A building construction quality analysis method and system based on visual recognition

Abstract

The application discloses a kind of based on visual identification's building construction quality analysis method and system, it is related to the technical field of visual identification, to improve the automation level and precision of building construction quality monitoring, this method includes: acquisition construction image, identify and extract initial feature, geometric feature and texture feature of construction target;Through calculating the brightness difference value and hamming distance of target and associated object or virtual associated object, the fusion similarity is obtained, and is spliced into comprehensive feature vector;Based on comprehensive feature vector, utilize cosine similarity to construct cross-frame similarity matrix;Similarity matrix is optimized to multistage affinity matrix, constructs three-level affinity matrix, and obtains optimal matching pair by optimal matching algorithm;Finally, based on the feature time sequence change of optimal matching pair, quality anomaly is judged.The application realizes accurate tracking and automatic quality monitoring to construction target, effectively improves analysis precision and efficiency.

Description

Technical Field

[0001] This invention relates to the field of visual recognition technology, and in particular to a method and system for analyzing the quality of building construction based on visual recognition. Background Technology

[0002] Construction quality analysis is a crucial step in ensuring the safety, performance, and service life of engineering structures. By conducting real-time monitoring and evaluation of material laying, component installation, and other aspects of the construction process, quality defects can be identified in a timely manner, avoiding rework costs and safety hazards.

[0003] Chinese patent application CN202510323991.5 discloses a method and system for inspecting the construction quality of road engineering based on machine vision. This solution acquires construction images of the road surface, performs grayscale processing, and combines optical flow vector analysis and asphalt particle density calculation to identify abnormal areas on the road surface. Compared with traditional manual inspection, this technology realizes automated evaluation of road construction quality and improves inspection efficiency.

[0004] However, this technology has significant limitations. Its detection is based on pixel-level or region-level static features and instantaneous motion vectors, making it unable to independently identify and track specific construction targets such as steel bars, concrete, and embedded parts across frames. It is also difficult to extract multi-dimensional quality features such as diameter deviation, rust texture, and binding spacing. Furthermore, it does not consider the spatial relationships between different construction targets, making it impossible to construct the temporal changes of features for a single target. This results in insufficient early identification capability for progressive defects. In addition, its quality judgment relies on simple threshold comparisons, which are easily affected by noise in complex construction scenarios. Moreover, it lacks an optimization mechanism for cross-frame target matching relationships, making it difficult to guarantee the continuity and accuracy of tracking. Therefore, the existing technology cannot meet the needs of modern building construction for precise quality control of multiple targets, multiple dimensions, and long time series. There is an urgent need for a new method that can achieve dynamic tracking, correlation analysis, and evolutionary evaluation of specific construction targets. Summary of the Invention

[0005] The technical problem solved by this invention is that existing detection technologies are based on pixel-level or region-level static features and instantaneous motion vectors, which cannot independently identify and track specific construction targets such as steel bars, concrete, and embedded parts across frames. It is also difficult to extract multi-dimensional quality features such as diameter deviation, rust texture, and binding spacing. Furthermore, it does not consider the spatial correlation between different construction targets, which makes it impossible to construct the feature temporal changes of a single target. It also lacks the ability to identify progressive defects in their early stages. In addition, its quality judgment relies on simple threshold comparison, which is easily affected by noise in complex construction scenarios. Moreover, it lacks an optimization mechanism for cross-frame target matching relationships, making it difficult to guarantee the continuity and accuracy of tracking.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a construction quality analysis method based on visual recognition, comprising the following steps:

[0007] Step S1: Collect construction data, identify and extract the initial features, geometric features and texture features of the construction target;

[0008] Step S2: Based on the construction target, calculate the brightness difference value and Hamming distance with the associated object and the virtual associated object respectively, obtain the similarity of the construction target across frames, and construct a similarity matrix by concatenating feature vectors and cosine similarity;

[0009] Step S3: Construct a three-level affinity matrix based on the similarity matrix, and then obtain the optimal matching pair in the three-level affinity matrix through the optimal matching algorithm;

[0010] Step S4: Determine quality anomalies based on the temporal changes of the features of the optimal matching pair.

[0011] Preferably, step S1 includes:

[0012] The construction data includes construction images and coordinates of the construction targets;

[0013] A lightweight target detection algorithm is used to identify construction targets in construction images, extract the initial features of the construction targets, and label the construction targets with target types and target codes;

[0014] The target types include reinforcing bars, concrete, and embedded parts;

[0015] The initial characteristics of the target specifically include:

[0016] The initial characteristics of reinforcing bars include their diameter measurements, the percentage of surface rust area, and the spacing between binding points;

[0017] The initial characteristics of concrete include the average surface gray value, bubble diameter, bubble density, and number of surface cracks;

[0018] The initial characteristics of embedded parts include center coordinate deviation, distance from template edge, and perpendicularity;

[0019] Based on the initial features, further extract the geometric and texture features of the construction target;

[0020] The geometric features include the deviation value of the rebar diameter calculated based on the measured value and the design value of the rebar diameter, the deviation of the lateral spacing calculated based on the spacing between the binding points and the design value, and the deviation of the verticality of the rebar.

[0021] The flatness deviation, crack length, and width are calculated based on the measured and designed values ​​of concrete surface flatness.

[0022] The center coordinate deviation rate, the spacing deviation rate with template edge, and the verticality deviation rate are calculated based on the center coordinate deviation of the embedded part and the design value.

[0023] The texture features include steel bar corrosion texture features extracted using local binary mode and concrete surface texture features extracted using gray-level co-occurrence matrix.

[0024] Principal component analysis is used to reduce the dimensionality of texture features and obtain texture feature vectors.

[0025] Preferably, step S2 includes:

[0026] Extract the associated objects of each construction target in the construction image. The associated objects are the adjacent construction targets within a radius of 50 pixels centered on the construction target.

[0027] Set a threshold for the number of associated objects of a construction target. If the number of associated objects of a construction target is lower than the threshold, the construction target is determined to be a low-degree node.

[0028] For non-low-degree nodes in the construction target, calculate the fusion similarity between the non-low-degree node and construction targets of the same type, as well as between the non-low-degree node and associated objects.

[0029] Preferably, calculating the fusion similarity specifically includes:

[0030] Calculate the absolute values ​​of the grayscale differences between non-low-saturation nodes and construction targets of the same type, and between non-low-saturation nodes and associated objects in the R, G, and B channels, respectively. Add the absolute values ​​to obtain the brightness difference value between non-low-saturation nodes and construction targets of the same type, and the brightness difference value between non-low-saturation nodes and associated objects.

[0031] An adaptive neighborhood window is constructed with non-low-degree nodes as the center. Within the adaptive neighborhood window, the gray value of each pixel is compared with the gray value of the center pixel. If the gray value of a neighboring pixel is less than that of the center pixel, it is recorded as 0; if the gray value of a neighboring pixel is greater than that of the center pixel, it is recorded as 1. A fixed-length binary string is formed, and the binary strings of non-low-degree nodes and construction targets of the same type, as well as the Hamming distance between non-low-degree nodes and associated objects, are obtained.

[0032] The adaptive neighborhood window is a rectangular region surrounding the center of the target level, and its range matches the target type.

[0033] The brightness difference and Hamming distance are normalized and fused to obtain the fusion similarity.

[0034] The mathematical expression for the fusion process is:

[0035] ;

[0036] ;

[0037] in, To integrate similarity, This represents the brightness difference value. For Hamming distance, This is the currently calculated maximum brightness difference value. The maximum Hamming distance calculated so far. To integrate weights, ) represents the information entropy of the adaptive neighborhood window. Let A be the maximum information entropy of the adaptive neighborhood window for all construction targets, and let A be a non-low-degree node. For related objects or similar construction targets;

[0038] The initial features, geometric features, texture feature vectors, average fusion similarity with all associated objects, and average fusion similarity with all construction targets of the same type of non-low-degree node are sequentially concatenated to obtain the comprehensive feature vector of the non-low-degree node.

[0039] Preferably, step S2 further includes:

[0040] A database of typical associated objects for construction targets is constructed based on big data. The database stores typical associated objects and the number of associated objects for various construction targets.

[0041] For low-degree nodes, based on the type of the low-degree node, query and virtually construct the associated objects of the low-degree node from the typical associated object library of the construction target, and obtain the fusion similarity between the low-degree node and the construction target of the same type, as well as between the low-degree node and the virtually constructed associated objects.

[0042] The initial features, geometric features, texture feature vectors, average fusion similarity of all virtual associated objects, and average fusion similarity with all construction targets of the same type of low-degree node are sequentially concatenated to obtain the comprehensive feature vector of the low-degree node.

[0043] Preferably, step S2 further includes:

[0044] The construction images are sorted by time. Two frames with a fixed interval are selected and defined as the front reference frame and the back matching frame in chronological order. The same construction target is extracted from the front reference frame and the back matching frame. The cosine similarity is calculated between the comprehensive vector of the construction target in the front reference frame and the comprehensive vector of the same construction target in the back matching frame. The cosine similarity is used as the basic similarity between the construction target in the front reference frame and the construction target in the back matching frame.

[0045] The comprehensive vector of the frame construction target includes the comprehensive feature vector of low-degree nodes and the comprehensive feature vector of non-low-degree nodes.

[0046] Preferably, step S2 further includes: concatenating the comprehensive feature vectors of all construction targets in the previous reference frame and performing dimensionality reduction processing through 1×1 convolution to generate the feature matrix of the previous reference frame;

[0047] The comprehensive feature vectors of all construction targets in the subsequent matching frame are concatenated and then dimensionality is reduced by 1×1 convolution to generate the feature matrix of the subsequent matching frame.

[0048] The similarity matrix is ​​obtained by multiplying the transpose of the feature matrix of the previous reference frame and the feature matrix of the subsequent matching frame.

[0049] The rows of the similarity matrix correspond to the construction targets of the previous reference frame, sorted by target code, and the columns correspond to the construction targets of the subsequent matching frame, sorted by target code. The elements in the matrix are the basic similarity of the construction target pairs formed by the construction targets in the corresponding rows and columns.

[0050] Preferably, step S3 includes:

[0051] The similarity matrix is ​​used as the first-level affinity matrix. The first-level affinity matrix is ​​dynamically adjusted based on the geometric feature deviation rate to obtain the second-level affinity matrix. The second-level affinity matrix is ​​then dynamically adjusted based on the kurtosis coefficient and neighborhood consistency to obtain the third-level affinity matrix.

[0052] The second-order affinity matrix is ​​the product of the first-order affinity matrix, the attenuation coefficient, and the correction coefficient;

[0053] The attenuation coefficient is 1 divided by the sum of 1 and the power of the natural constant e;

[0054] The power is 5 multiplied by the geometric feature deviation rate of the construction target;

[0055] The geometric feature deviation rate is the difference between the measured value and the designed value of the geometric feature divided by the designed value of the geometric feature.

[0056] The correction factor is 1 minus 0.1 multiplied by the ratio of the brightness difference value of the construction target to the maximum brightness difference value of all construction targets in the current construction image;

[0057] The process of obtaining the third-level affinity matrix specifically includes:

[0058] For each element in the second-level affinity matrix, a 9×9 neighborhood window is constructed corresponding to the construction target. The kurtosis coefficient of the pixel gray value within the 9×9 neighborhood window is obtained. If the kurtosis coefficient is less than 0, the penalty factor is set to 1.3. If the kurtosis coefficient is greater than or equal to 0, the penalty factor is set to 0.7. Based on the penalty factor, the neighborhood consistency correction of the second-level affinity is performed to obtain the third-level affinity matrix.

[0059] The corrected mathematical expression is:

[0060] ;

[0061] in, It is a three-level affinity matrix. It is a second-order affinity matrix. As a penalty factor, A collection of related objects for the construction objective. For the number of associated objects, For construction objectives Its associated objects Corresponding elements in the second-order affinity matrix;

[0062] The Hungarian algorithm is used to perform optimal matching on the three-level affinity matrix to obtain the optimal construction target pair between the previous reference frame and the subsequent matching frame. The affinity value corresponding to the optimal construction target pair is divided by the maximum affinity value in the three-level affinity matrix as the matching confidence.

[0063] Preferably, step S4 includes: when the matching confidence is greater than or equal to the confidence threshold, the construction target pair corresponding in the three-level affinity matrix is ​​determined as a high-confidence matching pair, a unique tracking code is assigned to each high-confidence matching pair for tracking the construction target in continuous video frames, the initial features, geometric features, and texture feature vectors of the construction target corresponding to the tracking code in the continuous frame images are recorded, the initial features, geometric features, and texture feature vectors are compared with the corresponding quality standard thresholds, and if any feature exceeds the quality standard threshold for a consecutive preset number of frames, it is determined that the construction target has a quality abnormality and an abnormality alarm is output.

[0064] A construction quality analysis system based on visual recognition includes a data acquisition module, a similarity module, an optimization module, and a judgment module.

[0065] The acquisition module is used to collect construction data and identify and extract the initial features, geometric features and texture features of the construction target;

[0066] The similarity module is used to calculate the brightness difference value and Hamming distance between the construction target and the associated object and the virtual associated object, respectively, to obtain the similarity of the construction target across frames, and to construct a similarity matrix by concatenating feature vectors and cosine similarity.

[0067] The optimization module constructs a three-level affinity matrix based on the similarity matrix, and then obtains the optimal matching pair in the three-level affinity matrix through the optimal matching algorithm;

[0068] The judgment module is used to judge quality anomalies based on the temporal changes of the features of the optimal matching pair.

[0069] The beneficial effects of this invention are as follows: By constructing a multi-level, context-aware visual recognition method, high-precision, automated, and dynamic monitoring of construction quality is achieved. Its core advantage lies first in the comprehensiveness of feature extraction. The solution not only collects the initial features of the target but also deeply mines the geometric and texture features that reflect key quality aspects. In the cross-frame target matching stage, the solution introduces the concepts of associated objects and virtual associated objects. By fusing the brightness differences and Hamming distances between the target and its neighbors and similar objects, a comprehensive feature vector containing contextual information is constructed. This greatly improves the robustness and accuracy of target recognition in complex scenarios such as object occlusion and similar appearances. By constructing and progressively optimizing the affinity matrix, the solution integrates geometric deviation and neighborhood consistency into the matching process, ensuring high confidence of the final matching pair. This stable and reliable long-term tracking capability enables the system to achieve early detection and accurate warning of quality anomalies through real-time comparison with standard thresholds, thereby significantly improving the efficiency and reliability of construction quality control and reducing the cost and error of manual inspection. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the basic process of a visual recognition-based building construction quality analysis method provided in one embodiment of the present invention. Detailed Implementation

[0071] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0072] Reference Figure 1 As an embodiment of the present invention, a method for analyzing building construction quality based on visual recognition is provided, comprising the following steps:

[0073] Step S1: Collect construction data, identify and extract the initial features, geometric features and texture features of the construction target;

[0074] Step S2: Based on the construction target, calculate the brightness difference value and Hamming distance with the associated object and the virtual associated object respectively, obtain the similarity of the construction target across frames, and construct a similarity matrix by concatenating feature vectors and cosine similarity;

[0075] Step S3: Construct a three-level affinity matrix based on the similarity matrix, and then obtain the optimal matching pair in the three-level affinity matrix through the optimal matching algorithm;

[0076] Step S4: Determine quality anomalies based on the temporal changes of the features of the optimal matching pair.

[0077] This invention provides an overall framework for a visual recognition-based construction quality analysis method. Compared with traditional methods that rely on manual sampling or single image analysis, this method constructs a similarity matrix of construction targets across frames and utilizes multi-level affinity matrix optimization and optimal matching algorithms to achieve accurate tracking of construction targets. This not only solves the problem of target identification and matching under interference such as target occlusion, lighting changes, and viewpoint shifts in construction scenarios, but also lays a reliable data foundation for subsequent quality anomaly judgment based on feature temporal changes. This method realizes the transformation from single-point static detection to full-process dynamic tracking, significantly improving the automation, accuracy, and timeliness of quality analysis.

[0078] Step S1 includes:

[0079] Construction data includes construction images and coordinates of construction targets;

[0080] A lightweight target detection algorithm is used to identify construction targets in construction images, extract the initial features of the construction targets, and label the construction targets with target types and target codes;

[0081] Target types include reinforcing steel, concrete, and embedded parts;

[0082] The initial characteristics of the target specifically include:

[0083] The initial characteristics of reinforcing bars include their diameter measurements, the percentage of surface rust area, and the spacing between binding points;

[0084] The initial characteristics of concrete include the average surface gray value, bubble diameter, bubble density, and number of surface cracks;

[0085] The initial characteristics of embedded parts include center coordinate deviation, distance from template edge, and perpendicularity;

[0086] Based on the initial features, further extract the geometric and texture features of the construction target;

[0087] Geometric features include the deviation of the rebar diameter calculated based on the measured and designed rebar diameter values, the deviation of the lateral spacing calculated based on the tying point spacing and the designed value, and the deviation of the rebar verticality.

[0088] The flatness deviation, crack length, and width are calculated based on the measured and designed values ​​of concrete surface flatness.

[0089] The center coordinate deviation rate, the spacing deviation rate with template edge, and the verticality deviation rate are calculated based on the center coordinate deviation of the embedded part and the design value.

[0090] Texture features include steel rust texture features extracted using local binary mode and concrete surface texture features extracted using gray-level co-occurrence matrix.

[0091] Principal component analysis is used to reduce the dimensionality of texture features and obtain texture feature vectors.

[0092] This invention clarifies that construction data includes images and coordinates, and defines specific and quantifiable initial features, geometric features, and texture features for three core construction targets: steel bars, concrete, and embedded parts. This classification makes feature extraction more targeted and can accurately reflect the key quality attributes of each type of target. By comparing measured values ​​with design values, geometric features such as deviation values ​​and deviation rates are generated, realizing a quantitative assessment of construction quality and providing an objective basis for subsequent automated judgment. Principal component analysis is used to reduce the dimensionality of high-dimensional texture features, which not only preserves key texture information but also reduces the complexity of subsequent calculations and improves the system's operating efficiency.

[0093] In one specific embodiment of the present invention, a drone equipped with a high-definition camera is used to capture construction images. YOLOv5s is used to identify all steel bars in the images, and each steel bar is labeled with the target type and a unique target code. For each steel bar, its diameter is measured by the image, the proportion of its surface rust area is calculated by the color segmentation algorithm, the spacing of its binding points is measured by the corner detection algorithm, the steel bar diameter deviation and lateral spacing deviation are calculated, and the angle between the steel bar and the vertical direction is calculated by fitting a straight line to obtain the verticality deviation. The LBP algorithm is used to extract the rust texture on the surface of the steel bar to obtain a texture feature vector, and then PCA is used for dimensionality reduction to obtain the final texture feature vector.

[0094] For concrete areas, the mean grayscale value of the surface is calculated, bubbles are identified and measured through morphological operations to obtain the average bubble diameter and bubble density, surface cracks are identified through edge detection algorithms, and flatness deviation is obtained by combining laser scanning data with images. Crack length and width are measured, and texture feature energy, contrast, correlation and homogeneity of the concrete surface are extracted using grayscale co-occurrence matrix, and dimensionality reduction is performed by PCA.

[0095] Step S2 includes:

[0096] Extract the associated objects of each construction target in the construction image. The associated objects are the adjacent construction targets within a radius of 50 pixels centered on the construction target.

[0097] Set a threshold for the number of associated objects of a construction target. If the number of associated objects of a construction target is lower than the threshold, the construction target is judged as a low-degree node.

[0098] For non-low-degree nodes in the construction target, calculate the fusion similarity between the non-low-degree node and construction targets of the same type, as well as between the non-low-degree node and related objects.

[0099] By introducing adjacent associated objects, similarity calculation no longer considers individual targets in isolation, but takes into account their local spatial relationships and contextual information. This greatly improves the robustness of target recognition in complex scenes. By setting a threshold to define low-degree nodes, the system can intelligently identify targets that lack sufficient neighborhood information due to occlusion, location at the edge of the image, etc. This provides a prerequisite for subsequently using virtual associated objects to handle these edge cases and avoids matching failures caused by missing information.

[0100] In a specific embodiment of the present invention, the system searches within a preset radius of 50 pixels with the target rebar A as the center and finds other rebars and concrete embedded parts around it. Therefore, the other rebars and embedded parts are defined as associated objects of rebar A. The system sets the threshold for the number of associated objects of rebars to 2. For another rebar located at the edge of the image, the system only finds 1 adjacent rebar in its neighborhood. Since the number of associated objects is lower than the threshold, the rebar is determined to be a low-degree node.

[0101] For a non-low-degree node rebar A, the system will calculate its fusion similarity with targets of the same type, all other rebars in the image, and its associated object rebars and embedded parts. For edge rebars, the system will trigger the subsequent virtual associated object process.

[0102] Calculating fusion similarity specifically includes:

[0103] Calculate the absolute values ​​of the grayscale differences between non-low-saturation nodes and similar construction targets, and between non-low-saturation nodes and associated objects in the R, G, and B channels, respectively. Add the absolute values ​​to obtain the brightness difference value between non-low-saturation nodes and similar construction targets, and the brightness difference value between non-low-saturation nodes and associated objects.

[0104] An adaptive neighborhood window is constructed with non-low-degree nodes as the center. Within the adaptive neighborhood window, the gray value of each pixel is compared with the gray value of the center pixel. If the gray value of a neighboring pixel is less than that of the center pixel, it is recorded as 0; if the gray value of a neighboring pixel is greater than that of the center pixel, it is recorded as 1. A fixed-length binary string is formed, and the binary strings of non-low-degree nodes and construction targets of the same type, as well as the Hamming distance between non-low-degree nodes and associated objects, are obtained.

[0105] The adaptive neighborhood window is a rectangular region surrounding the center of the target level, and its extent matches the target type.

[0106] The brightness difference and Hamming distance are normalized and fused to obtain the fusion similarity.

[0107] The mathematical expression for the fusion process is:

[0108] ;

[0109] ;

[0110] in, To integrate similarity, This represents the brightness difference value. For Hamming distance, This is the currently calculated maximum brightness difference value. The maximum Hamming distance calculated so far. To integrate weights, ) represents the information entropy of the adaptive neighborhood window. Let A be the maximum information entropy of the adaptive neighborhood window for all construction targets, and let A be a non-low-degree node. For related objects or similar construction targets;

[0111] The initial features, geometric features, texture feature vectors, average fusion similarity with all associated objects, and average fusion similarity with all similar construction targets of non-low-degree nodes are sequentially concatenated to obtain the comprehensive feature vector of non-low-degree nodes.

[0112] This invention fuses brightness differences, which reflect global color information, and Hamming distance, which reflects local texture structure. This combination allows the similarity measure to resist slow changes in illumination while capturing subtle texture differences on the surface, making it more discriminative than a single feature. This invention introduces adaptive weights based on information entropy. When the target region has complex texture (i.e., high information entropy), the system relies more on the Hamming distance, which captures details. When the target region has simple texture (i.e., low information entropy), it relies more on brightness differences. This adaptability ensures that similarity calculation maintains optimal performance in different scenarios. The average value of the fused similarity is concatenated with the initial, geometric, and texture features to form a comprehensive feature vector that includes the target's own attributes and neighborhood relationships. This greatly enriches the target's descriptive information and provides high-quality data input for subsequent high-precision cross-frame matching.

[0113] In a specific embodiment of the present invention, the average RGB value is calculated in the corresponding area of ​​the rebar and the associated object. An adaptive neighborhood window is constructed with the center point of the rebar, and the LBP binary string of the window is calculated. The same operation is performed on the associated object to obtain two binary strings, and the Hamming distance is calculated. The information entropy of the neighborhood window is calculated, and the maximum information entropy of all target neighborhood windows in the current frame is used. The weight is calculated based on the ratio of the information entropy of the neighborhood window to the maximum information entropy. The system calculates the fusion similarity between the rebar and all associated objects and takes the average value. At the same time, the average fusion similarity between rebar A and other rebars of the same type is calculated. Finally, the initial feature vector, geometric feature vector, texture feature vector, average fusion similarity with all associated objects, and average fusion similarity with all construction targets of the same type are sequentially concatenated to form a long-dimensional comprehensive feature vector.

[0114] Step S2 also includes:

[0115] A database of typical associated objects for construction targets is constructed based on big data. The database stores typical associated objects and the number of associated objects for various construction targets.

[0116] For low-degree nodes, based on the type of low-degree node, query and virtually construct associated objects of low-degree nodes from the typical associated object library of construction targets, and obtain the fusion similarity between low-degree nodes and construction targets of the same type, as well as between low-degree nodes and the virtually constructed associated objects.

[0117] The initial features, geometric features, texture feature vectors, average fusion similarity of all virtual associated objects, and average fusion similarity with all construction targets of the same type of low-degree node are sequentially concatenated to obtain the comprehensive feature vector of the low-degree node.

[0118] By constructing a typical associated object library for construction targets and virtually constructing associated objects, the problem of missing neighborhood information caused by isolated targets, occlusion, or location at the edge of the image is effectively solved. This enables the system to perform effective feature extraction and matching even when the data is incomplete, greatly enhancing the robustness and applicability of the algorithm in real and complex construction environments. The virtual associated objects are derived from a typical library based on big data, ensuring the rationality and representativeness of their feature distribution. This ensures that the comprehensive feature vectors generated for low-degree nodes are consistent with those for non-low-degree nodes in terms of format and connotation, facilitating the standardization of subsequent processing procedures.

[0119] In one specific embodiment of the present invention, the system queries a typical associated object library of construction targets based on the target type of the rebar. The library is generated based on historical construction image data and includes the typical spatial distribution, texture features and quantity statistics of various targets. The library stores typical structural data of standard rebar cages. Based on the position and posture of the rebar in the image, the system retrieves data from the typical library and virtually generates models of associated objects around it. These virtual objects have standard geometric dimensions and typical texture features. The system calculates the fusion similarity between the rebar and these virtual associated objects. The initial features, geometric features, texture feature vectors of the rebar, the average fusion similarity with all virtual associated objects, and the average fusion similarity with all real targets of the same type are concatenated to obtain the comprehensive feature vector of the rebar.

[0120] Step S2 also includes:

[0121] The construction images are sorted by time. Two frames with a fixed interval are selected and defined as the front reference frame and the back matching frame in chronological order. The same construction target is extracted from the front reference frame and the back matching frame. The cosine similarity is calculated between the composite vector of the construction target in the front reference frame and the composite vector of the same construction target in the back matching frame. The cosine similarity is used as the basic similarity between the construction target in the front reference frame and the construction target in the back matching frame.

[0122] The comprehensive vector of the frame construction target includes the comprehensive feature vector of low-degree nodes and the comprehensive feature vector of non-low-degree nodes.

[0123] By selecting two consecutive frames at a fixed interval and matching the same construction target based on the target code, the static single-frame feature analysis is extended to dynamic temporal analysis. This is a key step in realizing target tracking and quality evolution analysis. Cosine similarity is used to measure the similarity between two comprehensive feature vectors. Cosine similarity is not sensitive to the absolute value of the vector, but focuses more on its direction. This can effectively resist feature scaling caused by changes in shooting distance and angle, thereby obtaining a more stable and accurate basic similarity.

[0124] In a specific embodiment of the present invention, the system sorts the video frames in chronological order, selects a previous reference frame and a subsequent matching frame at 30-frame intervals, and searches for construction targets with the same target code in the two frames. There is a rebar in the previous reference frame and a rebar with the same target code in the subsequent matching frame. The system determines that they are images of the same construction target at different times. The system retrieves the comprehensive feature vectors generated by the rebar in the two frames in step S2. The comprehensive feature vector of the rebar in the previous reference frame is V1, and the comprehensive feature vector of the rebar in the subsequent matching frame is V2. By calculating the cosine similarity between V1 and V2, their basic similarity is obtained.

[0125] Step S2 also includes: concatenating the comprehensive feature vectors of all construction targets in the previous reference frame and performing dimensionality reduction processing through 1×1 convolution to generate the feature matrix of the previous reference frame;

[0126] The comprehensive feature vectors of all construction targets in the subsequent matching frame are concatenated and then dimensionality is reduced by 1×1 convolution to generate the feature matrix of the subsequent matching frame.

[0127] The similarity matrix is ​​obtained by multiplying the feature matrix of the previous reference frame by the transpose of the feature matrix of the subsequent matching frame.

[0128] The rows of the similarity matrix correspond to the construction targets in the previous reference frame, sorted by target code, and the columns correspond to the construction targets in the subsequent matching frame, also sorted by target code. The elements in the matrix represent the basic similarity of the construction target pairs formed by the construction targets in the corresponding row and column. By concatenating the comprehensive feature vectors of all targets into a matrix and using 1×1 convolution for dimensionality reduction, the data dimensionality can be significantly reduced while retaining the main information, thus reducing the computational load of subsequent matrix multiplication operations and improving processing speed. Constructing the similarity matrix unifies the pairwise similarities between all targets in two frames in a structured form, i.e., a matrix, providing a global matching perspective. This provides the necessary data structure foundation for subsequent optimization using affinity matrix and optimal matching algorithms, as well as the Hungarian algorithm for global optimal matching, avoiding the local optima and matching conflicts that may result from matching one by one.

[0129] In one specific embodiment of the present invention, the system concatenates the comprehensive feature vectors of the targets in the previous reference frame into a matrix according to the target code order, processes the matrix using a 1×1 convolution kernel to reduce its dimensionality, and obtains the feature matrix of the previous reference frame; similarly, the system concatenates and reduces the dimensionality of the comprehensive feature vectors of the targets in the subsequent matching frame to obtain the feature matrix of the subsequent matching frame, and calculates the product of the transpose of the feature matrix of the previous reference frame and the feature matrix of the subsequent matching frame to obtain the similarity matrix; the element in the i-th row and j-th column of this matrix represents the basic similarity between the i-th target in the previous reference frame and the j-th target in the subsequent matching frame. The entire matrix intuitively displays all possible matching relationships and their confidence levels between the two frames.

[0130] Step S3 includes:

[0131] The similarity matrix is ​​used as the first-level affinity matrix. The first-level affinity matrix is ​​dynamically adjusted based on the geometric feature deviation rate to obtain the second-level affinity matrix. The second-level affinity matrix is ​​then dynamically adjusted based on the kurtosis coefficient and neighborhood consistency to obtain the third-level affinity matrix.

[0132] The second-order affinity matrix is ​​the product of the first-order affinity matrix, the attenuation coefficient, and the correction coefficient;

[0133] The attenuation coefficient is 1 divided by the sum of 1 and the power of the natural constant e;

[0134] The power is 5 multiplied by the geometric characteristic deviation rate of the construction target;

[0135] The geometric feature deviation rate is the difference between the measured value and the designed value of the geometric feature, divided by the designed value of the geometric feature.

[0136] The correction factor is 1 minus 0.1 multiplied by the ratio of the brightness difference value of the construction target to the maximum brightness difference value of all construction targets in the current construction image;

[0137] The process of obtaining the third-order affinity matrix specifically includes:

[0138] For each element in the second-level affinity matrix, a 9×9 neighborhood window is constructed corresponding to the construction target. The kurtosis coefficient of the pixel gray value within the 9×9 neighborhood window is obtained. If the kurtosis coefficient is less than 0, the penalty factor is set to 1.3. If the kurtosis coefficient is greater than or equal to 0, the penalty factor is set to 0.7. Based on the penalty factor, the neighborhood consistency correction of the second-level affinity is performed to obtain the third-level affinity matrix.

[0139] The corrected mathematical expression is:

[0140] ;

[0141] in, It is a three-level affinity matrix. It is a second-order affinity matrix. As a penalty factor, A set of related objects that form a construction target, wherein the number of related objects in the set does not exceed 5, and the 5 closest related objects are selected. For the number of associated objects, For construction objectives Its associated objects Corresponding elements in the second-order affinity matrix;

[0142] when When the value is 0, skip the neighborhood correction. ;

[0143] The Hungarian algorithm is used to perform optimal matching on the three-level affinity matrix to obtain the optimal construction target pair between the previous reference frame and the subsequent matching frame. The affinity value corresponding to the optimal construction target pair is divided by the maximum affinity value in the three-level affinity matrix as the matching confidence.

[0144] The construction of the second-level affinity matrix innovatively introduces the key construction quality domain knowledge of geometric feature deviation rate into the matching process. By applying attenuation to targets with large geometric deviations, it effectively suppresses the matching weights of targets that are similar in appearance but may have problems in actual construction quality, making the matching results more consistent with engineering reality. The construction of the third-level affinity matrix uses the kurtosis coefficient to judge the sharpness of the texture and combines it with neighborhood consistency for correction. This makes the matching algorithm consider not only the target itself, but also its coordination with the surrounding environment. If the matching relationship of a target is inconsistent with the matching relationship of its neighbors, it will be penalized, thereby effectively avoiding mismatches caused by local occlusion or noise, and significantly improving the accuracy and stability of matching. Finally, the Hungarian algorithm is used to solve the optimized third-level affinity matrix, which can find the one-to-one matching result with the maximum total affinity from a global perspective, avoiding the matching conflicts that may be caused by local search methods such as greedy algorithms, and ensuring the optimality of the matching pair.

[0145] In a specific embodiment of the present invention, the similarity matrix is ​​used as the first-level affinity matrix Λ1. The verticality deviation rate of the target steel bar in the previous reference frame is examined, the attenuation coefficient is calculated, and the correction coefficient is calculated based on the brightness difference ratio between the steel bar and a candidate target in the subsequent matching frame. All elements corresponding to the steel bar in the first-level affinity matrix are updated. This operation is performed on all targets to obtain the second-level affinity matrix Λ2.

[0146] A neighborhood window is constructed around the rebar, the kurtosis coefficient of the pixel gray value is calculated, the neighborhood consistency correction term is calculated, and the affinity between rebar A and the optimal candidate target is updated. The pixel gray value range of the window is 0~255. When calculating, the gray value is first normalized to 0~1.

[0147] Step S4 includes: when the matching confidence is greater than or equal to the confidence threshold, the construction target pair corresponding in the three-level affinity matrix is ​​determined as a high-confidence matching pair. A unique tracking code is assigned to each high-confidence matching pair for tracking the construction target in continuous video frames. The initial features, geometric features, and texture feature vectors of the construction target corresponding to the tracking code in the continuous frame images are recorded. The initial features, geometric features, and texture feature vectors are compared with the corresponding quality standard thresholds. If any feature exceeds the quality standard threshold for a consecutive preset number of frames, it is determined that the construction target has a quality abnormality, and an abnormality alarm is output.

[0148] By assigning unique tracking codes to high-confidence matching pairs, the system can stably track each key construction target in a continuous video stream. This enables the analysis of the dynamic evolution of their quality characteristics, achieving a leap from discrete point detection to continuous process monitoring. By comparing with preset quality standard thresholds, the system can automatically identify the continuous deterioration trend of quality indicators. By using the judgment logic of exceeding the threshold for a consecutive preset number of frames, false alarms caused by instantaneous noise or measurement errors can be effectively filtered out, ensuring the accuracy and reliability of abnormal alarms. The output abnormal alarms include specific tracking codes, locations, and abnormal data, enabling managers to quickly locate problem points, analyze causes, and take corrective measures. This forms a closed-loop management system of automatic monitoring, intelligent diagnosis, timely early warning, and precise handling, significantly improving the level of construction quality management.

[0149] In one specific embodiment of the present invention, a high-confidence matching pair is determined when the matching confidence level is ≥0.7; if the width of the concrete crack exceeds the threshold for two consecutive frames, it is determined to be a quality anomaly.

[0150] A construction quality analysis system based on visual recognition includes a data acquisition module, a similarity module, an optimization module, and a judgment module.

[0151] The data acquisition module is used to collect construction data and identify and extract the initial features, geometric features, and texture features of the construction target.

[0152] The similarity module is used to calculate the brightness difference and Hamming distance between the construction target and the associated object and the virtual associated object, respectively, to obtain the similarity of the construction target across frames, and to construct a similarity matrix by concatenating feature vectors and cosine similarity.

[0153] The optimization module constructs a three-level affinity matrix based on the similarity matrix, and then obtains the optimal matching pair in the three-level affinity matrix through the optimal matching algorithm;

[0154] The judgment module is used to judge quality anomalies based on the temporal changes of the features of the optimal matching pair.

[0155] By constructing a multi-layered, context-aware visual recognition method, high-precision, automated, and dynamic monitoring of construction quality is achieved. Its core advantage lies first in the comprehensiveness of feature extraction. The solution not only collects the initial features of the target but also deeply mines the geometric and texture features that reflect key quality aspects. In the cross-frame target matching stage, the solution introduces the concepts of associated objects and virtual associated objects. By fusing the brightness differences and Hamming distances between the target and its neighbors and similar objects, a comprehensive feature vector containing contextual information is constructed. This greatly improves the robustness and accuracy of target recognition in complex scenarios such as object occlusion and similar appearances. By constructing and progressively optimizing the affinity matrix, the solution incorporates geometric deviation and neighborhood consistency into the matching process, ensuring high confidence of the final matching pair. This stable and reliable long-term tracking capability enables the system to achieve early detection and accurate warning of quality anomalies through real-time comparison with standard thresholds, thereby significantly improving the efficiency and reliability of construction quality control and reducing the cost and error of manual inspection.

[0156] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0157] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A construction quality analysis method based on visual recognition, characterized in that, Includes the following steps: Step S1: Collect construction data, identify and extract the initial features, geometric features and texture features of the construction target; Step S2: Based on the construction target, calculate the brightness difference value and Hamming distance of the associated object and the virtual associated object respectively, obtain the similarity of the construction target across frames, and construct a similarity matrix by concatenating feature vectors and cosine similarity; Step S3: Construct a three-level affinity matrix based on the similarity matrix, and then obtain the optimal matching pair in the three-level affinity matrix through the optimal matching algorithm; Step S4: Determine quality anomalies based on the temporal changes of the features of the optimal matching pair; The process of constructing the virtual associated object specifically includes: A database of typical associated objects for construction targets is constructed based on big data. The database stores typical associated objects and the number of associated objects for various construction targets. By constructing a library of typical related objects for construction targets and then constructing virtual related objects; The process of obtaining the third-level affinity matrix specifically includes: The similarity matrix is ​​used as the first-level affinity matrix. The first-level affinity matrix is ​​dynamically adjusted based on the geometric feature deviation rate to obtain the second-level affinity matrix. The second-level affinity matrix is ​​then dynamically adjusted based on the kurtosis coefficient and neighborhood consistency to obtain the third-level affinity matrix. The second-order affinity matrix is ​​the product of the first-order affinity matrix, the attenuation coefficient, and the correction coefficient; The attenuation coefficient is 1 divided by the sum of 1 and the power of the natural constant e; The power is 5 multiplied by the geometric feature deviation rate of the construction target; The geometric feature deviation rate is the difference between the measured value and the designed value of the geometric feature divided by the designed value of the geometric feature. The correction factor is 1 minus 0.1 multiplied by the ratio of the brightness difference value of the construction target to the maximum brightness difference value of all construction targets in the current construction image; For each element in the second-level affinity matrix, a 9×9 neighborhood window is constructed corresponding to the construction target. The kurtosis coefficient of the pixel gray value within the 9×9 neighborhood window is obtained. If the kurtosis coefficient is less than 0, the penalty factor is set to 1.

3. If the kurtosis coefficient is greater than or equal to 0, the penalty factor is set to 0.

7. Based on the penalty factor, the neighborhood consistency correction of the second-level affinity is performed to obtain the third-level affinity matrix. The corrected mathematical expression is: ； in, It is a three-level affinity matrix. It is a second-order affinity matrix. As a penalty factor, A collection of related objects for the construction objective. For the number of associated objects, For construction objectives Its associated objects The corresponding element in the second-order affinity matrix.

2. The construction quality analysis method based on visual recognition as described in claim 1, characterized in that, Step S1 includes: The construction data includes construction images and coordinates of the construction targets; A lightweight target detection algorithm is used to identify construction targets in construction images, extract the initial features of the construction targets, and label the construction targets with target types and target codes; The target types include reinforcing bars, concrete, and embedded parts; The initial characteristics of the target specifically include: The initial characteristics of reinforcing bars include their diameter measurements, the percentage of surface rust area, and the spacing between binding points; The initial characteristics of concrete include the average surface gray value, bubble diameter, bubble density, and number of surface cracks; The initial characteristics of embedded parts include center coordinate deviation, distance from template edge, and perpendicularity; Based on the initial features, further extract the geometric and texture features of the construction target; The geometric features include the deviation value of the rebar diameter calculated based on the measured value and the design value of the rebar diameter, the deviation of the lateral spacing calculated based on the spacing between the binding points and the design value, and the deviation of the verticality of the rebar. The flatness deviation, crack length, and width are calculated based on the measured and designed values ​​of concrete surface flatness. The center coordinate deviation rate, the spacing deviation rate with template edge, and the verticality deviation rate are calculated based on the center coordinate deviation of the embedded part and the design value. The texture features include steel bar corrosion texture features extracted using local binary mode and concrete surface texture features extracted using gray-level co-occurrence matrix. Principal component analysis is used to reduce the dimensionality of texture features and obtain texture feature vectors.

3. The construction quality analysis method based on visual recognition as described in claim 2, characterized in that, Step S2 includes: Extract the associated objects of each construction target in the construction image. The associated objects are the adjacent construction targets within a radius of 50 pixels centered on the construction target. Set a threshold for the number of associated objects of a construction target. If the number of associated objects of a construction target is lower than the threshold, the construction target is determined to be a low-degree node. For non-low-degree nodes in the construction target, calculate the fusion similarity between the non-low-degree node and construction targets of the same type, as well as between the non-low-degree node and associated objects.

4. The construction quality analysis method based on visual recognition as described in claim 3, characterized in that, Calculating the fusion similarity specifically includes: Calculate the absolute values ​​of the grayscale differences between non-low-saturation nodes and construction targets of the same type, and between non-low-saturation nodes and associated objects in the R, G, and B channels, respectively. Add the absolute values ​​to obtain the brightness difference value between non-low-saturation nodes and construction targets of the same type, and the brightness difference value between non-low-saturation nodes and associated objects. An adaptive neighborhood window is constructed with non-low-degree nodes as the center. Within the adaptive neighborhood window, the gray value of each pixel is compared with the gray value of the center pixel. If the gray value of a neighboring pixel is less than that of the center pixel, it is recorded as 0; if the gray value of a neighboring pixel is greater than that of the center pixel, it is recorded as 1. A fixed-length binary string is formed, and the binary strings of non-low-degree nodes and construction targets of the same type, as well as the Hamming distance between non-low-degree nodes and associated objects, are obtained. The adaptive neighborhood window is a rectangular region surrounding the center of the target level, and its range matches the target type. The brightness difference and Hamming distance are normalized and fused to obtain the fusion similarity. The mathematical expression for the fusion process is: ； ； in, To integrate similarity, This represents the brightness difference value. For Hamming distance, This is the currently calculated maximum brightness difference value. The maximum Hamming distance calculated so far. To integrate weights, ) represents the information entropy of the adaptive neighborhood window. The maximum information entropy of the adaptive neighborhood window for all construction targets. For non-low-degree nodes, For related objects or similar construction targets; The initial features, geometric features, texture feature vectors, average fusion similarity with all associated objects, and average fusion similarity with all construction targets of the same type of non-low-degree node are sequentially concatenated to obtain the comprehensive feature vector of the non-low-degree node.

5. The construction quality analysis method based on visual recognition as described in claim 4, characterized in that, Step S2 further includes: A database of typical associated objects for construction targets is constructed based on big data. The database stores typical associated objects and the number of associated objects for various construction targets. For low-degree nodes, based on the type of the low-degree node, query and virtually construct the associated objects of the low-degree node from the typical associated object library of the construction target, and obtain the fusion similarity between the low-degree node and the construction target of the same type, as well as between the low-degree node and the virtually constructed associated objects. The initial features, geometric features, texture feature vectors, average fusion similarity of all virtual associated objects, and average fusion similarity with all construction targets of the same type of low-degree node are sequentially concatenated to obtain the comprehensive feature vector of the low-degree node.

6. The construction quality analysis method based on visual recognition as described in claim 5, characterized in that, Step S2 further includes: The construction images are sorted by time. Two frames with a fixed interval are selected and defined as the front reference frame and the back matching frame in chronological order. The same construction target is extracted from the front reference frame and the back matching frame. The cosine similarity is calculated between the comprehensive vector of the construction target in the front reference frame and the comprehensive vector of the same construction target in the back matching frame. The cosine similarity is used as the basic similarity between the construction target in the front reference frame and the construction target in the back matching frame. The comprehensive vector of the frame construction target includes the comprehensive feature vector of low-degree nodes and the comprehensive feature vector of non-low-degree nodes.

7. The construction quality analysis method based on visual recognition as described in claim 6, characterized in that, Step S2 further includes: concatenating the comprehensive feature vectors of all construction targets in the previous reference frame and performing dimensionality reduction processing through 1×1 convolution to generate the feature matrix of the previous reference frame; The comprehensive feature vectors of all construction targets in the subsequent matching frame are concatenated and then dimensionality is reduced by 1×1 convolution to generate the feature matrix of the subsequent matching frame. The similarity matrix is ​​obtained by multiplying the transpose of the feature matrix of the previous reference frame and the feature matrix of the subsequent matching frame. The rows of the similarity matrix correspond to the construction targets of the previous reference frame, sorted by target code, and the columns correspond to the construction targets of the subsequent matching frame, sorted by target code. The elements in the matrix are the basic similarity of the construction target pairs formed by the construction targets in the corresponding rows and columns.

8. The construction quality analysis method based on visual recognition as described in claim 7, characterized in that, Step S3 includes: The Hungarian algorithm is used to perform optimal matching on the three-level affinity matrix to obtain the optimal construction target pair between the previous reference frame and the subsequent matching frame. The affinity value corresponding to the optimal construction target pair is divided by the maximum affinity value in the three-level affinity matrix as the matching confidence.

9. The construction quality analysis method based on visual recognition as described in claim 8, characterized in that, Step S4 includes: when the matching confidence is greater than or equal to the confidence threshold, the construction target pair corresponding in the three-level affinity matrix is ​​determined as a high-confidence matching pair, a unique tracking code is assigned to each high-confidence matching pair for tracking the construction target in continuous video frames, the initial features, geometric features and texture feature vectors of the construction target corresponding to the tracking code in continuous frame images are recorded, the initial features, geometric features and texture feature vectors are compared with the corresponding quality standard thresholds, if any feature exceeds the quality standard threshold for a consecutive preset number of frames, it is determined that the construction target has a quality abnormality and an abnormality alarm is output.

10. A construction quality analysis system based on visual recognition, characterized in that, It includes a data collection module, a similarity module, an optimization module, and a judgment module; The acquisition module is used to collect construction data and identify and extract the initial features, geometric features and texture features of the construction target; The similarity module is used to calculate the brightness difference value and Hamming distance of the associated object and the virtual associated object based on the construction target, respectively, to obtain the similarity of the construction target across frames, and to construct a similarity matrix by concatenating feature vectors and cosine similarity. The optimization module constructs a three-level affinity matrix based on the similarity matrix, and then obtains the optimal matching pair in the three-level affinity matrix through the optimal matching algorithm; The judgment module is used to judge quality anomalies based on the temporal changes of the features of the optimal matching pair. The process of constructing the virtual associated object specifically includes: A database of typical associated objects for construction targets is constructed based on big data. The database stores typical associated objects and the number of associated objects for various construction targets. By constructing a library of typical related objects for construction targets and then constructing virtual related objects; The process of obtaining the third-level affinity matrix specifically includes: The similarity matrix is ​​used as the first-level affinity matrix. The first-level affinity matrix is ​​dynamically adjusted based on the geometric feature deviation rate to obtain the second-level affinity matrix. The second-level affinity matrix is ​​then dynamically adjusted based on the kurtosis coefficient and neighborhood consistency to obtain the third-level affinity matrix. The second-order affinity matrix is ​​the product of the first-order affinity matrix, the attenuation coefficient, and the correction coefficient; The attenuation coefficient is 1 divided by the sum of 1 and the power of the natural constant e; The power is 5 multiplied by the geometric feature deviation rate of the construction target; The geometric feature deviation rate is the difference between the measured value and the designed value of the geometric feature divided by the designed value of the geometric feature. The correction factor is 1 minus 0.1 multiplied by the ratio of the brightness difference value of the construction target to the maximum brightness difference value of all construction targets in the current construction image; For each element in the second-level affinity matrix, a 9×9 neighborhood window is constructed corresponding to the construction target. The kurtosis coefficient of the pixel gray value within the 9×9 neighborhood window is obtained. If the kurtosis coefficient is less than 0, the penalty factor is set to 1.

3. If the kurtosis coefficient is greater than or equal to 0, the penalty factor is set to 0.

7. Based on the penalty factor, the neighborhood consistency correction of the second-level affinity is performed to obtain the third-level affinity matrix. The corrected mathematical expression is: ； in, It is a three-level affinity matrix. It is a second-order affinity matrix. As a penalty factor, A collection of related objects for the construction objective. For the number of associated objects, For construction objectives Its associated objects The corresponding element in the second-order affinity matrix.

Images