Construction site steel bar stacking form intelligent identification method and system

Abstract

The application discloses a construction site steel bar stacking form intelligent identification method and system, specifically relates to the field of image processing, and is used for solving the steel bar stacking identification error problem. The steel bar image and depth data under uniform illumination conditions are acquired by using a ring-shaped multi-view lens array, the rust texture and color information on the surface of the steel bar are combined, the real boundary and spatial position of the steel bar are accurately identified, the gap size between the steel bars and the surface rust reflection degree are analyzed, and a structure relationship network of the steel bar stacking form is constructed. The stability and stacking risk position between different steel bars are further evaluated by an intelligent algorithm, the steel bar identification result is more accurate and reliable, the misjudgment problem caused by traditional methods under conditions such as rust, reflection and stacking complexity is effectively avoided, and stable data support is provided for automatic inventory and mechanical grabbing of the construction site.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, more particularly, the present application relates to a construction site steel bar stacking form intelligent identification method and system. BACKGROUND

[0002] In the construction site, steel bars from different batches are often mixed and stacked together. Some of the steel bars have relatively new surfaces and bright colors, while some have obvious rust and dark colors. Due to the influence of the sun's angle and time change, the bright places in the picture are brighter, and the dark places are darker. The color difference of the steel bar surface and the interlaced light and shadow are mixed together, making the image show obvious color inconsistency.

[0003] However, the current commonly used identification method mainly distinguishes the steel bar boundary through color. In the face of such color mixed situation, it is easy to misjudge, either dividing a steel bar into two segments or identifying several steel bars as a whole. Since the rust degree and color difference will affect the final identification result, relying only on color distinction is unreliable, and the identification result is prone to error. This kind of problem needs to analyze the changes of rust texture and color characteristics, extract these information in advance as auxiliary judgment basis, and help the identification model to find the real boundary and quantity of the steel bar more accurately.

[0004] In order to solve the above problems, a technical scheme is provided. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a construction site steel bar stacking form intelligent identification method and system. By using a ring-shaped multi-view lens array to obtain steel bar images and depth data under uniform lighting conditions, combining the rust texture and color information of the steel bar surface, the real boundary and spatial position of the steel bar are accurately identified, and the gap size between the steel bars and the surface rust reflection degree are analyzed to construct the structural relationship network of the steel bar stacking form. Through intelligent algorithm, the stability and stacking risk position between different steel bars are further evaluated, so that the steel bar identification result is more accurate and reliable, effectively avoiding the misjudgment problem caused by traditional methods in the case of rust, reflection and stacking complexity, providing stable data support for automatic inventory and mechanical grabbing in the construction site, to solve the problems proposed in the above background technology.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] The construction site steel bar stacking form intelligent identification method comprises the following steps:

[0008] S1: sequentially trigger the polarization flash of the ring array lens, collect multi-angle images and depth point cloud, and output the aligned light balanced image and depth point cloud;

[0009] S2: Texture self-supervised encoding is performed on the image, a rust texture matrix and a spectral shift field are generated in combination with a color gamut histogram, and a rust index field is fused;

[0010] S3: The rust index field is projected and fused with the depth point cloud, a continuous edge band of the steel bar is extracted through a cross-scale edge texture segmentation network, the edge band breakage is repaired using the spectral shift field, and a complete boundary set is output;

[0011] S4: The complete boundary set is used to reconstruct the center curve, the distance between adjacent curves is calculated to generate a sequence entropy of the void, and the micro-rust reflection divergence is calculated according to the reflection direction variance of the multi-angle polarization image;

[0012] S5: A graph structure is constructed with the center curve as the node and the void sequence entropy difference value as the edge weight, the micro-rust reflection divergence is input into the graph attention encoder to obtain a coupling coefficient, and a potential collapse risk area is marked;

[0013] S6: A statistical information report is generated in combination with the center curve and the risk area, and is pushed to a dispatch terminal.

[0014] In a preferred embodiment, step S1 includes the following:

[0015] The illumination balanced image and the depth point cloud are collected by a ring array lens and a polarization flash technology, the ring array lens is composed of high-resolution cameras uniformly distributed around the steel bar stacking area and is equipped with a polarization filter, the polarization flash source corresponds to the high-resolution camera and is triggered in sequence, the flash intensity is adjusted according to the ambient light, each high-resolution camera collects a two-dimensional image and generates an illumination balanced image through gray histogram equalization and color temperature correction, a depth sensor collects depth information and generates a filtered depth point cloud through voxel filtering, a spatial transformation matrix is used to project the filtered depth point cloud to the high-resolution camera image plane, the illumination balanced image and the depth point cloud are aligned, and the illumination balanced image and the filtered depth point cloud are output.

[0016] In a preferred embodiment, step S2 includes the following:

[0017] The texture self-supervised encoding is performed on the illumination balanced image, a convolutional neural network is trained using self-supervised learning, the network parameters are optimized through contrastive learning, the convolutional neural network learns the texture difference between the rust area and the non-rust area, and a texture feature map is generated.

[0018] In a preferred embodiment, step S2 further includes the following:

[0019] The light balance image is converted from an RGB color space to an HSV color space, histograms of hue channels, saturation channels and brightness channels are calculated respectively to generate a color gamut histogram; a texture feature map and the color gamut histogram are fused to generate a corrosion texture matrix; the light balance image and the reference image are subjected to two-dimensional Fourier transform to calculate a phase shift in a frequency domain and generate a spectral shift field; and the corrosion texture matrix and the spectral shift field are fused to generate a corrosion index field.

[0020] In a preferred embodiment, step S3 comprises the following:

[0021] The corrosion index field is fused with the depth point cloud projection to generate a fused feature map; the fused feature map is input into a cross-scale edge texture segmentation network to generate a multi-scale feature map and perform edge detection to generate an edge probability map; the edge probability map is subjected to threshold segmentation to generate a binary edge map; the spectral shift field is used to repair a broken area in the binary edge map to generate a repaired edge map; and the repaired edge map is subjected to connected domain analysis to extract a continuous boundary of the steel bar and perform smoothing processing to generate a complete boundary set.

[0022] In a preferred embodiment, step S4 comprises the following:

[0023] Based on the complete boundary set, a center curve is reconstructed through skeletonization processing and B-spline curve interpolation; the center curve is sorted by layers according to an average height, a horizontal distance between adjacent center curves in each layer is calculated to generate a distance sequence and calculate an information entropy to represent a layering density; a reflection value sequence in a multi-angle polarization image is extracted for a same point on the surface of each steel bar, an angular divergence of the reflection value sequence is calculated and fused with the corrosion index field to generate a micro-rust reflection divergence.

[0024] In a preferred embodiment, step S4 further comprises the following:

[0025] The complete boundary set is used to generate the center curve, the center curve is used to calculate the layering density, and the layering density and the reflection value sequence are used to generate the micro-rust reflection divergence.

[0026] In a preferred embodiment, step S5 comprises the following:

[0027] The center curve is defined as a node in the graph, and edges are established between the nodes based on spatial proximity to construct a graph structure; the absolute difference of the gap sequence entropy between adjacent nodes is calculated as the edge weight; the micro-rust reflection divergence is input as the initial feature vector of the node into the graph attention encoder, and the node feature is updated through the multi-head attention mechanism; the sum of the inner product of the adjacent node features and the edge weight is mapped by the Sigmoid function to obtain the matching coupling coefficient, and then the comprehensive risk index of each node is calculated and compared with the preset risk threshold, and the steel bar area corresponding to the node whose comprehensive risk index exceeds the preset risk threshold is marked as the potential collapse risk area.

[0028] In a preferred embodiment, step S6 comprises the following:

[0029] Based on the center curve set, the total number of steel bars is calculated, the diameter of each steel bar is calculated through the boundary point set, and a specification distribution table is generated, which contains the number of steel bars in each diameter range and the proportion of the total number of steel bars; the risk level of each steel bar is determined through the spatial attribution relationship between the key nodes of the center curve and the potential collapse risk area, the number of high-risk steel bars and the proportion of the total number of steel bars, the number of low-risk steel bars and the proportion of the total number of steel bars are counted, and a stability statistical table is generated; the total number of steel bars, the specification distribution table and the stability statistical table are integrated into a statistical information report, which is transmitted to the dispatching terminal through a wireless communication module.

[0030] The construction site steel bar stacking form intelligent identification system comprises:

[0031] The illumination depth acquisition module sequentially triggers the polarization flash through the annular array lens, acquires multi-angle images and depth point clouds, and outputs the aligned illumination balanced images and depth point clouds;

[0032] The corrosion feature fusion module: texture self-supervised encoding is performed on the image, a corrosion texture matrix and a spectral shift field are generated by combining a color gamut histogram, and a corrosion index field is fused;

[0033] The edge contour extraction module: the corrosion index field and the depth point cloud are projected and fused, the continuous edge band of the steel bar is extracted through the cross-scale edge texture segmentation network, the spectral shift field is used to repair the edge band breakage, and a complete boundary set is output;

[0034] The structural material quantification module: the center curve is reconstructed based on the complete boundary set, the distance between adjacent curves is calculated and the layer gap sequence entropy is generated, and the micro-rust reflection divergence is calculated according to the reflection direction variance of the multi-angle polarization image;

[0035] The risk area identification module: the center curve is taken as a node, and the gap sequence entropy difference is taken as an edge weight to construct a graph structure, the micro-rust reflection divergence is input into the graph attention encoder to obtain a matching coupling coefficient, and a potential collapse risk area is marked;

[0036] Statistical information pushing module: generate statistical information report combining central curve and risk area, and push to dispatch terminal for inventory and grabbing planning.

[0037] The technical effects and advantages of the construction site steel bar stacking form intelligent recognition method and system of the present application are as follows:

[0038] The present application effectively avoids the boundary misjudgment problem caused by the difference in corrosion degree and reflection interference by constructing uniform view field under the basis data of illumination balanced image and depth point cloud, combining corrosion texture mapping and spectral shift analysis. On this basis, the edge texture linkage segmentation method is used to accurately extract the continuous contour of the steel bar, and the graph structure based on the central curve is introduced to model the stacking gap and surface reflection characteristics in the form of node attributes and edge weights. Through the graph attention mechanism, the joint influence of structural relationship and material change is excavated, and the coupling coefficient is generated to determine the position of the potential sliding or collapse risk in the stacking area. This method realizes the multi-dimensional cross modeling of form stability and surface state, significantly improves the stability and accuracy of steel bar recognition under complex illumination and corrosion interference, and provides a complete structure and risk-sensitive recognition result for subsequent quantity statistics and grabbing path planning, with good engineering adaptability and intelligent processing capability. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The standard flowchart of the construction site steel bar stacking form intelligent recognition method of the present application is shown.

[0040] Figure 2 The standard flowchart of the construction site steel bar stacking form intelligent recognition system of the present application is shown. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0042] Embodiment 1: Figure 1 The construction site steel bar stacking form intelligent recognition method of the present application is given, including:

[0043] S1: sequentially trigger the polarization flash of the ring array lens, collect multi-angle images and depth point clouds, and output the aligned illumination balanced image and depth point cloud.

[0044] S2: texture self-supervised encoding is performed on the image, corrosion texture matrix and spectral shift field are generated combining the color gamut histogram, and the corrosion index field is obtained by fusion.

[0045] S3: Fuse the rust index field with the depth point cloud projection, extract the reinforcement continuous edge band through the cross-scale edge texture segmentation network, and repair the edge band breakage with the spectral shift field to output the complete boundary set.

[0046] S4: Reconstruct the center curve based on the complete boundary set, calculate the distance between adjacent curves and generate the sequence entropy of the pile space, and calculate the micro-rust reflection divergence according to the reflection direction variance of the multi-angle polarization image.

[0047] S5: Construct a graph structure with the center curve as the node and the difference value of the sequence entropy of the pile space as the edge weight, input the graph attention encoder with the micro-rust reflection divergence to obtain the coupling coefficient, and mark the potential collapse risk area.

[0048] S6: Generate a statistical information report combining the center curve and the risk area, and push it to the dispatch terminal for inventory and grabbing planning.

[0049] In the construction site, steel reinforcement is the core material of the building structure, and the accurate identification of its stacking form is directly related to the construction safety and efficiency. However, in the actual scene, the steel reinforcement comes from various sources, and the surface state differs significantly: some steel reinforcement is coated with an epoxy layer, reflecting cold light; some steel reinforcement generates a reddish-brown rust film due to long-term exposure. In addition, the light and dark textures in the picture are interlaced due to the alternation of day and night light, forming a complex color distribution. These factors make it difficult for traditional tone segmentation methods to effectively distinguish between rust depth and stacking shadows, and the identification results often have number statistics deviation and posture fitting errors. To solve this problem, the present application proposes an intelligent identification method for the stacking form of steel reinforcement in the construction site, which realizes accurate identification and risk assessment of the stacking form of steel reinforcement through multi-step image acquisition and processing, combined with depth information and texture analysis. Step S1 is the first step of the whole method, which aims to use a ring array lens and polarization flash technology to collect light balanced images and depth point clouds, providing high-quality basic data for subsequent rust texture coding and boundary extraction.

[0050] Step S1 realizes the alignment of light balanced images and depth point clouds through multi-angle visual acquisition and light control, combined with image processing and depth data filtering. The specific calculation logic includes: projecting the depth point cloud to the image plane using the spatial transformation matrix, adjusting the image brightness and tone through gray histogram equalization and white balance, and filtering the depth data using the average value of the voxel grid.

[0051] Step S1 includes the following content:

[0052] Arrangement and parameter calibration of ring array lens:

[0053] The ring array lens is composed of multiple high-resolution cameras evenly distributed around the steel reinforcement stacking area, forming a closed ring structure. Each camera covers a part of the steel reinforcement stacking area, and there is a certain overlap between the fields of view of adjacent cameras to ensure complete capture of visual information from multiple angles. Each camera is equipped with a polarized filter to filter out non-polarized light interference and reduce the reflection effect on the steel reinforcement surface. After the arrangement is completed, a calibration board is used to determine the internal and external parameters of each camera: the internal parameters include focal length and principal point coordinates, and the external parameters include position and orientation relative to the unified coordinate system. According to these parameters, a spatial transformation matrix is generated to align the image with the depth point cloud in the future.

[0054] Sequential triggering of polarized flash and light control:

[0055] The polarized flash source is installed near the ring array lens, with its polarization direction consistent with the polarization filter of the camera, ensuring that the light reflected on the steel reinforcement surface reduces the impact of high light areas. The flash source is triggered sequentially: for multiple cameras, the flash source is activated at different time points, and each time only illuminates the shooting area of the corresponding camera, avoiding interference caused by multiple light sources shining at the same time. The intensity of the flash source is dynamically adjusted according to the intensity of the ambient light: by comparing the intensity of the ambient light with the target brightness range, calculating the required compensation light intensity value and applying it to the flash source, the brightness of the collected image is kept balanced.

[0056] Multi-angle image acquisition and light balance processing:

[0057] Each camera acquires a two-dimensional image when the corresponding polarized flash is triggered. Due to the rust and coating differences on the surface of the steel reinforcement, the original image may have local over-bright or over-dark areas. To solve this problem, a light balance processing method is used: first, calculate the gray level histogram of each image to count the pixel distribution of each gray level; then, use the cumulative distribution function to adjust the gray value to make the gray distribution uniform, generating a light balanced image. After that, apply the color temperature correction technology to adjust the red, green and blue channel values through the white balance algorithm: according to the predefined reference white point in the image, calculate the gain coefficient of each channel and apply it to the pixel value, ensuring that the color of the steel reinforcement surface is consistent in different images.

[0058] Depth point cloud acquisition and preprocessing:

[0059] The depth information is collected by an independent depth sensor (such as a laser scanner), generating raw depth point cloud data containing three-dimensional coordinates. To eliminate noise, a voxel filtering method is used: the raw depth point cloud is divided into multiple small voxel grids, the average depth value of all points in each grid is calculated, and the average value is used to replace the original depth value in the grid to generate the filtered depth point cloud. This process reduces the noise of the depth data and improves the accuracy of the data.

[0060] Spatial alignment of illumination-balanced images and depth point cloud:

[0061] To ensure that the illumination-balanced images and the depth point cloud correspond in space, spatial alignment processing is performed: using the spatial transformation matrix of camera calibration and the external parameters of the depth sensor, the filtered depth point cloud is projected onto the image plane of each camera. The projection process is achieved by calculation: according to the spatial transformation matrix, the three-dimensional coordinates of each point in the depth point cloud are converted into two-dimensional pixel coordinates on the corresponding image plane, generating a depth map aligned with the illumination-balanced image. After alignment, the spatial consistency of the image and the depth map is verified to ensure that the color information and depth information of each pixel correspond.

[0062] Data output:

[0063] After the above processing is completed, two types of data are output: one is a set of multi-angle illumination-balanced images, which are two-dimensional images after illumination balancing and color temperature correction; the other is the filtered depth point cloud that is spatially aligned with these images. These data contain clear visual information and accurate three-dimensional structure information for subsequent analysis.

[0064] Step S1 uses a ring array lens and polarized flash technology, combined with illumination balancing and spatial alignment, to solve the problem of complex lighting and variable surface state in the reinforcement stacking area. The ring array lens achieves multi-angle visual coverage, ensuring that the reinforcement stacking area is collected without dead angles; the polarized flash suppresses the interference of reinforcement surface reflection, improving image quality; the illumination balancing process eliminates uneven brightness caused by rust and coating differences, improving image consistency; and the filtering and alignment of the depth point cloud provide accurate three-dimensional structure information.

[0065] Step S1 collects and outputs the aligned light-balanced images and depth point clouds by using the ring array lens and polarized flash technology, laying a data foundation for subsequent processing. The light-balanced images provide uniform brightness visual information, and the depth point clouds reflect the spatial distribution of the steel bars. However, due to the possible existence of high-reflective cold light of the newly painted epoxy layer or the red-brown rust film formed by long-term exposure on the surface of the steel bars on the construction site, and the interlaced bright and dark textures caused by the change of light and darkness, it is difficult to effectively distinguish the corrosion depth and stacking shadows only relying on the light-balanced images and depth point clouds, which is easy to cause boundary misjudgment and recognition error. To solve this problem, step S2 introduces texture self-supervised encoding and spectral analysis technology, aiming to accurately capture the corrosion information of the steel bar surface from the texture and spectral feature levels, and provide reliable basis for subsequent boundary extraction and risk assessment.

[0066] Step S2 includes the following contents:

[0067] Texture self-supervised encoding:

[0068] The light-balanced images are used as input data to train the convolutional neural network through self-supervised learning. During training, the light-balanced images are compared for learning, that is, by comparing the similarities and differences between different image regions, the network parameters are optimized to enable the network to distinguish the texture features of the corrosion area and the non-corrosion area.

[0069] Specifically, the network is based on unlabeled data to automatically learn the texture patterns in the images, such as the difference between the roughness of the corrosion area and the smoothness of the non-corrosion area. After training, the light-balanced images are input into the optimized convolutional neural network, and the network extracts features layer by layer to finally output the texture feature map. The texture feature map is a multi-dimensional data structure that records the texture information of each region in the image. The processing idea is to reduce the dependence on manually labeled data through self-supervised learning, use unlabeled data to mine texture differences, and improve the universality and accuracy of feature extraction.

[0070] Color gamut histogram analysis:

[0071] The light-balanced images are converted from the RGB color space to the HSV color space, which includes hue, saturation, and brightness channels. After conversion, the data of the hue channel, the saturation channel, and the brightness channel are respectively counted, and the histogram of each channel is calculated. The histogram divides the pixel values of each channel into multiple intervals, counts the number of pixels in each interval, and generates a quantitative representation of color distribution.

[0072] The hue histogram reflects the distribution of color categories of the image, the saturation histogram reflects the distribution of color brightness, and the lightness histogram reflects the distribution of brightness. These histograms together constitute the color gamut feature, providing difference information of the rusted area and the non-rusted area in color. The HSV color space separates color and brightness characteristics, quantitatively analyzes the color distribution of the image, and provides color support for the generation of the rust texture matrix, enhancing the comprehensiveness of the rust feature.

[0073] Rust texture matrix generation:

[0074] The texture feature map and the color gamut histogram information are fused. The specific steps are as follows: first, the feature data of the texture feature map and the color gamut histogram are weighted and summed, for example, the weight coefficient of the texture feature map can be set to 0.6, and the weight coefficient of the color gamut histogram can be set to 0.4. When weighted and summed, the value at each position in the texture feature map is multiplied by 0.6, the value at the corresponding position in the color gamut histogram is multiplied by 0.4, and then the two are added to obtain the fused value, forming the rust texture matrix. The rust texture matrix is a two-dimensional data structure that integrates the spatial texture information of the texture feature map and the color information of the color gamut histogram, and can more accurately represent the characteristics of the rusted area.

[0075] Through the weighted combination of texture and color features, the importance of the two types of information is balanced, and a matrix reflecting the comprehensive characteristics of rust is generated, providing multi-dimensional data support for the generation of the subsequent rust index field.

[0076] Spectral shift field calculation:

[0077] The illumination balanced image and the reference image are subjected to two-dimensional Fourier transform, and are converted from the spatial domain to the frequency domain. The reference image is a rust-free steel bar image, which serves as a comparison benchmark. In the frequency domain, the phase difference between the illumination balanced image and the reference image is calculated, that is, by comparing the phase information of the frequency domain data of the two images, the distribution of the phase shift is determined. The phase shift reflects the difference in frequency components between the two images, and a spectral shift field is generated. The spectral shift field is a two-dimensional data structure that records the frequency domain change characteristics caused by rust.

[0078] The Fourier transform is used to capture the frequency domain feature difference between the illumination balanced image and the reference image, and to quantify the influence of rust on the image spectrum, providing spectral level information support for the generation of the rust index field.

[0079] Rust index field generation:

[0080] The processing procedure of the rust index field generation is to perform nonlinear fusion of the rust texture matrix and the spectral shift field. The specific steps are as follows: first, weighted sum of the rust texture matrix and the spectral shift field is performed, for example, the weight coefficients of the two can be set to 0.5, that is, the value of each position in the rust texture matrix is multiplied by 0.5, the value of the corresponding position in the spectral shift field is multiplied by 0.5, and then added to obtain a preliminary fusion result. Then, the preliminary fusion result is input into the Sigmoid function for nonlinear mapping, the Sigmoid function converts the input value into an output value between 0 and 1, and generates the rust index field.

[0081] The rust index field is a two-dimensional data structure, whose value range is between 0 and 1, representing the degree of rust. Through weighted sum of texture and spectral information, and then using the nonlinear characteristics of the Sigmoid function, an index field is generated which can accurately represent the degree of rust, providing a reliable basis for subsequent boundary extraction.

[0082] Step S2 realizes multi-dimensional representation of rust features through the technical combination of texture self-supervised encoding, color domain histogram analysis, spectral shift field calculation and nonlinear fusion. Texture self-supervised encoding uses unlabelled data to learn texture differences, reducing the cost of manual annotation and improving generalization ability; color domain histogram analysis provides color feature support for rust areas by quantifying color distribution; spectral shift field calculation captures frequency domain changes and reflects the spectral characteristics of rust; nonlinear fusion integrates texture, color and spectral information to generate an accurate rust index field.

[0083] Step S2 generates a rust texture matrix and a spectral shift field based on the light balanced image by combining texture self-supervised encoding with color domain histogram, and finally obtains a rust index field, laying a foundation for the recognition of rust features. However, in the steel reinforcement stacking scene, complex shadows, rust depth differences and highlight reflections may still cause the extraction of steel reinforcement boundaries to be not accurate enough, especially when adjacent steel reinforcement boundaries are fused or rust spots are misjudged as background, affecting the integrity of continuous edges. Step S3 aims to accurately extract the continuous edges of steel reinforcement and output a complete boundary set by fusing the rust index field and the depth point cloud, combining the cross-scale edge texture segmentation network and the spectral shift field repair technology, to solve the above boundary recognition problems.

[0084] Step S3 fuses the rust index field generated in step S2 with the depth point cloud generated in step S1, uses the cross-scale edge texture segmentation network to extract the continuous edges of steel reinforcement, and combines the spectral shift field generated in step S2 to repair the edge breaks, and finally outputs a complete boundary set. This step effectively improves the accuracy and integrity of boundary extraction for the problems of rust depth difference, shadow interference and highlight reflection in the steel reinforcement stacking scene on construction site, providing reliable boundary data for the center curve reconstruction and risk assessment in subsequent step S4.

[0085] Step S3 includes the following:

[0086] Rust index field and depth point cloud projection fusion:

[0087] The depth point cloud data is projected into the two-dimensional coordinate system of the light balance image to generate a depth image. The depth image records the depth value corresponding to each pixel position, and this process is completed through the geometric mapping between the point cloud data and the image. Next, the rust index field is fused with the depth image to generate a fused feature map. The specific way of fusion is: for each pixel position, take the product of the rust index value and the depth value, and then divide by the sum of the rust index value and the depth value; to avoid the case of zero divisor, a very small positive value is added in the denominator.

[0088] The fused feature map reflects the rust distribution characteristics and depth information through this harmonic average. The common influence of rust and depth is amplified through multiplication, the weights of the two are balanced in the form of division, and the synergistic effect of rust characteristics and spatial structure is highlighted through harmonic average, thereby enhancing the accuracy of subsequent edge extraction.

[0089] Cross-scale edge texture segmentation network:

[0090] The fused feature map is input into a multi-scale convolutional neural network. The network includes multiple branches, each branch uses different size convolution kernels to extract features of different scales to generate multi-scale feature maps. Then, edge detection is performed on the multi-scale feature maps, and edge probability maps are generated through convolution operation and activation function processing. The generation method of the edge probability map is: weighted sum of feature maps of different scales, and the weight coefficients of each scale are determined according to experimental optimization. Finally, threshold segmentation is performed on the edge probability map to generate a binary edge map. The specific way of threshold segmentation is: the part of the edge probability map whose pixel value is greater than or equal to the preset edge threshold is marked as edge and assigned a value of 1; the part whose value is less than the edge threshold is marked as non-edge and assigned a value of 0.

[0091] The diversity of edges is captured through multi-scale feature extraction, and the information of different scales is integrated through weighted sum, aiming to improve the robustness of edge detection and convert continuous probability into explicit edge labels through threshold segmentation.

[0092] Spectrum offset field repair sideband break:

[0093] The broken area is identified in the binarized edge map. The broken area refers to the discontinuous part of the edge pixel value changing from 1 to 0, which is determined by detecting the eight-neighbor connectivity of each edge pixel. Then, the broken area is repaired using the spectral offset field. The specific way of repair is as follows: in the broken area, the value of the spectral offset field is checked, if the value is greater than the preset spectral offset threshold and there is an edge pixel in the neighborhood, the position is marked as an edge and assigned a value of 1; otherwise, the original value remains unchanged.

[0094] The broken area is identified by neighborhood connectivity, the repair condition is judged by the value of the spectral offset field, and the broken area caused by corrosion interference is identified by the significant feature of the spectral offset field, ensuring the continuity and integrity of the edge.

[0095] Complete boundary set generation:

[0096] The repaired edge map is subjected to connected component analysis to extract the continuous boundary of each steel bar. This process uses a graph-based connectivity algorithm to mark independent boundary regions. Then, the extracted boundary is smoothed using a cubic spline interpolation method to generate a smooth boundary curve. Finally, the smoothed boundary curve is converted into a boundary point set to generate a complete boundary set, which is represented in the form of two-dimensional coordinates. The independent boundary is separated by connected component analysis, and the boundary curve is smoothed by interpolation, generating an accurate and smooth steel bar boundary to provide reliable boundary data for subsequent analysis.

[0097] Step S3 realizes the accurate extraction of the continuous boundary of the steel bar through the fusion of the corrosion index field and the depth point cloud projection, the processing of the cross-scale edge texture segmentation network, the spectral offset field repair technology, and the generation of the complete boundary set. The fusion of the corrosion index field and the depth information enhances the representation ability of the edge features, and can comprehensively reflect the corrosion distribution and spatial structure; the cross-scale edge texture segmentation network improves the robustness of edge detection through multi-scale feature extraction and weighted fusion, and adapts to edge changes of different scales; the spectral offset field repair technology effectively solves the boundary breakage problem caused by corrosion interference by using the significance of spectral features; the complete boundary set generation process ensures the continuity and smoothness of the boundary through connected component analysis and curve smoothing. The accuracy and integrity of the boundary extraction are significantly improved, which can adapt to the complex corrosion and lighting conditions on the construction site, and the generated boundary data has high quality, providing reliable support for the center curve reconstruction and risk assessment of the steel bar.

[0098] Step S3 fuses the rust index field with the depth point cloud projection, and combines it with a cross-scale edge texture segmentation network and spectral offset field repair technology to extract the complete boundary set of the steel bars, overcoming the problems of misidentification of rust spots and edge fractures. However, the complete boundary set only provides the outline information of the steel bars and cannot directly reflect the density and surface reflection characteristics of the steel bar stack, which are crucial for assessing the stability of the stack and the potential risk of collapse. Step S4 further extracts structural and material characteristics based on the complete boundary set and multi-angle polarization images, providing a quantitative basis for subsequent risk assessment.

[0099] Based on the complete boundary set output from step S3, step S4 reconstructs the central curve set through skeletonization and B-spline interpolation. Combined with height stratification and spacing calculations, this method generates the stacking void sequence entropy, quantifying the density of the rebar stack. Simultaneously, using the multi-angle polarized imagery provided by step S1 and the corrosion index field from step S2, the micro-rust reflectance divergence is calculated to characterize the corrosion-illumination heterogeneity of the rebar surface.

[0100] Step S4 includes the following contents:

[0101] Reconstruct the central curve based on the complete boundary set:

[0102] The boundary point set of each steel bar is skeletonized.

[0103] The skeletonization process generates an initial centerline by gradually removing the outer points of the boundary point set until the remaining point set forms a centerline with a width of one pixel. Specifically, the skeletonization process starts from the outside of the boundary point set, peeling off the outermost points layer by layer, retaining the inner points until the width of the point set is reduced to a single pixel, forming a line that can represent the topological center of the boundary point set. Next, the initial centerline is smoothed using the B-spline curve interpolation method. By inserting additional points between the ordered points of the initial centerline and adjusting the distribution of points according to the mathematical properties of the B-spline curve, a smooth center curve is generated. The center curve consists of a series of ordered points that can continuously and accurately characterize the spatial form of the steel bar.

[0104] Skeletonization is used to extract the geometric and topological centers of the boundary point set. The initial centerline is then smoothed and optimized using B-spline interpolation to ensure the continuity and smoothness of the center curve. Accurate geometric center features are extracted from the boundary information of the steel bar, providing a reliable representation for subsequent spatial structural analysis.

[0105] Compute the spacing between adjacent curves and generate the entropy of the stack gap sequence:

[0106] Firstly, all the center curves are sorted by their average height to generate a hierarchical sequence, where each layer contains several center curves. Specifically, the average height value of each center curve is calculated, and the center curves are grouped in order of height from low to high, and each group forms a layer. Then, within each layer, the average spacing between adjacent center curves in the horizontal direction is calculated to generate the spacing sequence of the layer. The average spacing is calculated by measuring the point-to-point distance of the two center curves in the horizontal projection and taking the average.

[0107] Then, the information entropy of the spacing sequence of each layer is calculated. The method of calculating information entropy is to first count the frequency of each spacing value in the spacing sequence, and then input these frequencies into the calculation formula of information entropy to obtain the entropy value representing the uniformity of the spacing distribution of the layer. The smaller the information entropy value, the more uniform the spacing distribution, and the higher the compactness of the pile; on the contrary, the larger the entropy value, the more random the spacing distribution, and the lower the compactness.

[0108] The compactness of the reinforcement pile is quantified by height layering and horizontal spacing calculation, and the distribution characteristics of the spacing sequence are analyzed by information entropy to reflect the uniformity of the structure of the pile. The spatial characteristics of the reinforcement pile are revealed by the compactness, providing data support for subsequent structural characteristic analysis.

[0109] According to the multi-angle polarization image, the micro-rust reflection divergence is calculated:

[0110] For each point on the surface of each reinforcement, the reflection value sequence of the point in different angle polarization images is extracted. Specifically, for each fixed point on the surface of the reinforcement, the reflection value of the point is obtained from the polarization images taken at different angles to form an ordered reflection value sequence. Then, the angular divergence of the reflection value sequence is calculated. The method of calculating the angular divergence is to nonlinearly weight the deviation of each reflection value in the reflection value sequence from the average value of the sequence, where the nonlinear weighting amplifies the deviation by a preset weighting function to highlight the unevenness of the reflection value distribution.

[0111] Then, the rust index field is taken as a weight and area-integrated with the angular divergence to generate the micro-rust reflection divergence. Specifically, the rust index field is a pre-defined distribution field representing the corrosion degree of the reinforcement surface. By multiplying the angular divergence point-by-point with the value of the rust index field on the reinforcement surface and integrating the entire surface, the micro-rust reflection divergence is obtained, which is used to represent the corrosion degree of the reinforcement surface and the unevenness of the illumination.

[0112] The unevenness of the optical properties of the surface is quantified by the divergence analysis of the reflection value sequence, and the contribution of the rust area to the divergence is highlighted by the fusion with the rust index field. Through the analysis of the reflection information of the multi-angle polarization image, the micro material characteristics of the reinforcement surface are revealed, providing a basis for subsequent material property evaluation.

[0113] By combining the reconstruction of the center curve based on the complete boundary set, the calculation of the adjacent curve spacing and the generation of the stack gap sequence entropy, and the calculation of the micro-rust reflection divergence from the multi-angle polarization image, the overall quantitative characterization of the steel reinforcement stacking form is realized in the structure and material level. The reconstruction of the center curve through skeletonization and B-spline interpolation provides an accurate geometric representation of the spatial form of the steel reinforcement, ensuring the spatial accuracy of subsequent analysis. The calculation of the adjacent curve spacing and the generation of the stack gap sequence entropy quantifies the density and uniformity of the steel reinforcement stacking through highly layered and information entropy analysis, revealing the compactness and distribution characteristics of the stacking structure. The calculation of the micro-rust reflection divergence from the multi-angle polarization image characterizes the corrosion degree and illumination non-uniformity of the steel reinforcement surface through the fusion of reflection value divergence and corrosion index field, providing micro-characteristic analysis in the material level. Further, it can accurately describe the steel reinforcement stacking form from both geometric structure and surface material dimensions, significantly improving the accuracy and comprehensiveness of stacking form recognition.

[0114] Step S4 reconstructs the center curve based on the complete boundary set, and calculates the stack gap sequence entropy and micro-rust reflection divergence, which quantifies the density and surface corrosion-illumination non-uniformity of the steel reinforcement stacking. However, the stability of the steel reinforcement stacking not only depends on the form and surface state of a single steel reinforcement, but also closely related to its spatial structure and interaction. Step S5 aims to use graph structure modeling and attention mechanism to comprehensively analyze the center curve, stack gap sequence entropy and micro-rust reflection divergence, identify potential collapse risk areas, and provide technical support for construction safety.

[0115] Step S5 is based on the center curve set, stack gap sequence entropy and micro-rust reflection divergence generated in step S4, by constructing a graph structure with center curves as nodes and gap sequence entropy difference as edge weights, and combining micro-rust reflection divergence input into graph attention encoder, it comprehensively analyzes the spatial distribution, stacking density and surface state characteristics of the steel reinforcement. Finally, by calculating the coupling coefficient and risk value, it realizes the accurate marking of potential collapse risk areas in the steel reinforcement stacking on the construction site. This step effectively explores the influence of spatial interaction and material characteristics between steel reinforcements on stability.

[0116] Step S5 includes the following content:

[0117] Build a graph structure with center curves as nodes:

[0118] Each center curve is defined as a node in the graph, and each center curve represents the spatial form of a steel reinforcement, which is derived from the data generated in the previous step that can represent the spatial position and form characteristics of the steel reinforcement. On this basis, edges are constructed between nodes through spatial proximity. The specific method is to judge the positional relationship of two center curves in space.

[0119] An edge is established between the corresponding nodes when the spatial distance of the two center curves reaches the minimum or there is an overlapping area in the projection of the spatial distance. The calculation of the spatial distance is achieved by comparing the shortest distance between the two center curves, and the detection of the projection overlap is based on whether the projection range of the center curve on the specified plane intersects.

[0120] The spatial distribution relationship of the steel bars is converted into the node and edge relationship in the graph structure, and the spatial topological characteristics of the steel bar stacking are captured to support the subsequent structured analysis.

[0121] The gap sequence entropy difference value is taken as the edge weight:

[0122] An edge weight is given to each edge in the graph structure, and the calculation of the edge weight is based on the absolute difference value of the gap sequence entropy between the layers where the adjacent nodes are located. The specific calculation method of the edge weight is to take the absolute value of the difference between the gap sequence entropy values of the layers where the adjacent nodes are located. If two nodes are located in the same layer, the edge weight is defined as zero. This process quantifies the density difference between the adjacent steel bar layers by reflecting the difference size as the edge weight value. The larger the edge weight, the more significant the density difference between the layers.

[0123] Joint micro-rust reflection divergence input graph attention encoder:

[0124] First, the micro-rust reflection divergence is taken as the initial feature vector of the node to input the graph attention encoder. The graph attention encoder adopts a multi-head attention mechanism to update the node features by calculating the attention weights between nodes. The specific calculation method is to concatenate the linear transformation of the initial feature vector of the node, then calculate the attention coefficient through the activation function and the exponential function, and finally weight the sum of the features of the neighbor nodes according to the attention coefficient to update the feature vector of the current node.

[0125] The weight of feature aggregation between nodes is dynamically adjusted through the attention mechanism to highlight the influence of important neighbor nodes, thereby fusing the surface features and spatial structure features of the steel bars and improving the accuracy and richness of the feature expression.

[0126] Obtain the matching coupling coefficient and mark the potential collapse risk area:

[0127] The process of obtaining the matching coupling coefficient and marking the potential collapse risk area first calculates the matching coupling coefficient of each edge, which is quantified based on the feature inner product of adjacent nodes and the edge weight. The specific calculation method is to add the inner product of the feature vectors of adjacent nodes to the edge weight value, and then perform nonlinear mapping through the Sigmoid function to obtain a matching coupling coefficient between 0 and 1, which represents the coupling strength between nodes. Then, the risk value of each node is calculated by multiplying the matching coupling coefficient of all connected edges and the corresponding edge weight, and the comprehensive risk index is obtained. Finally, according to the preset risk threshold, the steel reinforcement area corresponding to the node whose comprehensive risk index exceeds the risk threshold is marked as the potential collapse risk area.

[0128] By quantifying the interaction strength between nodes through the matching coupling coefficient, and combining the edge weight to evaluate the collapse risk of each node, the high-risk area in the steel reinforcement stacking is accurately identified.

[0129] By constructing a graph structure with central curves as nodes, using gap sequence entropy difference as edge weight, combining micro-rust reflection divergence input graph attention encoder, and obtaining matching coupling coefficient and marking potential collapse risk area, a comprehensive analysis of the steel reinforcement stacking form is realized in the structure and material level. The construction of the graph structure with the central curve as the node converts the spatial distribution relationship of the steel reinforcement into nodes and edges, providing structured data for analysis; the gap sequence entropy difference as the edge weight quantifies the difference in interlayer density, reflecting the stability characteristics of the stacking structure; the joint micro-rust reflection divergence input graph attention encoder fuses the steel surface features and spatial features through the attention mechanism, improving the accuracy of feature expression; the matching coupling coefficient and the marking of the potential collapse risk area accurately identify the high-risk area by quantifying the interaction between nodes and comprehensive risk assessment.

[0130] Step S5 identifies the potential collapse risk area by modeling the graph structure and the attention mechanism, combining the central curve, the stacking gap sequence entropy, and the micro-rust reflection divergence. However, steel reinforcement management on the construction site not only requires risk identification, but also needs to provide accurate statistical information to support inventory and grabbing planning. Step S6 aims to generate detailed statistical information using the central curve and risk area data and push it to the dispatch terminal to achieve intelligent management.

[0131] Step S6 generates a comprehensive statistical information report by accurately calculating the number, specification, and stability of the steel reinforcement based on the central curve set generated in step S4 and the potential collapse risk area marker set identified in step S5. This step uses the boundary point set to calculate the diameter distribution, and quantifies the stability in combination with the risk area marker, and finally pushes the results to the dispatch terminal, providing data support for steel reinforcement inventory and grabbing planning on the construction site. Under complex lighting and rust conditions, this method ensures the accuracy and practicality of the statistical information, and improves the intelligent level of construction management.

[0132] Step S6 includes the following:

[0133] Generate steel bar quantity and specification statistics based on center curves:

[0134] First, count the number of center curves, and determine the total number of steel bars by calculating the total number of elements in the center curve set. The center curve set is derived from the steel bar center path data extracted in the previous step, and the total number of elements directly corresponds to the number of steel bars. On this basis, for each center curve, calculate the diameter of the steel bar using the corresponding boundary point set. The specific method is to extract the boundary point pairs in the boundary point set along the cross-sectional direction of the steel bar, calculate the straight-line distance between each pair of boundary points in three-dimensional space to obtain multiple distance values, and take the average of these distance values as the diameter of the steel bar. The selection of boundary point pairs is achieved by identifying the relative position points on both sides of the cross-section in the boundary point set, and the calculation of straight-line distance is based on the three-dimensional coordinate values of the boundary points. Next, classify the diameters of all steel bars, divide them into different intervals according to the preset diameter range, and count the number of steel bars in each interval to generate a specification distribution table.

[0135] Directly determine the number of steel bars using the center curve set, and accurately calculate the diameter by analyzing the spatial position relationship of the boundary point set, aiming to provide accurate steel bar quantity and specification information through spatial form data analysis.

[0136] Generate stability statistics in combination with risk zones:

[0137] For each center curve, check whether its key node is located within the potential collapse risk zone. The key node refers to a feature point representing the position of the steel bar, such as the midpoint or endpoint of the center curve. If the key node is located within the potential collapse risk zone, the corresponding steel bar is marked as a high-risk steel bar; if the key node is not within the potential collapse risk zone, it is marked as a low-risk steel bar. The potential collapse risk zone is derived from the region data identified based on spatial analysis in the previous step.

[0138] Next, count the number of high-risk steel bars and low-risk steel bars by calculating the total number of steel bars marked as high-risk and low-risk, respectively, and further calculate the proportion of the number of high-risk steel bars and low-risk steel bars in the total number of steel bars to generate a stability statistics table.

[0139] Determine the risk level of the steel bar by analyzing the spatial attribution relationship between the key node and the potential collapse risk zone, and quantify the stability of the steel bar stacking through number statistics and proportion calculation.

[0140] Generate statistics information and push to dispatch terminal:

[0141] The steel bar quantity, specification distribution table and stability statistics table are integrated into one statistical information report, and the report content includes the total number of steel bars, the number of steel bars in each diameter range and its proportion in the total number, the number of high-risk steel bars and its proportion, and the number of low-risk steel bars and its proportion. Then, the statistical information report is transmitted to the dispatch terminal through the wireless communication module for the construction management personnel to use for inventory management and grabbing planning.

[0142] Through the combination of generating steel bar quantity and specification statistics based on the center curve, generating stability statistics combined with risk areas, and generating statistical information and pushing to the dispatch terminal, the overall statistics and risk assessment of the steel bar stacking form are realized. The generation of steel bar quantity and specification statistics based on the center curve uses spatial data analysis of the center curve set and the boundary point set to provide accurate information on the distribution of steel bar quantity and specification; the generation of stability statistics combined with risk areas accurately assesses the stability of the steel bar stacking through the relationship check of the key nodes and the potential collapse risk area; the generation of statistical information and the pushing to the dispatch terminal realize the intelligent management of the construction site by integrating the statistical data and transmitting them to the management end. The accuracy and practicality of the statistical information under complex environmental conditions are ensured, and the efficiency and intelligent level of construction management are improved, providing reliable data support for steel bar inventory and grabbing planning, which has significant engineering application value.

[0143] Embodiment 2: Figure 2 The present application provides a construction site steel bar stacking form intelligent identification system, which comprises:

[0144] The illumination depth acquisition module sequentially triggers the polarization flash of the annular array lens to acquire multi-angle images and depth point clouds, and outputs the aligned illumination balanced images and depth point clouds.

[0145] The corrosion feature fusion module performs texture self-supervised encoding on the images, generates a corrosion texture matrix and a spectral shift field combined with a color gamut histogram, and fuses to obtain a corrosion index field.

[0146] The edge contour extraction module projects and fuses the corrosion index field and the depth point cloud, extracts the steel bar continuous edge band through the cross-scale edge texture segmentation network, repairs the edge band breakage with the spectral shift field, and outputs the complete boundary set.

[0147] The structural material quantification module reconstructs the center curve based on the complete boundary set, calculates the distance between adjacent curves and generates the layer gap sequence entropy, and calculates the micro-corrosion reflection divergence according to the reflection direction variance of the multi-angle polarization image.

[0148] The risk area identification module constructs a graph structure with the center curve as the node and the gap sequence entropy difference value as the edge weight, inputs the graph attention encoder combined with the micro-corrosion reflection divergence to obtain the coupling coefficient, and marks the potential collapse risk area.

[0149] Statistical information pushing module: generate statistical information report combining the center curve and the risk area, and push to the dispatch terminal for inventory and capture planning.

[0150] The above formulas are dimensionless values calculated, and the formulas are obtained by collecting a large amount of data to simulate a formula of the nearest real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0151] It should be noted that the system of the present application can be deployed on the device itself to realize embedded application, or run on PC or other terminal with user interface, so as to meet various hardware environments and use requirements.

[0152] The above only describes some exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

[0153] It should be noted that in this paper, if there are relationship terms such as first and second, they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "include" "contain" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or equipment. Without more limitations, the element defined by the sentence "including a" does not exclude the presence of other identical elements in the process, method, article or equipment including the element.

[0154] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An intelligent identification method for steel bar stacking morphology at a construction site, characterized in that: Including steps: S1: The annular array lens sequentially triggers polarized flash operations to capture multi-angle images and depth point clouds, and outputs aligned illumination-balanced images and depth point clouds; S2: Perform texture self-supervised encoding on the illumination-balanced image, combine the color gamut histogram to generate a rust texture matrix and a spectral offset field, and fuse them to obtain a rust index field, which is used to characterize the degree of rust. S3: Project and fuse the corrosion index field with the depth point cloud, extract the continuous edge bands of the steel bars through the cross-scale edge texture segmentation network, and use the spectral offset field to repair the broken edge bands and output a complete boundary set; S4: Reconstruct the central curve based on the complete boundary set, calculate the distance between adjacent curves and generate the layer gap sequence entropy, and calculate the micro-rust reflection divergence based on the reflectance value sequence of multi-angle polarization images; Step S4 includes: The complete boundary set is skeletonized and the central curve is reconstructed using B-spline interpolation. The steel bars are sorted in layers according to the average height of the central curves. The average horizontal distance between adjacent central curves in each layer is calculated to generate a spacing sequence. The information entropy is calculated based on the spacing sequence to quantify the stacking density. For the same point on the surface of each steel bar, a reflectance value sequence is extracted from the multi-angle polarization image. The micro-rust reflectance divergence is generated using the stacking density and the reflectance value sequence. The micro-rust reflectance divergence is used to characterize the unevenness of rust and light on the steel bar surface. S5: A graph structure is constructed with the central curve as the node and the difference between the entropies of the stack gap sequences as the edge weight. The micro-rust reflection divergence is combined with the input graph attention encoder to obtain the coordination coupling coefficient to mark the potential collapse risk area. The coordination coupling coefficient is used to characterize the coupling strength between nodes. Step S5 includes: The central curve is defined as a node in the graph, and edges are established between nodes based on spatial proximity to construct the graph structure. The absolute value of the difference in the entropy of the stack gap sequence of adjacent nodes is used as the edge weight. The micro-rust reflection divergence is input as the initial feature vector of the node into the graph attention encoder, and the node features are updated through the multi-head attention mechanism. The inner product of the feature vectors of adjacent nodes is added to the edge weight value, and then nonlinearly mapped using the Sigmoid function to obtain a coordination coupling coefficient between 0 and 1. The comprehensive risk index of each node is then calculated and compared with the preset risk threshold. The steel bar area corresponding to the node whose comprehensive risk index exceeds the preset risk threshold is marked as a potential collapse risk area. S6: Generate a statistical information report based on the central curve and potential collapse risk area, and push it to the dispatch terminal.

2. The intelligent identification method for steel bar stacking morphology at a construction site according to claim 1, characterized in that: Step S1 includes the following contents: Light-balanced images and depth point clouds are collected through an annular array lens and polarized flash technology. The annular array lens consists of high-resolution cameras evenly distributed around the steel bar stacking area and is equipped with a polarization filter. The polarized flash source corresponds to the high-resolution cameras and is triggered sequentially. The flash intensity is adjusted according to the ambient light. Each high-resolution camera collects a two-dimensional image and generates a light-balanced image through grayscale histogram equalization and color temperature correction. The depth sensor collects depth information and generates a filtered depth point cloud through voxel filtering. The filtered depth point cloud is projected onto the image plane of each camera using a spatial transformation matrix to achieve alignment between the light-balanced image and the depth point cloud, and the aligned light-balanced image and depth point cloud are output.

3. The intelligent identification method for steel bar stacking morphology at a construction site according to claim 2, characterized in that: Step S2 includes the following contents: Texture self-supervised encoding is performed on the illumination-balanced image, and the convolutional neural network is trained using self-supervised learning. The network parameters are optimized through comparative learning, so that the convolutional neural network can learn the texture differences between rusted and non-rusted areas and generate a texture feature map.

4. The intelligent identification method for steel bar stacking morphology at a construction site according to claim 3 is characterized in that: Step S3 includes the following contents: Fuse the corrosion index field with the depth point cloud projection to generate a fused feature map; The fused feature map is input into the cross-scale edge texture segmentation network to generate a multi-scale feature map and perform edge detection to generate an edge probability map; The edge probability map is threshold segmented to generate a binary edge map. The broken areas in the binary edge map are repaired using the spectral offset field to generate a repaired edge map. The repaired edge map is subjected to connected domain analysis to extract the continuous boundaries of the steel bars and smooth them to generate a complete boundary set.

5. The intelligent identification method for steel bar stacking morphology at a construction site according to claim 4 is characterized in that: Step S6 includes the following contents: The total number of steel bars is calculated based on the central curve, and the diameter of each steel bar is calculated using the complete boundary set, generating a specification distribution table that includes the number of steel bars within each diameter range and their proportion to the total number of steel bars. The risk level of each steel bar is determined by the spatial relationship between the key nodes of the central curve and the potential collapse risk area. The number of high-risk steel bars and their proportion to the total number of steel bars, as well as the number of low-risk steel bars and their proportion to the total number of steel bars, are counted, and a stability statistics table is generated. The total number of steel bars, specification distribution table, and stability statistics table are integrated into a statistical information report and transmitted to the dispatching terminal via a wireless communication module.

6. A system for intelligently identifying the stacking form of steel bars at a construction site, for implementing the method for intelligently identifying the stacking form of steel bars at a construction site as claimed in any one of claims 1 to 5, characterized in that: include: Lighting depth acquisition module: The annular array lens sequentially triggers polarized flash operations to collect multi-angle images and depth point clouds, and outputs aligned lighting-balanced images and depth point clouds; The corrosion feature fusion module performs texture self-supervision encoding on the illumination-balanced image, combines the color gamut histogram to generate a corrosion texture matrix and a spectral offset field, and fuses them to obtain a corrosion index field, which is used to characterize the degree of corrosion. The edge contour extraction module projectively fuses the corrosion index field with the depth point cloud, extracts continuous steel bar edge bands through a cross-scale edge texture segmentation network, and repairs broken edge bands using the spectral offset field, outputting a complete boundary set. Structural material quantification module: reconstructs the central curve based on the complete boundary set, calculates the distance between adjacent curves and generates the entropy of the stacking layer gap sequence, and calculates the micro-rust reflection divergence based on the reflection value sequence of multi-angle polarization images; Risk area identification module: A graph structure is constructed using central curves as nodes and the difference between the entropies of the stack gap sequences as edge weights. The attention encoder of the micro-rust reflection divergence input graph is combined with the obtained coordination coupling coefficient to mark potential collapse risk areas. The coordination coupling coefficient is used to characterize the coupling strength between nodes. Statistical information push module: Generates statistical information reports based on the central curve and potential collapse risk areas, and pushes them to the dispatching terminal.

Images