An image recognition-based construction quality intelligent acceptance system

Abstract

The application discloses a construction quality intelligent acceptance system based on image recognition, and relates to the technical field of intelligent monitoring of construction quality, comprising an acceptance management center, which is communicatively connected with the following modules: a self-perception decision module, which comprehensively applies deep reinforcement learning decision and image recognition technology to adaptively perceive a target construction site and identify a reinforcing steel target, and determine each reinforcing steel intersection point and its pixel coordinates in an image. The scene self-adaptive acquisition unit driven by deep reinforcement learning can perceive dynamic interference such as illumination change and equipment shielding of the construction site in real time, autonomously decide optimal image acquisition parameters and paths, and break away from the dependence on fixed markers, so that the system can actively avoid interference in an unstructured scene, accurately position to a key area, and ensure that clear and complete reinforcing mesh images can be captured under various complex environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for building construction quality, specifically to an intelligent construction quality acceptance system based on image recognition. Background Technology

[0002] With the continuous development of the construction industry, construction quality management is facing increasingly higher requirements. Traditional manual acceptance methods have problems such as low efficiency, strong subjectivity, and susceptibility to interference from environmental and human factors, which cannot meet the refined quality control requirements of modern construction projects. The introduction of image recognition technology provides a new solution for intelligent acceptance of construction quality. By combining computer vision and deep learning algorithms, image recognition can efficiently and accurately detect various quality problems in the construction process, such as material defects and deviations in construction techniques.

[0003] For example, Chinese Patent Publication No. CN110163843A describes an auxiliary acceptance method for floor slab reinforcement engineering based on image acquisition. This method automatically judges the construction quality of floor slab reinforcement engineering, quickly identifies areas with poor reinforcement binding quality, helps inspectors reduce workload, and ultimately improves the construction quality of engineering projects.

[0004] In existing technologies, traditional methods for construction quality acceptance often rely on fixed marker boards for dimensional reference, which is difficult to adapt to the dynamic changes in the construction site environment, such as equipment obstruction and uneven lighting, leading to inaccurate rebar identification. Usually, only static images are analyzed after the fact, which cannot detect in real time the uneven spacing of local rebars and loosening of ties caused by improper binding operations or material deformation during construction. At the same time, it lacks the ability to predict the evolution trend of quality during construction and cannot provide early warning before the gradual degradation of rebar binding quality occurs. Therefore, it is difficult to achieve intelligent monitoring and proactive intervention of construction quality throughout the entire process. To address these problems, an intelligent construction quality acceptance system based on image recognition is proposed. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention is implemented through the following technical solution: an intelligent construction quality acceptance system based on image recognition, including an acceptance management center, wherein the acceptance management center is communicatively connected to the following modules:

[0006] The autonomous perception and decision-making module comprehensively utilizes deep reinforcement learning decision-making and image recognition technologies to dynamically and adaptively perceive complex construction sites and robustly identify steel reinforcement targets. It determines the intersection points of steel reinforcement and their pixel coordinates in the image, enabling the system to break free from dependence on static markers, autonomously adapt to complex environments, and achieve real-time, proactive, and accurate capture of steel reinforcement targets.

[0007] The topology defect detection module is used to abstract the steel mesh into a graph structure and use graph neural networks for analysis, which elevates visual detection to the level of spatial relationship understanding, and performs quantitative evaluation and defect location of structural attributes such as connection relationships and spacing, surpassing the limitations of traditional image processing.

[0008] The quality time series modeling module is used to organize the data on the location and topological relationship of rebar nodes collected in multiple batches according to the construction time sequence, and to organize the rebar topology map status obtained at different time points in sequence to construct a time series sequence that reflects the quality evolution and describes the dynamic evolution process of binding quality.

[0009] The degradation trend early warning module is used to learn the inherent laws of time sequence using long short-term memory networks, predict the degradation trend of rebar binding quality in the next construction cycle, and generate early warning signals.

[0010] The intelligent monitoring and intervention module is used to integrate defect detection results and quality trend predictions to generate a comprehensive construction quality assessment report. Based on the assessment results, it generates targeted rectification suggestions and intervention measures to guide on-site quality correction.

[0011] Preferably, the autonomous perception and decision-making module includes a scene adaptive acquisition unit and a rebar node accurate identification unit;

[0012] The scene adaptive acquisition unit is used to use a deep Q-network based on reinforcement learning to perceive the dynamic interference of the target construction site, including changes in ambient light, equipment movement and occlusion, and autonomously decide the optimal image acquisition parameters and path to obtain the original image sequence of the target construction site.

[0013] The rebar node precision identification unit is used to accurately identify and locate each rebar intersection in the image based on the adaptively acquired target construction site image, and combine deep learning target detection and image enhancement algorithms to overcome the recognition interference caused by uneven lighting, rebar reflection and partial occlusion, determine its pixel coordinates, and transform the rebar mesh in the dynamic environment into a high-precision rebar intersection pixel coordinate data set.

[0014] Preferably, the scene adaptive acquisition unit performs the following steps:

[0015] Based on multi-source sensors mounted on mobile acquisition devices, the environmental status characteristics of the target construction site are acquired in real time. These environmental status characteristics include changes in light intensity, the dynamic position and movement trajectory of large construction equipment, and the degree of temporary occlusion of key areas of the steel mesh, providing a dynamic data foundation for subsequent environmental perception.

[0016] The environmental state features are used as input to drive a reinforcement learning decision model based on a deep Q-network. Through interactive learning with the environment, the model outputs the current optimal image acquisition control parameters and device movement path planning. The control parameters include focal length, exposure compensation, and shooting angle to achieve adaptive acquisition decision-making against dynamic interference.

[0017] Based on the control parameters output by the reinforcement learning decision model and the device movement path planning, the pose and optical system of the mobile acquisition device are autonomously adjusted to dynamically avoid interference in unstructured scenes and actively locate the acquisition point that can clearly and completely capture the key area of ​​the steel mesh to be inspected, so as to obtain the original image sequence of the target construction site and ensure that the original image sequence that can be used for high-precision analysis is obtained.

[0018] Preferably, the rebar node accurate identification unit performs the following steps:

[0019] The acquired original image sequence is subjected to adaptive illumination equalization processing based on Retinex theory and polarization filtering simulation algorithm for metal reflection to suppress uneven illumination and specular noise, enhance the texture and contour features of steel bars in the image, and the processed image has uniform overall brightness and significantly improved contrast between steel bar edges and background, which facilitates stable recognition by subsequent algorithms.

[0020] The preprocessed image sequence is input into a trained deep learning target detection network. Image recognition is performed with the rebar intersection as the central target, and the bounding box of each identified rebar intersection and its precise pixel coordinates in the image coordinate system are output, forming a preliminary set of node positions. The detection network can quickly locate most intersections and initially form a comprehensive set of node coordinates, laying the foundation for subsequent accurate measurement.

[0021] Based on a multi-frame image sequence, multi-view geometric constraints are used to deduplicate, eliminate false detections, and refine coordinates of the initial set of node positions. This results in a high-confidence set of pixel coordinate data for rebar intersections in a unified coordinate system, providing a foundation for topology modeling. The final output coordinate data excludes duplicate and erroneous detection points, exhibiting high positional consistency.

[0022] Preferably, the topology defect detection module includes a spatial relationship modeling unit and a local anomaly analysis unit;

[0023] The spatial relationship modeling unit is used to automatically construct the identified discrete steel bar intersections into nodes of a graph, abstract the steel bars themselves (i.e. the connection relationship between nodes) into edges, and assign the edges with attributes of theoretical spacing and actual visual connection strength, thereby upgrading the two-dimensional image information into a steel bar mesh topology graph containing rich spatial relationships and semantics.

[0024] The local anomaly analysis unit is used to perform deep learning on the constructed rebar mesh topology using a graph convolutional network to perceive the relationship between each node in the graph and its neighborhood, and accurately detect the abnormal length of the edges and the weakening of the connection properties of the edges within a local range. Among them, the abnormal length of the edges within a local range corresponds to uneven spacing of the rebars, and the weakening of the connection properties of the edges corresponds to loosening of the binding points.

[0025] Preferably, the spatial relationship modeling unit performs the following steps:

[0026] Each coordinate point in the output set of pixel coordinate data of steel bar intersections is mapped and defined as an independent node in the graph structure, and its spatial location information is recorded, thus transforming the disordered point cloud data into a structured unit with a clear identity.

[0027] Based on the preset proximity criterion and directional consistency constraint, the defined nodes are automatically connected and reasoned. An edge is established between two adjacent nodes that are on the same straight line. This edge abstractly represents a single physical steel bar segment, realizing the automatic reconstruction of the physical connection relationship of the steel bars, so that the abstract set of nodes forms a topology that reflects the real steel bar network.

[0028] Each established edge is assigned a structured attribute, which includes the actual visual length calculated based on the node coordinates, the theoretical design spacing value of the steel bar segment represented by the edge, and an initialized connection strength coefficient that represents the stability of the connection between nodes. This constructs a complete and attributed steel mesh topology. By adding engineering semantic attributes to the edges, the topology is upgraded from a purely geometric relationship to a digital model that includes design benchmarks and quality assessment indicators.

[0029] Preferably, the local anomaly analysis unit performs the following steps:

[0030] The attributed rebar mesh topology is input into a graph convolutional network. This graph convolutional network learns the local structural patterns and attribute distribution features of each node and its first-order neighborhood edges in the graph by performing message passing and feature aggregation on the nodes and edges of the graph. This enables the model to fully understand the local structural dependencies of the rebar mesh, thereby significantly improving the sensitivity to subtle geometric deviations and abnormal connection states.

[0031] Based on the learned local patterns, the graph convolutional network performs anomaly scoring on the attributes of each edge, specifically identifying two types of anomalies: edge with a spacing anomaly whose actual visual length attribute deviates significantly from its theoretical design spacing attribute, and edge with a weak connection strength coefficient lower than the dynamic adaptive threshold. This enables automated and quantitative diagnosis of two core defects: uneven rebar spacing and loose binding points, significantly reducing the subjectivity and false negative rate of manual identification.

[0032] Based on the identification results of the abnormal spacing edges and weak connection edges, they are mapped back to the original rebar mesh topology and the corresponding construction site image area. This accurately locates the specific defect locations where there are uneven rebar spacing or loose binding points, and outputs the defect type, location coordinates, and deviation quantification value. Through the spatial mapping mechanism, the abstract digital anomalies are directly mapped to specific physical locations, providing clear guidance for rapid on-site location and accurate repair.

[0033] Preferably, the quality time series modeling module performs the following steps:

[0034] The autonomous perception and decision-making module and the topology defect detection module are periodically triggered at preset fixed time intervals to obtain the data for the same construction area at different times t1, t2, ..., t3. n The snapshot of the topology of the steel mesh contains the set of node coordinates and the complete topology structure at that moment, enabling continuous and structured monitoring of construction quality, ensuring the integrity and traceability of data throughout the process, and providing a stable and reliable data foundation for time series analysis.

[0035] The topology snapshots of the steel mesh obtained at different times are spatially aligned based on the unchanging construction area benchmark to ensure that nodes at the same physical location are comparable in different snapshots. The aligned topology snapshots are then organized into a time-series sequence according to the timestamps to eliminate data offset caused by equipment movement and positioning fluctuations. This ensures that the data at different time points are strictly aligned in space, forming a continuous sequence that accurately reflects the dynamic changes in quality.

[0036] From the time-series state sequence, key time-series features characterizing the dynamic evolution of tying quality are extracted. These key time-series features include: displacement trajectories of key nodes, time series of specific side length attributes, and change curves of local connection strength coefficients. The original time-series data is transformed into quantitative features that can be directly used for quality assessment and prediction, revealing the hidden gradual quality change process and supporting the early identification and warning of degradation trends.

[0037] Preferably, the degradation trend early warning module performs the following steps:

[0038] The time-series state sequence containing evolutionary features output by the quality time-series modeling module is used as training data input into the long short-term memory network model. Through training, the model learns the potential laws and periodic patterns of the evolution of various indicators of rebar binding quality (spacing uniformity, connection stability) over time during the construction process, enabling the system to deeply understand the dynamic trend of construction quality changes over time.

[0039] The historical time-series state sequence of the most recent 4 days is input into the trained long short-term memory network model to perform multi-step forward inference and predict the state probability distribution of key quality indicators (average spacing deviation, proportion of weak connection edges) of the steel mesh in the target area after a future construction cycle, so that the system can identify potential risks in advance before quality anomalies actually occur.

[0040] The predicted key quality indicators are compared with the preset quality control thresholds. If the predicted values ​​indicate a risk of degradation exceeding the quality control thresholds, an early warning signal is generated that includes the location of the risk, the type of degradation predicted, and the expected time of occurrence. This effectively supports early intervention and proactive control of quality issues at the construction site.

[0041] Preferably, the intelligent monitoring and intervention module performs the following steps:

[0042] It receives current defect detection results and future quality degradation trend prediction information in real time, integrates and analyzes the two in spatial and temporal dimensions, and generates a panoramic assessment report that comprehensively reflects the current status of construction quality and future risks. After the panoramic assessment report is generated, it realizes the visual integrated display of the current status of construction quality and future risks, which makes it easy for managers to quickly grasp the overall situation.

[0043] Based on the aforementioned panoramic assessment report, combined with a pre-built construction specification knowledge base and rectification case library, specific rectification process suggestions are matched for the identified current defects, and preventive intervention measures are generated for the predicted risk areas. The intervention measures include specific operation locations, methods and standards. The automatic matching of rectification and intervention suggestions improves the standardization of problem handling and ensures that the measures comply with specifications and are targeted.

[0044] The generated rectification and intervention measures suggestions are structured and packaged into executable, visual task sheets, which are then distributed to the corresponding on-site management personnel terminals through the acceptance management center. This drives on-site quality corrections and tracks feedback data after corrections to initiate the next round of acceptance. The distribution and status tracking of task sheets form a closed-loop task system, promoting timely response and effective handling of on-site quality issues.

[0045] This invention provides an intelligent construction quality acceptance system based on image recognition. It has the following beneficial effects:

[0046] (I) This intelligent construction quality acceptance system based on image recognition, through a scene adaptive acquisition unit driven by deep reinforcement learning, can perceive dynamic interferences such as changes in lighting and equipment obstruction at the construction site in real time, autonomously decide the optimal image acquisition parameters and path, get rid of the dependence on fixed markers, actively avoid interference in unstructured scenes, accurately locate key areas, and ensure that clear and complete steel mesh images can be captured in various complex environments.

[0047] (II) This intelligent construction quality acceptance system based on image recognition, combined with multi-scale Retinex enhancement, virtual polarization filtering and other image preprocessing technologies and target detection network, can effectively overcome interference such as uneven lighting and steel bar reflection, and achieve accurate identification and positioning of steel bar intersections. Furthermore, through multi-view geometric constraints, the identification results are transformed into high-precision three-dimensional coordinates and constructed into an attributed steel bar mesh topology map. This not only improves the measurement accuracy of node coordinates, but also upgrades visual information into structured data containing rich spatial relationships and engineering semantics, laying the foundation for analyzing structural defects such as uneven steel bar spacing and loose binding.

[0048] (III) This intelligent construction quality acceptance system based on image recognition utilizes graph convolutional networks to perform deep learning on the topology of the steel mesh. Through the message passing mechanism between nodes and edges, it can accurately learn the statistical laws of attributes such as the length distribution, design deviation, and connection stability of steel segments within a local area. It can automatically identify edges with abnormal spacing and weak connections, and make quantitative judgments based on specifications and dynamic adaptive thresholds. This surpasses the limitations of traditional image processing that only focuses on pixel features, and realizes intelligent understanding of the structural attributes of the steel mesh and defect location, greatly improving the structured, automated, and objective nature of quality assessment. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the workflow of an intelligent construction quality acceptance system based on image recognition according to the present invention.

[0050] Figure 2 This is a schematic diagram of the data flow of an intelligent construction quality acceptance system based on image recognition according to the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: an intelligent construction quality acceptance system based on image recognition, including an acceptance management center, which has the following communication connections:

[0053] The autonomous perception and decision-making module integrates deep reinforcement learning decision-making and image recognition technologies to perform dynamic adaptive perception and robust identification of steel reinforcement targets in complex construction sites. It determines the intersection points of steel reinforcement and their pixel coordinates in the image, enabling the system to break free from dependence on static markers, autonomously adapt to complex environments, and achieve real-time, proactive, and accurate capture of steel reinforcement targets. The autonomous perception and decision-making module includes a scene adaptive acquisition unit and a steel reinforcement node accurate identification unit.

[0054] The scene-adaptive acquisition unit employs a deep Q-network based on reinforcement learning to perceive dynamic interference at the target construction site, including changes in ambient lighting, equipment movement, and occlusion. It autonomously decides the optimal image acquisition parameters and path to obtain the original image sequence of the target construction site, replacing reliance on fixed, easily damaged marker boards. This ensures that key areas usable for analysis are always captured even in unstructured scenes. Based on multi-source sensors mounted on a mobile acquisition device, it acquires real-time environmental state characteristics of the target construction site, including changes in light intensity, the dynamic position and trajectory of large construction equipment, and the degree of temporary occlusion of key areas of the rebar mesh. This provides dynamic data for subsequent environmental perception. Based on the data foundation, environmental state characteristics are used as input to drive a reinforcement learning decision model based on a deep Q-network. Through interactive learning with the environment, the model outputs the current optimal image acquisition control parameters and device movement path planning. The control parameters include focal length, exposure compensation, and shooting angle to achieve adaptive acquisition decisions against dynamic interference. Based on the control parameters and device movement path planning output by the reinforcement learning decision model, the model autonomously adjusts the pose and optical system of the mobile acquisition device to dynamically avoid interference in unstructured scenes and actively locates the acquisition point that can clearly and completely capture the key area of ​​the steel mesh to be inspected, so as to obtain the original image sequence of the target construction site and ensure that the original image sequence that can be used for high-precision analysis is obtained.

[0055] The specific work involves: mobile data acquisition devices deployed at the target construction site integrating multispectral ambient light sensors, an ultra-wideband positioning base station network, and a 3D structured light depth camera to construct an on-site environmental perception system. This system acquires environmental characteristics of the target construction site, including changes in light intensity, the dynamic position and movement trajectory of large construction equipment, and the degree of temporary occlusion of key areas of the steel mesh. Specifically, the multispectral ambient light sensor monitors changes in light intensity in the 500-800nm ​​visible light band in real time at a sampling frequency of 10 times per second, covering a measurement range of 10-10000 lux, ensuring accurate quantification of ambient light intensity under various sunlight and artificial lighting conditions; the ultra-wideband positioning base station network... A 30Hz refresh rate enables centimeter-level precision tracking of all large construction equipment within a 50-meter radius. Simultaneously, a 9-axis inertial measurement unit calculates the equipment's motion trajectory and attitude angles in real time. A 3D structured light depth camera actively projects 940nm infrared structured light patterns, generating depth point cloud data at 15 frames per second to accurately calculate the spatial distance between temporary obstructions and the steel mesh. Environmental state feature data is transmitted to the edge computing unit via a 5G industrial gateway and input into a reinforcement learning decision model based on a deep Q-network. The model's state space contains six environmental parameters: normalized illumination intensity, equipment occlusion index, target area visibility score, and relative equipment position (three-dimensional coordinates). The action space is defined by 17 discrete control commands: focal length adjustment is divided into 7 levels (24mm to 70mm), exposure compensation is divided into 5 levels (-2EV to +2EV), and shooting angle is divided into 5 azimuth angles (-45° to +45°). An ε-greedy strategy is used for exploration, with ε linearly decreasing from 0.9 to 0.1 during the training phase. The discount factor γ is set to 0.95. Through a combination of offline pre-training and online fine-tuning, the model can output the current optimal combination of acquisition parameters within 5 seconds. According to the decision command, the camera attitude is adjusted by a high-precision gimbal control unit, with azimuth angle control accuracy of ±0.5° and pitch angle adjustment range of -30° to +90°. The motorized zoom lens adjusts according to the focal length command. With millisecond-level response and a combination of 3 aperture settings (F2.8-F16) and shutter speeds ranging from 1 / 8000 to 30 seconds, it ensures an image signal-to-noise ratio higher than 35dB in dynamic lighting environments. The mobile chassis moves along the optimal trajectory at a cruising speed of 0.3m / s based on a path planning algorithm, completing positioning at a distance of 1.5-3 meters from the target area with a positioning error of less than 5cm. It also adopts an adaptive frame rate acquisition strategy, using a base frame rate of 2fps under stable lighting conditions, and automatically increasing to 10fps when a sudden change in lighting or device movement is detected, ensuring that at least 5 effective images are captured for each key area, and the image resolution is always maintained above 3840×2160 pixels.

[0056] The rebar node precision identification unit, based on adaptively acquired images of the target construction site, combines deep learning object detection and image enhancement algorithms to overcome recognition interference caused by uneven lighting, rebar reflection, and partial occlusion. It accurately identifies and locates each rebar intersection in the image, determining its pixel coordinates. This transforms the rebar mesh in the dynamic environment into a high-precision set of rebar intersection pixel coordinates. The unit then applies Retinex-based adaptive lighting equalization and a polarization filtering simulation algorithm to suppress uneven lighting and specular noise, enhancing the texture and contour features of the rebar in the image. The processed image exhibits uniform overall brightness, significantly improved contrast between rebar edges and background, facilitating the stability of subsequent algorithms. The image recognition process involves inputting the preprocessed image sequence into a trained deep learning object detection network. Image recognition is performed with the rebar intersections as the central target, and the bounding box of each identified rebar intersection and its precise pixel coordinates in the image coordinate system are output, forming a preliminary set of node locations. The detection network can quickly locate most intersections, initially forming a comprehensive set of node coordinates, laying the foundation for subsequent accurate measurements. Based on multiple image sequences, multi-view geometric constraints are used to deduplicate, eliminate false detections, and refine the coordinates of the preliminary node location set, outputting a high-confidence set of pixel coordinate data for rebar intersections in a unified coordinate system, providing a foundation for topology modeling. The final output coordinate data excludes duplicate and erroneous detection points, exhibiting high positional consistency.

[0057] The specific work involves: after acquiring the original image sequence from the construction site, performing preprocessing enhancement based on a physical imaging model. A multi-scale Retinex algorithm is employed, with a three-level weighting ratio (0.3:0.5:0.2) set for Gaussian wrapping scales of 15, 80, and 250 pixels. Logarithmic domain transformation and color restoration are then performed on the images. Simultaneously, considering the specular reflection characteristics of the steel reinforcement surface, a virtual polarization filter is designed to simulate linearly polarized light illumination conditions in four polarization directions: 0°, 45°, 90°, and 135°. Optimal polarization composite images are generated through Stokes vector calculations, ensuring the images are within an illumination range of 10-10000 lux. The local contrast is improved by no less than 40%, and the peak signal-to-noise ratio of the metallic reflective area is improved by more than 15dB, effectively suppressing overexposed areas caused by changes in the angle of sunlight and welding arc light, ensuring the uniform presentation of the steel bar texture; the pre-processed 4K resolution image is input into a pre-trained deep learning object detection network. The deep learning object detection network adopts an improved YOLOv7 architecture and is optimized for the structural features of the steel bar intersections: the input resolution is adjusted to 1280×1280 pixels, and the anchor frame size is set to three groups of [8, 16], [16, 32], and [32, 64], corresponding to the projection size of the intersections under different viewing angles, and the training... During training, the FocalLoss function was used with α=0.25 and γ=2.0 to balance the ratio of positive and negative samples. The output included the confidence score (threshold set to 0.85), bounding box coordinates, and classification probability for each detection box. In actual deployment, the processing time for each frame was controlled within 120ms. Under complex backgrounds, the recall rate for intersections of φ12-φ32mm steel bars reached over 95%, with a false positive rate below 3%. Based on multiple image sequences (each sequence containing at least 5 frames) acquired by the acquisition device within a working distance of 1.5-3 meters, data refinement was performed using multi-view geometric constraints, and inter-frame pairings were established through ORB feature matching. Based on the relationship, the RANSAC algorithm is used to estimate the fundamental matrix and remove outliers with matching errors greater than 0.8 pixels. For the initially detected node set, its three-dimensional spatial coordinates are calculated by triangulation. The reprojection error threshold is set to 2.5 pixels. On this basis, the DBSCAN clustering algorithm is applied to deduplicate the spatial point cloud. The neighborhood radius ε=10mm and the minimum number of points MinPts=3 are set. The final output node coordinate set has a positional accuracy of ±1.5mm in the global coordinate system, which meets the accuracy requirement of ±2mm for rebar spacing measurement in GB50204-2015 "Code for Acceptance of Construction Quality of Concrete Structures".

[0058] The topology defect detection module is used to abstract the steel mesh into a graph structure and use graph neural networks for analysis, which elevates visual detection to the level of spatial relationship understanding. It quantitatively evaluates and locates defects in the structural attributes of connection relationships and spacing, surpassing the limitations of traditional image processing. The topology defect detection module includes a spatial relationship modeling unit and a local anomaly analysis unit.

[0059] The spatial relationship modeling unit automatically constructs the identified discrete rebar intersections into nodes of a graph. It abstracts the rebars themselves (i.e., the connections between nodes) as edges, assigning them attributes such as theoretical spacing and actual visual connection strength. This elevates the two-dimensional image information into a rebar mesh topology graph containing rich spatial relationships and semantics. Each coordinate point in the output set of rebar intersection pixel coordinates is mapped and defined as an independent node in the graph structure, recording its spatial location information. This transforms unordered point cloud data into structured units with clear identifiers. Based on preset proximity criteria and directional consistency constraints, it performs automated connection reasoning on the defined nodes, connecting nodes located in the same... An edge is established between two adjacent nodes in a straight line. This edge abstractly represents a single physical rebar segment, enabling the automated reconstruction of the physical connection relationship of the rebars. This allows the abstract set of nodes to form a topology that reflects the real rebar network. Each established edge is assigned a structured attribute, which includes the actual visual length calculated based on the node coordinates, the theoretical design spacing value of the rebar segment represented by the edge, and an initialized connection strength coefficient that represents the stability of the connection between nodes. This constructs a complete and attributed rebar network topology. By adding engineering semantic attributes to the edges, the topology is upgraded from a purely geometric relationship to a digital model that includes design benchmarks and quality assessment indicators.

[0060] The specific work involves: after accurately identifying and calculating the 3D coordinates of the rebar intersections, initiating the topology graph construction process. This maps the high-precision 3D coordinate data (accuracy ±1.5mm) of each rebar intersection into an independent node in graph theory. Each node object is created and stored in the system, its core attribute being its 3D spatial position (X, Y, Z) in the global coordinate system of the construction site. This 3D spatial position directly originates from the output of the multi-view geometry refinement module, completing the transformation from an unordered point cloud to a structured node set. The accuracy of the node coordinates directly determines the reliability of all geometric calculations and must ensure that it meets the relevant requirements for rebar position measurement in the "Code for Acceptance of Construction Quality of Concrete Structures" GB50204-2015. Based on the defined node set, automated connection reasoning is performed according to preset geometric rules to construct edges representing the physical connections of the rebars, following a dual constraint criterion: first, the proximity criterion, calculating the Euclidean distance between any two nodes, only when the distance is less than the preset maximum reasonable span of the rebar (based on the rebar spacing in the design drawings). The connection possibility is considered only when the following conditions are met: First, the direction consistency constraint is set. For node pairs that meet the distance threshold, the vector direction formed by them and their neighboring nodes is further analyzed. If the direction angle deviation is within the allowable range, they are determined to be on the same straight line. An undirected edge is established between the two nearest adjacent nodes that meet the above criteria. This edge is abstractly represented as a continuous steel bar segment in the physical world in the topology graph. After each edge is established, it is immediately assigned a set of structured attributes, which upgrades the visual geometric information into topological relationship data rich in engineering semantics. The core attributes include: actual visual length, which is obtained by calculating the three-dimensional Euclidean distance between the two ends of the edge; theoretical design spacing value, which is automatically obtained from the imported design drawings and serves as the benchmark for quality assessment; and connection strength coefficient, which is initialized. Its initial value is estimated based on the visual clarity of the intersection area in the image and the visibility of the binding wire. After all the attributes of the edges are assigned, a complete and attributed steel bar mesh topology graph is generated. This graph structure completely records the geometric layout and theoretical design information of the steel bar mesh.

[0061] The local anomaly analysis unit employs a graph convolutional network (GCNN) to perform deep learning on the constructed rebar mesh topology. It perceives the relationship between each node and its neighborhood, accurately detecting anomalies in edge length and weakened edge connectivity within a local area. Specifically, anomalies in edge length correspond to uneven rebar spacing, and weakened edge connectivity corresponds to loose ties. The attributed rebar mesh topology is input into a GCNN. This GCNN learns the local structural patterns and attribute distribution characteristics of each node and its first-order neighborhood edges through message passing and feature aggregation on the nodes and edges of the graph. This allows the model to fully understand the local structural dependencies of the rebar mesh, significantly improving its sensitivity to subtle geometric deviations and connectivity anomalies. Based on the learned local patterns, the GCNN analyzes each edge... Anomaly scoring is performed on edge attributes, specifically identifying two types of anomalies: spacing anomalies where the actual visual length attribute of the edge deviates significantly from its theoretical design spacing attribute, and weak connection edges where the connection strength coefficient is lower than the dynamic adaptive threshold. This achieves automated and quantitative diagnosis of two core defects: uneven rebar spacing and loose tying points, significantly reducing the subjectivity and missed detection rate of manual identification. Based on the identification results of spacing anomalies and weak connection edges, they are mapped back to the original rebar mesh topology and the corresponding construction site image area, accurately locating the specific defect location of uneven rebar spacing or loose tying points, and outputting the defect type, location coordinates, and deviation quantification value. Through a spatial mapping mechanism, abstract digital anomalies are directly mapped to specific physical locations, providing clear guidance for rapid on-site location and accurate repair.

[0062] The specific work involves inputting the topology diagram of the steel mesh with completed attribute assignments into a graph convolutional network for analysis. The network architecture is designed as a three-layer graph convolutional network, with each layer followed by batch normalization and ReLU activation functions. Anomaly scores for edges are output through fully connected layers. The message passing mechanism of this network is defined as follows: node features are initialized to their three-dimensional coordinates, and edge features are initialized to their three-attribute vectors (actual visual length, theoretical design spacing, and connection strength coefficient). In each convolutional layer, nodes aggregate the features of their connecting edges and adjacent nodes, and edges aggregate the features of their two endpoints and adjacent edges. Through multiple iterations, the network accurately learns... Within the first-order neighborhood of a node, the local distribution patterns and statistical regularities of the length of the rebar segment, design deviation, and connection stability attributes are observed. During training, a labeled topology graph dataset containing various typical construction defects is used. The loss function combines the classification cross-entropy of edges and the regression mean square error. The Adam optimizer is selected, with an initial learning rate of 0.001, and an early stopping strategy is applied to prevent overfitting. Based on the learned local structural patterns, the graph convolutional network performs anomaly scoring on the attributes of each edge in the input topology graph. For the identification of edges with abnormal spacing, the standardized residual between the actual visual length of the edge and its theoretical design spacing is calculated. When the absolute value of the residual is greater than the preset threshold (according to GB50204-2015 standard, it is taken as ±5% of the design spacing, and the absolute value is not less than 2mm), it is judged as a spacing abnormality; for the identification of weak connection edges, the network evaluates its connection strength coefficient and compares it with a dynamic adaptive threshold; the adaptive threshold is calculated and determined in real time based on the distribution of connection strength coefficients of all edges in the current construction section (taken as the 10th percentile). If the coefficient of an edge is lower than the adaptive threshold, it is judged as a weak connection. Each edge marked as abnormal gets a specific score, and the quantitative deviation between its actual value and theoretical value is recorded; After identifying edges with abnormal spacing and weak connections, they are mapped back to the original rebar mesh topology and the corresponding construction site image. By indexing the edges, the global three-dimensional coordinates of the nodes at both ends are retrieved, thereby accurately locating the specific location of the defects in physical space. The defects are identified by their center coordinates (midpoint of the coordinates of the two end nodes). At the same time, the original image sequence on which the topology map was generated is called, and visual evidence of the defect area is automatically extracted and associated. Finally, a structured report containing the defect type, accurate three-dimensional coordinates, quantified deviation value, and associated image identifiers is generated. The report format meets the requirements of construction quality acceptance records.

[0063] The quality time series modeling module is used to organize the data on the location and topological relationship of rebar nodes collected in multiple batches according to the construction time sequence, and to organize the rebar topology map status obtained at different time points in sequence to construct a time series sequence that reflects the quality evolution and describes the dynamic evolution process of binding quality.

[0064] The degradation trend early warning module is used to learn the inherent laws of time sequence using long short-term memory networks, predict the degradation trend of rebar binding quality in the next construction cycle, and generate early warning signals.

[0065] The intelligent monitoring and intervention module is used to integrate defect detection results and quality trend predictions to generate a comprehensive construction quality assessment report. Based on the assessment results, it generates targeted rectification suggestions and intervention measures to guide on-site quality correction.

[0066] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the quality time series modeling module performs the following steps: periodically triggering the autonomous perception decision module and the topology defect detection module at preset fixed time intervals to obtain the same construction area at different times t1, t2, ..., t n The system captures snapshots of the rebar mesh topology, each snapshot containing the set of node coordinates and the complete topology structure at that moment. This enables continuous and structured monitoring of construction quality, ensuring complete and traceable data throughout the process. It provides a stable and reliable data foundation for time-series analysis. The system spatially aligns snapshots of the rebar mesh topology acquired at different times, based on a constant construction area benchmark, ensuring comparability of nodes at the same physical location across different snapshots. Following the chronological order of timestamps, a series of aligned topology snapshots are organized into a time-series state sequence, eliminating data offsets caused by equipment movement and positioning fluctuations. This ensures strict spatial alignment of data at different time points, forming a continuous sequence that accurately reflects dynamic quality changes. From this time-series state sequence, key time-series features characterizing the dynamic evolution of binding quality are extracted. These key features include: key node displacement trajectories, time series of specific side length attributes, and variation curves of local connection strength coefficients. This transforms the original time-series data into quantitative features that can be directly used for quality assessment and prediction, revealing hidden quality gradual changes and supporting early identification and warning of degradation trends.

[0067] The specific work content is as follows: In the implementation, the data acquisition and analysis cycle is automatically triggered according to a preset fixed time interval (set to every 4 hours based on the construction progress and supervision requirements). The autonomous perception and decision-making module drives the mobile acquisition device to reach the preset benchmark position. Within a working distance of 1.5-3 meters, image capture is completed with a resolution of ≥4K and an adaptive frame rate. Subsequently, the topology defect detection module processes the image to generate a snapshot of the current rebar mesh topology. This snapshot of the rebar mesh topology is stored in the form of structured data, completely containing the three-dimensional coordinate set of all rebar intersections at this moment (accuracy ±1.5mm) and the complete topology structure composed of nodes, edges and their attributes (actual length, design spacing, connection strength coefficient); when acquiring t1, t2, ..., t nAfter capturing a series of topology map state snapshots at different times, spatiotemporal alignment processing is performed. Specifically, the global coordinate system of the construction area established during the initial acquisition is used as an invariant reference. Through coordinate transformation using an iterative nearest-point algorithm, the coordinates of nodes in all subsequent snapshots are uniformly registered to this reference coordinate system. This ensures that nodes representing the same physical rebar intersection have consistent coordinate identifiers across different snapshots. The residual error after registration is controlled within ±2mm to meet the accuracy requirements of subsequent trend analysis. After alignment, the series of topology map states are organized into a standard temporal state sequence according to the timestamp order. Based on the aligned temporal state sequence, the dynamic evolution of the binding quality is automatically extracted and quantified. The key temporal characteristics of the process include the displacement trajectories of key nodes, the time series of specific edge length attributes, and the variation curves of local connection strength coefficients. Specifically, the three-dimensional coordinates of key nodes (located in the critical stress region of the structure) at each time point are tracked to form their displacement trajectories, and their cumulative displacement and instantaneous velocity are calculated. The actual visual length attributes of specific edges (stress bars with small design spacing) are monitored to generate a time series of their length attributes over time, and the length difference between adjacent periods is calculated to assess the stability of the spacing. At the same time, the mean value of the connection strength coefficients of all edges within a local area (1 square meter grid) is statistically analyzed, and the local connection strength variation curve is plotted to quantify the gradual process of the overall stability of the binding.

[0068] The degradation trend early warning module performs the following steps: The time-series state sequence containing evolutionary characteristics, output by the quality time-series modeling module, is input as training data into the Long Short-Term Memory (LSTM) network model. Through training, the model learns the potential patterns and cyclical variations in various rebar binding quality indicators (spacing uniformity, connection stability) over time during construction. This allows the system to deeply understand the dynamic trends of construction quality changes over time. The system then inputs the historical time-series state sequence of the last four days into the trained LTM network model for multi-step forward inference, predicting the state probability distribution of key quality indicators (average spacing deviation, weak connection ratio) of the rebar mesh in the target area after one construction cycle. This enables the system to identify potential risks before quality anomalies actually occur. The predicted key quality indicators are compared with preset quality control thresholds. If the predicted values ​​indicate a degradation risk exceeding the quality control threshold, an early warning signal is generated, including the risk location, predicted degradation type, and expected occurrence time. This effectively supports early intervention and proactive control of quality issues at the construction site.

[0069] The specific work involves converting the periodically collected and spatiotemporally aligned rebar mesh topology state sequence into standardized time-series data that can be directly processed by machine learning models. Each time step in this sequence is a set of multi-dimensional feature vectors, including: the average spacing deviation value of the region calculated from the time series of specific edge length attributes, the proportion of weakly connected edges statistically derived from the local connection strength change curve, and the average displacement rate of key nodes. Each feature is organized at a fixed frequency of one data point every 4 hours, forming a continuous time-series sample. During the training phase, historical time-series data covering the entire construction phase (no less than 30 consecutive working days) is used, with 70% of the data serving as the training set, 15% as the validation set, and 15% as the test set. The Long Short-Term Memory (LSTM) network model adopts a three-layer hidden layer structure, with each layer containing 128 LSTM units. The input time step is set to 24 steps (i.e., backtracking 4 days of data), and the output is the key indicator for the next step (i.e., predicting the next 4 hours). During training, mean squared error is used as the loss function, and Adam optimization is employed. The model is trained with an initial learning rate of 0.001 and a learning rate decay strategy. The learning rate is halved when the validation set loss no longer decreases within 10 consecutive epochs. An early stopping mechanism prevents overfitting. Training terminates when the validation set loss shows no improvement within 20 consecutive epochs. The trained model learns the evolution of rebar binding quality under the combined effects of construction load, environmental temperature and humidity, and human disturbance. During model deployment and online application, historical time-series feature data from the most recent 24 time steps (i.e., the most recent 4 days) are input into the trained Long Short-Term Memory (LSTM) network model in real-time using a sliding window. The model performs multi-step forward inference and outputs predicted values ​​and their 90% confidence intervals for two key quality indicators: the average spacing deviation and the proportion of weakly connected edges in the target area rebar mesh for the next construction cycle (4 hours later). The average spacing deviation is calculated according to the allowable deviation of rebar installation position specified in the "Code for Acceptance of Construction Quality of Concrete Structures" GB50204-2015, with the predicted deviation accurate to 0.The prediction of the proportion of weakly connected edges (5mm) is based on the probability distribution of the connection strength coefficient output by the network. It is calculated after being judged by a dynamic threshold. The prediction process is carried out in a containerized environment deployed on an edge server, with a single inference time controlled within 5 seconds to meet the timeliness requirements of real-time construction monitoring. The prediction results are automatically compared with preset three-level quality control thresholds: Level 1 quality control threshold (early warning line) is 80% of the allowable deviation of the specification; reaching this line indicates a degradation trend. Level 2 quality control threshold (warning line) is 100% of the allowable deviation of the specification; reaching this line indicates that the specification threshold has been reached. Level 3 quality control threshold (action line) is 120% of the allowable deviation of the specification; reaching this line indicates that the standard is about to be exceeded or has already been exceeded, requiring immediate intervention. When the key quality indicator value output by the Long Short-Term Memory network model exceeds the preset Level 1 quality control threshold, automatic... The early warning generation process is triggered. The core content of the early warning signal includes: the three-dimensional coordinates of the risk location, the predicted degradation type (spacing uniformity degradation or connection stability degradation), and the estimated occurrence time (4 ± 0.5 hours after the current time). Based on the threshold level reached, the early warning level is dynamically divided into three levels: Attention, Warning, and Severe. All early warning information, corresponding prediction data, historical sequence snapshots, and associated on-site image evidence are structurally encapsulated and pushed in real-time via HTTPS protocol in JSON format to the risk dashboard of the acceptance management center. Simultaneously, it is sent to the mobile terminal applications of relevant supervisors and construction managers. Each early warning information includes a unique traceability code, which can be used for closed-loop verification with subsequent actual inspection results, ensuring that the entire process from prediction calculation to early warning issuance is delayed by no more than 10 seconds, thus allowing sufficient time for on-site preventative intervention measures.

[0070] The intelligent monitoring and intervention module performs the following steps: It receives current defect detection results and future quality degradation trend predictions in real time, integrates and analyzes these two data across spatial and temporal dimensions, and generates a comprehensive panoramic assessment report reflecting the current construction quality status and future risks. After the panoramic assessment report is generated, it provides a visual integrated display of the current construction quality status and future risks, enabling managers to quickly grasp the overall situation. Based on the panoramic assessment report, combined with a pre-built construction specification knowledge base and rectification case library, it matches specific rectification process suggestions for identified current defects and generates preventative intervention measures for predicted risk areas. These intervention measures include specific operational locations, methods, and standards. The automatic matching of rectification and intervention suggestions improves the standardization of problem handling, ensuring that measures comply with specifications and are targeted. The generated rectification and intervention measures are structured and packaged into executable, visual task orders, which are distributed to the corresponding on-site management personnel terminals through the acceptance management center. This drives on-site quality correction and tracks feedback data after correction to initiate the next round of acceptance. The distribution and status tracking of task orders form a closed-loop task system, promoting timely response and effective handling of on-site quality issues.

[0071] The specific tasks are as follows: Real-time reception of current defect detection results from the topology defect detection module and future quality risk prediction information from the degradation trend early warning module. Spatially, the three-dimensional coordinates of the current defect (positioning accuracy ±1.5mm) are overlaid with the three-dimensional coordinate range of the predicted risk area (a spherical area with a radius of 1.5 meters centered on the predicted associated node) to generate a construction area quality heatmap that integrates current problem points and future risk domains. Temporally, the detection timestamp of the current defect is overlaid with the predicted occurrence time of the risk (4±0 seconds after the current time).Alignment is performed over a 5-hour period to construct a timeline view of quality evolution from the present to the near future. Based on this fusion analysis, a structured panoramic assessment report is automatically generated. This panoramic assessment report follows the format requirements of the "Unified Standard for Acceptance of Construction Quality of Building Engineering" GB50300-2013, including: a detailed list of currently confirmed defects (type, coordinates, deviation, and specification clause number) and a list of future quality risk predictions (risk type, area, predicted index value, confidence level, and expected occurrence time). Based on the generated panoramic assessment report, a pre-built construction specification knowledge base and historical rectification case library are invoked to automatically match specific requirements for each identified current defect. The system provides rectification process recommendations. The standard knowledge base is based on the current standards "Code for Acceptance of Construction Quality of Concrete Structures" (GB50204-2015) and "Technical Specification for Mechanical Connection of Reinforcing Steel" (JGJ107), providing standardized rectification process steps, required tools, and acceptance standards for each type of defect. The case library stores effective rectification cases annotated by experts and verified historically. Through similarity matching (based on defect type, reinforcing steel specifications, and working environment), it recommends the optimal practice solution. Simultaneously, for predicted risk areas, it generates preventative intervention suggestions, such as suggesting preventative positioning before concrete pouring for areas predicted to experience spacing degradation. Review and fine-tune; for areas where connection stability is predicted to decline, it is recommended to increase temporary reinforcement points or inspection frequency; all recommendations should clearly specify the operation location (3D coordinates), operation method, and execution standard; the generated rectification and intervention measures recommendations should be structured and packaged into an executable visual task sheet conforming to the ISO19650 standard. The task sheet uses a QR code as a unified identifier and includes: task number, associated defect / risk ID, execution location (screenshot of site plan with 3D coordinates), specific instructions, safety precautions, acceptance standards, and completion deadline (currently, defect rectification requires completion within 4 hours, and preventive intervention is recommended to be deployed 2 hours before the predicted risk occurs). The task order is distributed in real-time to the mobile terminal application of the designated on-site management personnel of the responsible unit via the 5G private network of the acceptance management center, using both push notifications and in-system messages. After receiving the task, the management personnel can view all details and provide feedback on the order status (acceptance, start, completion, etc.) on their terminals. Upon completion of the task, the personnel upload the rectified on-site images and measurement data as required, automatically triggering a targeted re-inspection process for that area. The feedback data is compared and verified with the original defect / predicted data, forming a digital closed loop of initiation-execution-verification. The verification results are also used to update the case library and optimize the early warning model, thus initiating the next round of more precise intelligent monitoring and acceptance cycle.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A construction quality intelligent acceptance system based on image recognition, comprising an acceptance management center, characterized in that, The acceptance management center communication connection includes the following modules: The autonomous perception and decision-making module comprehensively utilizes deep reinforcement learning decision-making and image recognition technologies to adaptively perceive the target construction site and identify rebar targets, determining each rebar intersection point and its pixel coordinates in the image. The autonomous perception and decision-making module includes a scene adaptive acquisition unit and a rebar node accurate identification unit. The scene adaptive acquisition unit is used to use a deep Q-network based on reinforcement learning to perceive the dynamic interference of the target construction site, autonomously decide the optimal image acquisition parameters and path, and obtain the original image sequence of the target construction site. Specifically, it includes: acquiring the environmental state characteristics of the target construction site in real time based on multi-source sensors mounted on the mobile acquisition device. The environmental state characteristics include changes in light intensity, the dynamic position and movement trajectory of large construction equipment, and the degree of temporary occlusion of key areas of the steel mesh. The environmental state features are used as input to drive a reinforcement learning decision model based on a deep Q-network. Through interactive learning with the environment, the model outputs the current optimal image acquisition control parameters and device movement path planning. The control parameters include focal length, exposure compensation, and shooting angle. Based on the control parameters output by the reinforcement learning decision model and the device movement path planning, the pose and optical system of the mobile acquisition device are autonomously adjusted to dynamically avoid interference in unstructured scenes and actively locate to the acquisition point that can completely capture the key area of ​​the steel mesh to be inspected, so as to obtain the original image sequence of the target construction site. The precise identification unit for rebar nodes is used to overcome recognition interference and accurately identify and locate each rebar intersection in the image based on the adaptively acquired target construction site image, combined with deep learning target detection and image enhancement algorithms, to determine its pixel coordinates and transform the rebar mesh in the dynamic environment into a set of rebar intersection pixel coordinate data. The topology defect detection module is used to abstract the steel mesh into a graph structure and use graph neural networks for analysis to quantitatively evaluate and locate defects in the structural properties of connection relationships and spacing. The quality time series modeling module is used to organize the data on the location and topological relationship of rebar nodes collected in multiple batches according to the construction time sequence, and to organize the rebar topology map status obtained at different time points in sequence to construct a time series sequence that reflects the quality evolution and describes the dynamic evolution process of binding quality. The degradation trend early warning module is used to learn the inherent laws of time sequence using long short-term memory networks, predict the degradation trend of rebar binding quality in the next construction cycle, and generate early warning signals. The intelligent monitoring and intervention module is used to integrate defect detection results and quality trend predictions to generate a comprehensive construction quality assessment report. Based on the assessment results, it generates targeted rectification suggestions and intervention measures to guide on-site quality correction.

2. The intelligent construction quality acceptance system based on image recognition according to claim 1, characterized in that: The precise identification unit for rebar nodes performs the following steps: The acquired original image sequence is subjected to adaptive illumination equalization processing based on Retinex theory and a polarization filtering simulation algorithm for metal reflection to suppress uneven illumination and specular noise, and enhance the texture and contour features of steel bars in the image. The preprocessed image sequence is input into a trained deep learning object detection network. Image recognition is performed with the rebar intersection as the central target, and the bounding box of each recognized rebar intersection and its precise pixel coordinates in the image coordinate system are output to form a preliminary set of node positions. Based on a multi-frame image sequence, and utilizing multi-view geometric constraints, the initial set of node positions is deduplicated, false detections are eliminated, and coordinates are refined, outputting a set of pixel coordinate data of rebar intersections in a unified coordinate system.

3. The intelligent construction quality acceptance system based on image recognition according to claim 1, characterized in that: The topology defect detection module includes a spatial relationship modeling unit and a local anomaly analysis unit; The spatial relationship modeling unit is used to automatically construct the identified discrete steel bar intersections into nodes of a graph, abstract the steel bars themselves into edges, and assign the edges with attributes such as theoretical spacing and actual visual connection strength, thereby upgrading the two-dimensional image information into a steel bar mesh topology graph. The local anomaly analysis unit is used to perform deep learning on the constructed rebar mesh topology using a graph convolutional network to perceive the relationship between each node in the graph and its neighborhood, and accurately detect the abnormal length of the edges and the weakening of the connection properties of the edges within a local range. Among them, the abnormal length of the edges within a local range corresponds to uneven spacing of the rebars, and the weakening of the connection properties of the edges corresponds to loosening of the binding points.

4. The intelligent construction quality acceptance system based on image recognition according to claim 3, characterized in that: The spatial relationship modeling unit performs the following steps: Map and define each coordinate point in the output set of pixel coordinate data of rebar intersections into an independent node in the graph structure, and record its spatial location information. Based on the preset proximity criterion and directional consistency constraint, the defined nodes are automatically connected and reasoned, and an edge is established between two adjacent nodes that are on the same straight line. This edge is abstractly represented as a single steel bar segment in physical terms. Each established edge is assigned a structured attribute, which includes the actual visual length calculated based on the node coordinates, the theoretical design spacing value of the steel bar segment represented by the edge, and an initialized connection strength coefficient that represents the stability of the connection between nodes, thereby constructing a complete and attributed steel mesh topology.

5. The intelligent construction quality acceptance system based on image recognition according to claim 4, characterized in that: The local anomaly analysis unit performs the following steps: The attributed steel mesh topology is input into a graph convolutional network. This graph convolutional network learns the local structural patterns and attribute distribution features of each node and its first-order neighborhood edges in the graph by performing message passing and feature aggregation on the nodes and edges of the graph. Based on the learned local patterns, the graph convolutional network performs anomaly scoring on the attributes of each edge, specifically identifying two types of anomalies: edge with a spacing anomaly where the actual visual length attribute of the edge deviates significantly from its theoretically designed spacing attribute, and edge with a weak connection edge where the connection strength coefficient is lower than the dynamic adaptive threshold. Based on the identification results of the abnormal spacing edges and weak connection edges, they are mapped back to the original rebar mesh topology and the corresponding construction site image area to accurately locate the specific defect locations where there are uneven rebar spacing or loose binding points, and output the defect type, location coordinates and deviation quantification value.

6. The intelligent construction quality acceptance system based on image recognition according to claim 3, characterized in that: The quality time series modeling module performs the following steps: The autonomous perception and decision-making module and the topology defect detection module are periodically triggered at preset fixed time intervals to obtain the data for the same construction area at different times t1, t2, ..., t3. n A snapshot of the state of the rebar mesh topology, each snapshot containing the set of node coordinates and the complete topology structure at that moment; The snapshots of the steel mesh topology obtained at different times are spatially aligned based on the unchanging construction area benchmark to ensure that nodes at the same physical location are comparable in different snapshots. The aligned topology states are then organized into a time-series state sequence according to the timestamp order. From the time-series state sequence, key time-series features characterizing the dynamic evolution of the binding quality are extracted. These key time-series features include: the displacement trajectory of key nodes, the time series of specific side length attributes, and the change curve of the local connection strength coefficient.

7. The intelligent construction quality acceptance system based on image recognition according to claim 6, characterized in that: The degradation trend early warning module performs the following steps: The time-series state sequence containing evolutionary features output by the quality time-series modeling module is used as training data input into the long short-term memory network model. Through training, the potential laws and periodic patterns of the evolution of various indicators of rebar binding quality over time during the construction process are learned. The historical time series state sequence of the most recent 4 days is input into the trained long short-term memory network model to perform multi-step forward inference and predict the state probability distribution of the key quality indicators of the steel mesh in the target area after a future construction cycle. The predicted key quality indicators are compared with the preset quality control thresholds. If the predicted values ​​indicate a risk of degradation exceeding the quality control thresholds, an early warning signal is generated that includes the location of the risk, the predicted type of degradation, and the expected time of occurrence.

8. The intelligent construction quality acceptance system based on image recognition according to claim 7, characterized in that: The intelligent monitoring and intervention module performs the following steps: It receives current defect detection results and future quality degradation trend prediction information in real time, integrates and analyzes the two in spatial and temporal dimensions, and generates a panoramic assessment report that comprehensively reflects the current status of construction quality and future risks. Based on the aforementioned panoramic assessment report, combined with a pre-built construction specification knowledge base and rectification case library, specific rectification process suggestions are matched for the identified current defects, and preventive intervention measures are generated for the predicted risk areas. The intervention measures include specific operation locations, methods and standards. The generated rectification and intervention measures suggestions are structured and packaged into an executable, visual task sheet, which is then distributed to the corresponding on-site management personnel's terminals through the acceptance management center. This drives on-site quality corrections and tracks the feedback data after corrections to initiate the next round of acceptance cycle.

Images