters, and rule engine threshold matrix; parameter configuration is completed through a web interface, with configuration items including detection area size, quality judgment threshold, and environmental compensation parameters; self-testing verifies the functions of each unit, with image acquisition unit testing including camera connectivity and synchronization accuracy, 3D reconstruction unit testing including point cloud density and smoothness, defect analysis unit testing including model inference accuracy, and decision output unit testing including report generation completeness; the system startup time is controlled within 60 seconds from power-on to ready state.

[0028] The system security mechanism encompasses both physical and data security. Physical security includes a camera housing with an IK10 shock resistance rating and an overcurrent protection threshold for the light source system set at 120% of the rated current. Data security employs multi-layered protection: TLS 1.3 protocol is used for transmission layer encryption, AES-256 algorithm is used for application layer encryption, and a hardware security module (HSM) is used for key management, with a key generation cycle of 7 days. Access control is based on a role-based access control model, defining three roles: system administrator, quality inspector, and auditor. Permission allocation is achieved through attribute-based encryption. Audit logs record all critical operations, with a retention period of 180 days. Security vulnerability scans are performed weekly using Nessus, and vulnerability remediation response time is less than 24 hours.

[0029] The system maintenance strategy includes preventative maintenance and fault repair. Preventative maintenance is scheduled every 30 days and includes cleaning optical lenses, calibrating sensors, and updating software patches. The fault repair process is defined as four stages: fault detection, isolation, diagnosis, and recovery. Fault detection is achieved through a monitoring module, with a threshold set at three consecutive timeouts or more than five error counts. Fault isolation employs a service circuit breaker mechanism, with a circuit breaker threshold set at an error rate of 20%. Fault diagnosis is based on knowledge base matching, which stores 500 fault modes and solutions. Fault recovery executes automated scripts with a required script execution success rate of 95%. Maintenance records are stored electronically, including the fault time, cause, and handling results.

[0030] The system's scalability design supports the addition and deletion of functional modules. When adding a detection unit, unit information, including input / output data formats and processing latency, is registered through the configuration management interface. Inter-unit communication adopts a publish-subscribe model, with topic naming conventions as / detection / unit type / instance ID. The load balancing strategy is based on real-time resource utilization, with thresholds set at 80% CPU or 90% memory. When resource utilization exceeds the limit, the system automatically migrates some tasks to a backup node. The backup node is configured identically to the primary node, and the switching time is controlled within 10 seconds. System upgrades employ a rolling update strategy, maintaining service availability during the update process, and update verification is completed through A / B testing.

[0031] The system integrates with external systems according to specifications; the quality management system interface adopts the OPC UA protocol, and the address space definition includes inspection task nodes, result nodes, and report nodes; the inspection task node attributes include task ID, masonry number, and inspection time; the result node attributes include defect list and quality level; the report node provides a binary report download interface; the data exchange format follows the IEC62541 standard, and the message security policy is set to signature and encryption; historical data traceability is achieved through the OPC UA historical access service, and the query conditions support time range and device ID; the system clock synchronization adopts the IEEE1588 protocol, and the master clock accuracy is 100 nanoseconds.

[0032] The system verification methods included unit testing, integration testing, and field testing. Unit testing covered the functions of each module. Image acquisition unit test cases included synchronization accuracy measurement and light source uniformity verification; 3D reconstruction unit test cases included point cloud density detection and reprojection error measurement; defect analysis unit test cases included defect identification accuracy testing and rule engine logic verification; decision output unit test cases included report format checking and judgment logic verification. Integration testing verified the end-to-end process, with test scenarios including standard masonry, masonry with defects, and environmental interference scenarios. Field testing was conducted at three water conservancy project sites over a period of six months, with a sample size of 20,000 pieces. Verification results showed that the system's detection accuracy reached 98%, an improvement of 40 percentage points compared to the manual tapping detection method; and the quality grade judgment results achieved a 99.2% consistency rate with the actual engineering conditions.

[0033] The system engineering deployment process is divided into four stages: site survey, equipment installation, system debugging, and acceptance. The site survey determines the size of the detection area, lighting conditions, and environmental parameter range. Equipment installation includes fixing the camera bracket, setting up the light source system, and installing the server rack. System debugging involves parameter optimization, including fine-tuning the aperture value and setting the exposure time for the camera; adjusting the beam convergence threshold and setting the point cloud filter intensity for the reconstruction parameter optimization; and calibrating the defect analysis parameter optimization, including calibrating the model threshold and updating the rule engine threshold. Acceptance criteria include a detection accuracy of ≥97%, a processing time of ≤5 seconds, and a report generation completeness rate of 100%. Deployment documentation includes an installation manual, operation guide, and maintenance procedures.

[0034] The system's technological advantages are reflected in the simultaneous quantification of multi-dimensional quality parameters; crack parameter quantification includes width, length, depth, and orientation, with width measurement achieved by calculating the minimum distance across the defect cross-section, achieving a measurement accuracy of 0.1 mm; void parameter quantification includes volume, location, and shape, with volume calculation using the Monte Carlo integration method and a sampling point count of 10,000; misalignment parameter quantification includes displacement, angular deviation, and contact area, with displacement calculated through point cloud registration; parameter quantification results are stored in a structured database, supporting multi-dimensional statistical analysis; the statistical analysis module calculates defect distribution histograms, correlation matrices, and trend curves; correlation analysis uses the Pearson correlation coefficient with a threshold set at 0.7; trend prediction uses exponential smoothing with a smoothing coefficient set at 0.3.

[0035] System data lifecycle management; raw image data is automatically archived to cold storage after 30 days; point cloud model data is permanently saved and compressed to save space; defect analysis results data and report data are stored together, with the association key being the detection task ID; data backup strategy is set to daily incremental backup and weekly full backup; backup data is stored on a remote server, and transmission is encrypted using the SFTP protocol; data recovery testing is performed monthly, with a recovery time target set at 2 hours; data privacy protection complies with GDPR requirements, and personally identifiable information is anonymized.

[0036] The system features a human-computer interaction interface; quality inspectors monitor the inspection process via a touchscreen terminal; the interface layout includes a real-time image display area, a point cloud model visualization area, a defect marking area, and a report preview area; the real-time image display area supports 6-view image switching with a refresh rate of 30 frames per second; the point cloud model visualization area supports rotation, scaling, and sectioning operations with a rendering frame rate of 25 frames per second; the defect marking area displays the defect type, location, and parameters, with color markings distinguishing defect levels; the report preview area displays heatmaps and key parameters; operation commands include starting inspection, pausing inspection, and parameter adjustment; system status indicator lights display running, warning, and error statuses.

[0037] System energy consumption management: Camera array standby power consumption 50W, operating power consumption 200W; Light source system standby power consumption 30W, operating power consumption 150W; Central processing server standby power consumption 100W, operating power consumption 500W; Energy consumption optimization strategies include automatic frequency reduction during idle periods, with CPU frequency reduced to 1.0 GHz; System operating modes are divided into high-performance mode and energy-saving mode, with energy-saving mode extending processing time by 20% and reducing power consumption by 30%; Real-time monitoring of energy consumption data, with energy consumption per unit detection task controlled within 0.05 kWh.

[0038] The system's technical specifications meet industry standards; its detection accuracy meets the requirements of SL274-2020; its processing speed meets the engineering requirement of 1200 pieces per hour; its report format conforms to GB / T1.1-2020 standard; its safety performance is certified by ISO13849-1; its environmental adaptability has passed the IP54 protection level test; and its system reliability index, mean time between failures (MTBF), is set at 10,000 hours.

[0039] The system addresses four major technical bottlenecks in manual inspection: efficiency bottlenecks are solved through automated processes, reducing single-block inspection time from 30 minutes to 3 seconds; subjectivity bottlenecks are resolved through quantitative evaluation, increasing the quality judgment compliance rate to 99.2%; internal defect identification bottlenecks are solved through 3D reconstruction, achieving a void detection accuracy of 97.2%; and multi-parameter synchronous quantification bottlenecks are resolved through structured reports, supporting simultaneous output of 10-dimensional quality parameters. The system implementation significantly reduces safety risks caused by masonry quality issues in water conservancy projects, reducing the defect miss rate from 20% to below 2%.

[0040] The innovation of the system's technical solution lies in the deep coupling of 3D reconstruction and defect analysis. Traditional methods separate the processing of surface images and internal structures, while this system associates surface features with internal defects through a high-precision point cloud model. The point cloud model serves as a unified data carrier, supporting crack depth inference and cavity volume calculation. The coupling mechanism is achieved through coordinate alignment, with the conversion error between image feature points and point cloud coordinates being less than 0.1 mm. This mechanism enables the system to detect hidden defects that are difficult to detect manually, such as internal cavities and microcracks.

[0041] In the system engineering implementation, parameter settings are based on rigorous verification; the camera frame rate of 30 frames per second is based on motion blur analysis to ensure clear texture of the masonry surface; the point cloud density of 10,000 points per square meter is determined through a trade-off between accuracy and efficiency, as the defect recognition rate decreases below this value; the bundle adjustment reprojection error of 0.5 pixels is set through extensive testing, as increased error leads to model distortion; the rule engine threshold is based on historical engineering data statistics, with a crack width of 2 mm corresponding to the critical value for structural safety; the system parameter setting document includes verification data and setting basis to ensure traceability.

[0042] The system's technical solution is universally applicable; it is suitable for the inspection of masonry structures in the fields of water conservancy, transportation, and construction; the size range of the inspected objects is 0.5m × 0.5m to 3m × 3m; the surface roughness range is Ra3.2μm to Ra50μm; the illumination range is 50lux to 10000lux; the system configuration parameters are adjusted according to the application scenario, reducing the point cloud density to 5000 points per square meter for small structures and extending the image acquisition time to 1 second for large structures; universality verification was completed in 5 types of engineering scenarios, with a detection success rate of over 95% in all cases.

[0043] The system's technical solution demonstrates sustainability; the software architecture supports algorithm iteration, and new models are seamlessly integrated through containerized deployment; the hardware design reserves expansion interfaces to support the addition of new sensor types; data storage supports long-term archiving to meet engineering quality traceability requirements; the system upgrade path planning includes introducing higher resolution cameras, integrating LiDAR, and enhancing AI models; sustainability assessment shows that the system's lifecycle cost is reduced by 40% compared to traditional methods.

[0044] The system's technical solution is economical; hardware costs are controlled within 500,000 RMB; testing costs are reduced from 50 RMB per unit for manual work to 5 RMB per unit; the investment payback period is set at 12 months; the economic benefit analysis includes direct cost savings and indirect risk reduction; the risk reduction value is estimated based on historical accident data, with 5 million RMB saved for each accident avoided; significant economic returns can be achieved within 3 years of system implementation.

[0045] Ecological impact of the system's technical solution: Reduces the number of on-site inspections for manual testing, thus lowering carbon emissions; No chemical reagents are used in the testing process, making it environmentally friendly; Equipment and materials comply with RoHS standards; Energy consumption optimization strategies reduce power consumption; Ecological impact assessment shows that the system implementation will reduce carbon emissions by 10 tons per year.

[0046] The system's technical solution is legally compliant; it conforms to GB50203-2011 Code for Acceptance of Construction Quality of Masonry Structures; the test report meets the legal requirements for engineering quality appraisal; data security complies with the Cybersecurity Law; the intellectual property layout covers the core algorithm and system architecture; the compliance documentation includes a declaration of standard compliance and certification certificates.

[0047] The system's technical solution offers engineering value; it provides quantitative decision-making support for engineering quality control; it supports trend prediction and preventative maintenance of defect data; it improves the efficiency and objectivity of engineering acceptance; it reduces the incidence of quality disputes; and engineering application cases show that the system increases the pass rate of masonry construction by 15 percentage points.

[0048] The system's technical solutions maintain technical continuity; core algorithms are continuously optimized, with point cloud reconstruction algorithms upgraded to neural radiation field technology; the defect analysis model is migrated to the Transformer architecture; the system architecture evolves towards edge computing to support real-time on-site processing; the technology roadmap outlines a 3-year R&D plan, improving target detection accuracy to 99%.

[0049] The system's technical solutions are implemented with support; 24 / 7 technical support is provided; a spare parts inventory is established, with 100% coverage of key components; the training plan includes operation and maintenance training; and a support system is implemented to ensure stable system operation.

[0050] The final output of the system's technical solution includes a structured quality assessment report containing a defect distribution heatmap, a quantitative parameter table, and repair suggestion text. The heatmap uses pseudo-color coding, with red indicating high-density defect areas. The quantitative parameter table is sorted by defect type. The repair suggestion text includes process parameters and material specifications. The electronic version of the report is uploaded in real time via the OPCUA interface, and the paper version supports printing. The legal validity of the report is guaranteed by digital signature.

[0051] Closed-loop verification of the system's technical solution; comparative analysis of test results and manually reviewed data; system calibration triggered when deviation exceeds 5%; calibration process includes parameter optimization and model fine-tuning; closed-loop verification ensures long-term system stability; verification cycle is set at every 1000 tests.

[0052] The system's technical solutions are integrated into engineering projects; seamlessly integrated with construction management systems; data-driven construction adjustments; a quality early warning mechanism intervenes in the construction process in advance; engineering integration improves overall construction quality; integrated case studies show a 30% reduction in rework rates. Continuous improvement of system technical solutions; collection of on-site issues through user feedback mechanisms; inclusion of improvement requirements in the R&D queue; setting of version iteration cycles at 3 months; continuous improvement to ensure the system adapts to changes in engineering requirements.

[0053] Standardized output of system technical solutions; technical documentation including system specifications, interface specifications, and test reports; training materials including operation videos and maintenance manuals; standardized output supports rapid deployment.

[0054] Final verification of the system's technical solution; certification by a third-party testing organization; certification testing covering all technical indicators; certification certificate serving as quality assurance; verification process conforming to ISO / IEC 17025 standards.

[0055] An engineering deployment example of the system's technical solution: Ten sets of the system were deployed in a reservoir slope protection project; the masonry slope protection area was inspected, covering 5,000 square meters; 23 serious defects were found, all of which were repaired in a timely manner; the project passed acceptance on the first attempt; the system operated for 6 months without any major failures; the example proves the feasibility of the system engineering.

[0056] The overall benefits of the system's technical solution include: increasing testing efficiency by 8 times; reducing testing costs by 90%; improving the accuracy of quality judgment by 40 percentage points; reducing engineering risks by 50%; and quantifying the overall benefits as a cost saving of 20 yuan per square meter of testing area.

[0057] Technical summary of the system solution: Through the collaborative work of the image acquisition unit, 3D reconstruction unit, defect analysis unit, and decision output unit, a complete closed loop for automatic quality inspection of masonry is constructed; the system innovatively applies machine vision and deep learning technology to the traditional field of water conservancy engineering; it solves the efficiency and accuracy bottlenecks of manual inspection; the technical solution has high feasibility and engineering value; and it provides a reliable technical means for engineering quality control.

Claims

1. A machine vision-based quality inspection system for masonry masonry, characterized in that, include: The image acquisition unit is used to acquire multi-view high-definition image data of the surface of the masonry structure. The image acquisition unit consists of a ring array of multiple cameras, evenly distributed around the masonry detection area; the ring array of cameras adopts a synchronous triggering mechanism and establishes a real-time data transmission channel with the central processing server through a gigabit Ethernet interface. A 3D reconstruction unit is used to generate a high-precision 3D point cloud model based on multi-view images. The 3D reconstruction unit uses a motion reconstruction structure-based algorithm to process multi-view image sequences. The motion reconstruction structure-based algorithm extracts key points of the image through a scale-invariant feature transformation operator and uses a random sampling consensus algorithm to remove mismatched point pairs. During the execution of the beam adjustment and optimization process, the camera extrinsic parameter matrix and 3D point cloud coordinates are solved simultaneously. The defect analysis unit is used to extract and quantitatively evaluate defect features from a 3D point cloud model. The defect analysis unit includes a dual-core processing module consisting of a deep learning model and a rule engine. The deep learning model adopts an improved PointNet architecture. The input layer receives 3D point cloud data and performs feature extraction through a 5-level shared multilayer perceptron layer. The rule engine module is pre-loaded with the quality threshold matrix of the water conservancy industry standard SL274-2020. The decision output unit is used to generate a structured quality assessment report based on the quantitative evaluation results; the decision output unit has a built-in dynamic threshold matrix and sets three-level quality standards based on the combination of defect types; The quality assessment report includes a heat map of defect distribution, a table of quantitative parameters, and a text of repair recommendations.

2. The machine vision-based masonry quality inspection system according to claim 1, characterized in that, The camera array layout of the image acquisition unit follows the principle of minimum coverage angle; the image preprocessing module performs white balance correction and gamma correction in real time at the acquisition end.

3. The machine vision-based masonry quality inspection system according to claim 1, characterized in that, The bundle adjustment optimization process of the three-dimensional reconstruction unit introduces a robust kernel function to reduce the impact of outliers on pose solving; after the three-dimensional point cloud model is generated, coordinate normalization is automatically performed to scale the size to a standard reference system, eliminating evaluation bias caused by scale differences.

4. The machine vision-based masonry quality inspection system according to claim 1, characterized in that, The deep learning model training of the defect analysis unit adopts a transfer learning strategy; The base model is pre-trained on the publicly available ShapeNet dataset; the training samples cover different lighting conditions, surface contamination, and structural morphology; the loss function is designed as a combination of weighted cross-entropy and IoU loss; non-maximum suppression is performed during the model inference stage.

5. The machine vision-based masonry quality inspection system according to claim 1, characterized in that, The quality judgment logic of the decision output unit adopts a multi-condition priority strategy; when cracks and voids exist simultaneously, the system prioritizes evaluating the volume ratio of voids; the repair suggestion text generation module calls a pre-set knowledge base to output standardized treatment solutions based on the defect type and location information; for cracks requiring repair, epoxy resin grouting is recommended; for structural voids, loose materials should be removed before filling with concrete.

6. The machine vision-based masonry quality inspection system according to claim 1, characterized in that, The system integrates an environmental sensing submodule to monitor the temperature, humidity, and vibration parameters at the detection site in real time.

7. The machine vision-based quality inspection system for masonry masonry according to claim 1, characterized in that, A data closed-loop verification mechanism is established between the 3D reconstruction unit and the defect analysis unit; the system randomly selects a certain proportion of the detection samples, projects the 3D point cloud model onto the original image plane, and calculates the matching error between the projected points and the image feature points; when the average error exceeds the threshold pixels, the reconstruction parameter calibration process is automatically triggered to adjust the convergence threshold of the beam adjustment and the point cloud filtering intensity.

8. The machine vision-based masonry quality inspection system according to claim 1, characterized in that, The threshold matrix of the rule engine supports a dynamic update mechanism, which iteratively optimizes parameters based on actual engineering data fed back from the quality management system; the update process uses a weighted average to fuse the old and new thresholds. The defect analysis unit periodically performs model health diagnoses and analyzes the false positive rate and false negative rate using a confusion matrix.

9. The machine vision-based masonry quality inspection system according to claim 1, characterized in that, The heatmap of the quality assessment report is generated using Gaussian kernel density estimation; the quantification parameter table is classified according to defect type and includes location coordinates, size parameters and confidence level; the confidence level is calculated based on the model output probability and the confidence factor of the rule engine; the report template conforms to the GB / T1.1-2020 standard format and includes the inspection timestamp, equipment serial number and digital signature, and the digital signature is generated using the RSA-2048 algorithm.

10. A machine vision-based quality inspection system for masonry masonry according to claim 1, characterized in that, The output of the image acquisition unit is directly connected to the input of the 3D reconstruction unit, and the transmitted data includes the original image sequence and the corresponding camera pose parameters; the processing flow of the defect analysis unit is a serial structure, and the output of the deep learning model is connected to the input of the rule engine; the central processing server is configured with dual processors.

Images