Asbestos-cement quality detection method based on multi-source sensing

Abstract

The application discloses a kind of based on multi-source sensing's asbestos cement quality detection method, belong to image data processing technical field, specifically include: first, the high-resolution topography and thermodynamic image of asbestos cement test piece are collected, and initial structure distribution matrix is constructed to align bimodal coordinate system;Subsequently, local pixel direction gradient is extracted, and the tensor feature map representing fiber direction is constructed;Then, the thermodynamic image is projected to the feature map, and the bias of heat conduction attenuation and fiber direction is calculated to locate the cross-modal thermal anomaly node;Then, the spatial topology graph is established with the abnormal node as the vertex, the geometric connectivity weight is calculated, and the normal fiber interference node is completely filtered out;Finally, the boundary profile of microcracking area is reconstructed along the connected path of the remaining vertex, and the physical coordinate sequence of structural defects is output;The present application can accurately extract and reconstruct the microstructure defect from the extremely complex base texture, effectively solve the serious misjudgment problem caused by the image feature being submerged.

Description

Technical Field

[0001] This invention relates to the field of image data technology, and more specifically to a method for quality detection of asbestos cement based on multi-source sensing. Background Technology

[0002] Asbestos cement, as a basic building and industrial material, requires core processes in its production, including pressing and curing. To control the structural strength and safety indicators of the final product, the industry commonly uses multi-source sensing technology to acquire physical characterization information of the material. Existing testing procedures typically combine structured light equipment and infrared thermal imagers to acquire high-resolution morphological images and thermodynamic distribution images of the specimen, respectively. After acquiring the dual-modal data, conventional computer vision systems analyze pixel grayscale and edge gradients based on image feature extraction models to extract the product's microstructure and complete quality assessment.

[0003] However, in the pressing, molding, and curing processes of asbestos cement products, the substrate texture formed by the random interlacing of asbestos fibers on the material surface and inside exhibits extremely high visual complexity and spatial chaos. In high-resolution morphological and thermodynamic distribution images captured by multi-source sensing devices, areas of normal fiber overlap and aggregation, as well as real defect areas such as micro-cracks and pores, show highly similar visual morphologies in pixel grayscale and edge gradients. Conventional image feature extraction models cannot extract discriminative multidimensional morphological features from strongly interfering background textures and lack the ability to capture image features of implicit spatial topological relationships in cross-modal images. This causes extremely subtle early structural defect image features to be completely submerged by normal fiber substrate features, leading to serious missed detections at the pattern recognition level. Summary of the Invention

[0004] The purpose of this invention is to provide a method for quality testing of asbestos cement based on multi-source sensing, thereby solving the problems in the background art.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for quality testing of asbestos cement based on multi-source sensing includes the following steps:

[0007] S1: Acquire high-resolution morphological and thermodynamic images of asbestos cement specimens, and construct an initial structural distribution matrix by aligning the dual-modal pixel coordinate system;

[0008] S2: Extract the local pixel orientation gradient of the topography image in the initial structure distribution matrix and construct a tensor feature map representing the interlacing direction of asbestos fibers;

[0009] S3: Project the thermodynamic image in the matrix onto the tensor feature map, calculate the local heat conduction attenuation and the offset of fiber orientation, and locate the cross-modal thermal anomaly node;

[0010] S4: Establish a spatial topology graph with cross-modal thermal anomaly nodes as network vertices, calculate the geometric connectivity weights between adjacent vertices, and filter out normal fiber interference nodes;

[0011] S5: Reconstruct the boundary contour of the micro-cracked region along the connected paths between the remaining network vertices, and output the physical coordinate sequence of the defects in the asbestos cement structure.

[0012] As a further aspect of the present invention: in step S1, the process of aligning the dual-modal pixel coordinate system to construct the initial structure distribution matrix is ​​as follows:

[0013] The structured light camera and infrared thermal imager are used to simultaneously scan the surface of the asbestos cement specimen, and high-resolution morphological and thermodynamic images of the asbestos cement specimen are acquired.

[0014] Extract matching feature points from high-resolution topographic images and thermodynamic images, perform spatial mapping to align the dual-modal pixel coordinate system based on the matching feature points, and merge pixel channels to construct an initial structure distribution matrix.

[0015] As a further aspect of the present invention: in step S2, the process of constructing the tensor feature map characterizing the interlacing orientation of asbestos fibers is as follows:

[0016] Spatial difference operation is performed on the topography image in the initial structure distribution matrix to extract the gray-level derivatives of pixel nodes in the topography image on the orthogonal axis and synthesize the local pixel orientation gradient.

[0017] Calculate the outer product matrix of the gradient of local pixel directions, perform spatial neighborhood integration on the outer product matrix, and aggregate the matrix elements after integration into a tensor feature map representing the interlacing direction of asbestos fibers.

[0018] As a further aspect of the present invention: the specific method for extracting the gray-level derivatives of pixel nodes in the topography image on orthogonal axes and synthesizing local pixel directional gradients is as follows:

[0019] By performing sliding convolution on the topography image of the initial structure distribution matrix using horizontal and vertical difference operators, the gray-level derivatives of pixel nodes in the topography image on the orthogonal axis are extracted.

[0020] The gradient magnitude is extracted by summing the squares of the gray-level derivatives on the orthogonal axes and taking the square root. The gradient direction is extracted by calculating the arctangent of the quotient of the gray-level derivatives on the orthogonal axes. The gradient magnitude and gradient direction are combined to synthesize the local pixel direction gradient.

[0021] As a further aspect of the present invention: in step S3, the process of locating the cross-modal thermal anomaly node is as follows:

[0022] Based on the pixel coordinates of the initial structure distribution matrix, bilinear interpolation is performed on the thermodynamic image in the matrix to project the thermodynamic image onto the spatial grid of the tensor feature map;

[0023] Calculate the temperature gradient of the thermodynamic image in the spatial grid, generate a local heat conduction attenuation vector, analyze the principal direction vector of the tensor feature map as the fiber direction, and calculate the angle offset between the two.

[0024] Calculate the offset difference between the included angle offset and the set reference offset threshold, extract the spatial grid nodes with positive offset differences, and locate them as cross-modal thermal anomaly nodes.

[0025] As a further aspect of the present invention: in step S4, the process of calculating the geometric connectivity weights between adjacent vertices and filtering out normal fiber interference nodes is as follows:

[0026] The located cross-modal thermal anomaly nodes are mapped to network vertices, the spatial coordinates of the network vertices are extracted and neighborhood connection operations are performed to establish a spatial topology graph containing adjacent vertices.

[0027] Extract the Euclidean distance between adjacent vertices in the spatial topology graph, calculate the sine value of the angle between the direction of the connection and the fiber direction, and divide the sine value by the sum of the Euclidean distance and a preset constant to generate the geometric connectivity weight.

[0028] Calculate the difference between the geometric connectivity weight and the set retention threshold, extract adjacent vertices with negative differences and determine them as normal fiber interference nodes, and filter out normal fiber interference nodes in the spatial topology graph.

[0029] As a further aspect of the present invention: the specific content of establishing a spatial topology graph containing adjacent vertices is as follows:

[0030] The data indexes of the located cross-modal thermal anomaly nodes are mapped to network vertices. The position components of the network vertices in the bimodal pixel coordinate system are read, and the spatial coordinates of the network vertices are extracted.

[0031] Based on spatial coordinates, extract adjacent vertices whose distance is lower than the connected boundary, perform neighborhood connection operations on adjacent vertices to generate spatial connected edges, and aggregate the spatial connected edges and network vertices to establish a spatial topology graph containing adjacent vertices.

[0032] As a further aspect of the present invention: in step S5, the process of outputting the physical coordinate sequence of asbestos cement structural defects is as follows:

[0033] Read the connected paths between the remaining network vertices, perform tracing operations along the outside of the connected paths to extract the outermost contour line, and based on the comparison results of the distance between the first and last nodes and the closure threshold, adaptively close or retain the extended shape to reconstruct the boundary contour of the micro-cracked region.

[0034] Extract the pixel coordinates of the pixel nodes corresponding to the boundary contour, obtain the spatial transformation matrix of the dual-modal pixel coordinate system, use the spatial transformation matrix to solve the pixel coordinates, and output the physical coordinate sequence of the defects in the asbestos cement structure.

[0035] The beneficial effects of this invention are:

[0036] The technical solution provided by this invention effectively overcomes the technical bottleneck of extremely high visual complexity of asbestos cement substrate texture by constructing a tensor feature map characterizing the interlacing orientation of asbestos fibers and projecting a thermodynamic image onto this tensor feature map to calculate the angular offset between local heat conduction attenuation and fiber orientation. Since normal fiber aggregation areas and micro-cracks are highly similar in conventional pixel grayscale and edge gradients, this application delves into the underlying physical correlation of cross-modal data. Utilizing the characteristic that heat is conducted unidirectionally in intact fibers but attenuates due to obstruction at cracks, it accurately extracts the offset difference between the heat attenuation vector and the main fiber direction. This mechanism successfully extracts discriminative multidimensional morphological features from strongly interfering background textures, accurately locates cross-modal thermal anomaly nodes, and prevents extremely small early structural defect features from being completely submerged by substrate features.

[0037] Furthermore, the technical solution provided by this invention establishes a spatial topology graph with cross-modal thermal anomaly nodes as network vertices, and introduces geometric connectivity weights based on sine values ​​and Euclidean distances between connections, thus compensating for the lack of implicit spatial topological relationships in conventional models. Microscopic cracks typically exhibit a tearing morphology spanning the fiber direction. This application uses the sine value of the angle between the connection direction and the fiber direction as the core molecule of the weights, ensuring that normal interference nodes along the fiber direction are completely filtered out due to their excessively small weight values. Finally, based on the comparison between the distance between the first and last nodes and the closure threshold, the boundary contour of the microscopic crack region is adaptively reconstructed. This topology filtering and adaptive morphology reconstruction mechanism accurately extracts and converts minute structural defects from spatial chaotic features into physical coordinate sequences, solving the serious problem of missed detection at the pattern recognition level. Attached Figure Description

[0038] The invention will now be further described with reference to the accompanying drawings.

[0039] Figure 1 This is a flowchart illustrating a method for detecting the quality of asbestos cement based on multi-source sensing, according to the present invention.

[0040] Figure 2 This is a schematic diagram of the modal acquisition and alignment process of an asbestos cement quality detection method based on multi-source sensing according to the present invention.

[0041] Figure 3 This is a schematic diagram of the tensor feature map construction process of an asbestos cement quality detection method based on multi-source sensing according to the present invention.

[0042] Figure 4 This is a schematic diagram of the cross-modal thermal anomaly node localization process of an asbestos cement quality detection method based on multi-source sensing according to the present invention.

[0043] Figure 5 This is a schematic diagram of the inter-topological filtering process of an asbestos cement quality detection method based on multi-source sensing according to the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1

[0046] Please see Figure 1 As shown, this invention is a method for quality testing of asbestos cement based on multi-source sensing, comprising the following steps:

[0047] S1: Acquire high-resolution morphological and thermodynamic images of asbestos cement specimens, and construct an initial structural distribution matrix by aligning the dual-modal pixel coordinate system;

[0048] S2: Extract the local pixel orientation gradient of the topography image in the initial structure distribution matrix and construct a tensor feature map representing the interlacing direction of asbestos fibers;

[0049] S3: Project the thermodynamic image in the matrix onto the tensor feature map, calculate the local heat conduction attenuation and the offset of fiber orientation, and locate the cross-modal thermal anomaly node;

[0050] S4: Establish a spatial topology graph with cross-modal thermal anomaly nodes as network vertices, calculate the geometric connectivity weights between adjacent vertices, and filter out normal fiber interference nodes;

[0051] S5: Reconstruct the boundary contour of the micro-cracked region along the connected paths between the remaining network vertices, and output the physical coordinate sequence of the defects in the asbestos cement structure.

[0052] Example 2

[0053] Please see Figure 2 As shown, this embodiment is based on Embodiment 1. Regarding step S1, the process of acquiring high-resolution morphological and thermodynamic images of the asbestos cement specimen and aligning the dual-modal pixel coordinate system to construct the initial structural distribution matrix is ​​as follows:

[0054] An asbestos cement specimen was placed flat on a high-precision motorized translation stage. A line laser 3D contour scanner was used as the structured light camera, and an uncooled vanadium oxide microbolometer infrared camera was used as the infrared thermal imager. A synchronous trigger pulse signal was sent from the main control circuit board to both the line laser 3D contour scanner and the uncooled vanadium oxide microbolometer infrared camera, ensuring that both began acquiring data at the same pulse edge. The line laser 3D contour scanner projected coded laser stripes onto the specimen surface and received the reflected light, calculating a high-resolution morphological image of the specimen surface using the principle of triangulation. The uncooled vanadium oxide microbolometer infrared camera received the infrared energy spontaneously emitted from the specimen surface, converting it into a temperature electrical signal and quantifying it into a thermodynamic image. The high-precision motorized translation stage moved at a constant speed, causing the specimen to perform a one-dimensional linear motion relative to the camera, completing a line-by-line scan of the entire field of view.

[0055] Gaussian blur denoising was applied to the acquired high-resolution topographic images, and the intersections of corners and edges were extracted as topographic feature points using a scale-invariant feature transform algorithm. Histogram equalization was performed on the thermodynamic images to enhance contrast, and thermal extrema were extracted as thermodynamic feature points using the same scale-invariant feature transform algorithm. The descriptor vectors of the topographic and thermodynamic feature points were compared using the nearest neighbor ratio matching method, eliminating mismatched point pairs and retaining spatially consistent feature point pairs as matching feature points. Based on the pixel coordinate pairs of the matching feature points, an affine transformation model between the topographic and thermodynamic images was constructed, and the six parameters of the affine transformation matrix were solved using the least squares method. The affine transformation matrix was applied to each pixel coordinate of the thermodynamic image, and the resampled value of that coordinate in the topographic image coordinate system was calculated through interpolation, completing the spatial mapping and alignment of the dual-modal pixel coordinate systems.

[0056] In an aligned dual-modal pixel coordinate system, the grayscale values ​​of the high-resolution topographic image at the same spatial coordinate are used as the first channel data, and the quantized temperature values ​​of the thermodynamic image are used as the second channel data. The first and second channel data are concatenated along the depth direction, so that each spatial grid node contains both light intensity and thermal physical property information. By traversing all aligned pixel nodes in the entire image, the concatenated dual-channel data is arranged in two dimensions according to row and column coordinates, generating a tensor structure with a dimension of two (number of rows x number of columns x two). This tensor structure is the initial structure distribution matrix.

[0057] Example 3

[0058] Please see Figure 3 As shown, based on the above embodiments, this embodiment, regarding step S2, involves extracting the local pixel orientation gradients of the morphology image in the initial structure distribution matrix and constructing a tensor feature map characterizing the interlacing orientation of asbestos fibers as follows:

[0059] The specific execution process of extracting local pixel orientation gradients from the topography image in the initial structure distribution matrix and constructing tensor feature maps relies on an embedded image processing accelerator card. The image acquisition module directly reads the topography image data from the initial structure distribution matrix into the video memory unit of the image processing accelerator card. The coprocessor inside the image processing accelerator card calls the horizontal and vertical difference operators to perform a sliding convolution operation on the topography image. This operation uses a preset 3rd-order convolution kernel to slide pixel by pixel on the two-dimensional pixel array of the topography image. The horizontal difference operator is used to extract the gray-level derivative of the pixel node in the horizontal dimension of the orthogonal axis. The vertical difference operator is used to extract the gray-level derivative of the pixel node in the vertical dimension of the orthogonal axis. The mathematical calculation process of the sliding convolution satisfies the given formula: ; ; where, in the formula The pixel grayscale matrix representing the shape image. Represents the horizontal difference operator matrix. This represents the vertical difference operator matrix, and the symbol * denotes the discrete spatial convolution operation. Represents the grayscale derivative on the horizontal orthogonal axis. This represents the grayscale derivative on the perpendicular orthogonal axis.

[0060] After obtaining the gray-level derivatives on the orthogonal axes, the computational logic unit begins the operation of synthesizing the local pixel orientation gradient. The computational logic unit reads the gray-level derivatives on the horizontal and vertical orthogonal axes and sums their squares. The square root operation is performed on the accumulated sums to extract the gradient magnitude of the current pixel node. The computational logic unit divides the gray-level derivative on the vertical orthogonal axis by the gray-level derivative on the horizontal orthogonal axis to obtain the gray-level derivative quotient. The arctangent function is applied to this gray-level derivative quotient to calculate the gradient direction of the current pixel node. The computational logic unit concatenates the gradient magnitude and gradient direction of the same pixel node to synthesize the complete local pixel orientation gradient. The specific calculation process for the above magnitude and direction satisfies the set mathematical relationship formula: ; In the formula, M represents the gradient magnitude. This represents the radian angle value of the gradient direction. The computational logic unit generates a two-dimensional directional gradient vector field in the feature memory array based on the synthesized magnitude and direction to accurately characterize the intensity of texture within the topographic image.

[0061] After obtaining the local pixel orientation gradient, the microprocessor chip needs to calculate the outer product matrix of the local pixel orientation gradient. The microprocessor chip reads the horizontal and vertical grayscale derivatives of each pixel node to construct an orientation gradient column vector. This orientation gradient column vector is then multiplied by its transpose row vector to generate the second-order outer product matrix corresponding to the current pixel node. The feature representation of the elements within this outer product matrix satisfies the set outer product generation formula: ; where, in the formula This represents the initial outer product matrix for a single pixel node. This represents the square component of the horizontal gray-level derivative. This represents the square component of the gray-level derivative in the vertical direction. This represents the cross-product term of the gray-level derivatives in the horizontal and vertical directions. Since the initial outer product matrix of a single pixel node is easily affected by environmental noise, leading to distortion in fiber orientation representation, the microprocessor chip must perform spatial neighborhood integration on this outer product matrix to eliminate abrupt interference.

[0062] The spatial neighborhood integration process is implemented by calling a two-dimensional Gaussian smoothing filter. The microprocessor chip sets a fifth-order spatial neighborhood sliding window centered on the current pixel node. A weighted summation is performed on all initial outer product matrix elements within the spatial neighborhood sliding window. The weight coefficients of the weighted summation are determined by the two-dimensional spatial distribution function; matrix nodes closer to the center pixel are assigned larger weight values. The specific spatial neighborhood integration operation satisfies the tensor integral mathematical formula: ; where, in the formula This represents the smooth structure tensor matrix after integration. This represents the two-dimensional Gaussian kernel function distribution matrix, where the symbol * indicates neighborhood-weighted convolution integration. The microprocessor chip traverses each pixel node in the topographic image and sequentially performs the aforementioned spatial neighborhood integration. The computational execution core aggregates all the smoothed structural tensor matrix elements after integration operations according to their original physical coordinate space mappings. The resulting overall tensor array directly serves as the tensor feature map representing the interlacing orientation of asbestos fibers.

[0063] Example 4

[0064] Please see Figure 4 As shown, this embodiment, based on the above embodiments, specifically addresses step S3, whereby the process of projecting the thermodynamic image in the matrix onto a tensor feature map, calculating the bias between local heat conduction attenuation and fiber orientation, and locating cross-modal thermal anomaly nodes is as follows:

[0065] The specific execution process of projecting the thermodynamic image from the matrix onto the tensor feature map is accomplished using a field-programmable gate array (FPGA). The data parsing module extracts the second channel data from the initial structure distribution matrix as the initial thermodynamic image array. Since the initial thermodynamic image array and the tensor feature map generated in step S2 have different grid scales in physical space resolution, the coordinate mapping circuit inside the FPGA extracts the pixel coordinates of the initial structure distribution matrix. The coordinate mapping circuit maps these pixel coordinates to the physical space grid of the tensor feature map to find a mapping anchor point. The interpolation logic unit extracts the four nearest discrete pixel nodes in the initial thermodynamic image array, centered on this mapping anchor point. The interpolation logic unit reads the temperature quantization values ​​of these four discrete pixel nodes and performs a linear weighted summation operation sequentially along the horizontal and vertical axes. The weight coefficients of the weighted summation operation are determined by the absolute Euclidean distance between the mapping anchor point and the four surrounding discrete pixel nodes; the closer the discrete pixel node, the larger the weight coefficient. Bilinear interpolation smoothly projects low-resolution temperature quantization values ​​onto a spatial grid of a high-resolution tensor feature map, thereby generating a spatially aligned projected thermodynamic image.

[0066] After acquiring the projected thermodynamic image, the multi-core digital signal processor (DSP) calculates the temperature gradient of the projected thermodynamic image across the spatial grid. The DSP reads the quantized temperature values ​​of each grid node in the projected thermodynamic image and uses a differential filter array to extract the rate of temperature change along the horizontal and vertical axes for each grid node. The DSP then orthogonally combines the rate of temperature change along the horizontal and vertical axes to directly generate a local heat conduction attenuation vector characterizing the intensity of heat transfer. The calculation logic for the local heat conduction attenuation vector satisfies the established thermal gradient extraction formula: ; where, in the formula Represents the local heat conduction attenuation vector. This represents the quantized temperature value of a spatial grid node. Represents the physical coordinate components along the horizontal axis. Represents the physical coordinate components along the vertical axis.

[0067] While generating the local heat conduction attenuation vector, the matrix decomposition unit inside the microcontroller begins to parse the principal direction vector of the tensor feature map. The matrix decomposition unit reads the smooth structure tensor matrix elements corresponding to each spatial grid node in the tensor feature map and performs eigenvalue decomposition on this smooth structure tensor matrix. The matrix decomposition unit solves for the two orthogonal eigenvectors and two scalar eigenvalues ​​corresponding to the smooth structure tensor matrix. The microcontroller selects the eigenvector corresponding to the smallest scalar eigenvalue as the principal direction vector representing the local texture direction and directly uses this principal direction vector as the fiber orientation of the current physical node region. The arithmetic logic unit inside the microcontroller reads the local heat conduction attenuation vector and the fiber orientation, and calculates the cosine value by taking their inner product and dividing by the product of their magnitudes. The arithmetic logic unit performs an inverse trigonometric function mapping operation on this cosine value to accurately obtain the angle offset between the local heat conduction attenuation and the fiber orientation. The calculation logic for the angle offset satisfies the set offset calculation formula: ; in the formula Indicates the included angle offset. The vertical line indicates the selected principal direction vector, i.e., the fiber orientation, and the vertical line symbol indicates the scalar modulus of the orientation vector.

[0068] After the arithmetic logic unit obtains the included angle bias, the anomaly detection circuit begins locating the cross-modal thermal anomaly node. The registers within the anomaly detection circuit are pre-programmed with a reference bias threshold characterizing the healthy fiber structure. The anomaly detection circuit reads the included angle bias corresponding to each spatial grid node and inputs it into the subtractor module. The subtractor module subtracts the reference bias threshold from the included angle bias to calculate the specific bias difference data. The sign bit detection module within the anomaly detection circuit reads the highest sign bit of this bias difference data. When the sign bit detection module detects a positive sign bit, it indicates that the calculated bias difference is greater than 0, and the local heat conduction attenuation direction of the current physical node deviates significantly from the fiber orientation at the physical level. The anomaly detection circuit extracts the spatial grid node with a positive bias difference value and pushes its three-dimensional coordinate data to the anomaly recording memory, accurately locating the spatial grid node as a cross-modal thermal anomaly node.

[0069] Example 5

[0070] Please see Figure 5 As shown, this embodiment, based on the above embodiments, specifically addresses step S4, where the process of establishing a spatial topology graph using cross-modal thermal anomaly nodes as network vertices, calculating the geometric connectivity weights between adjacent vertices, and filtering out normal fiber interference nodes is as follows:

[0071] The specific execution process for building a spatial topology graph containing adjacent vertices relies on a dedicated topology graph processing chip. The addressing control module inside the chip reads cross-modal hot anomaly node data from the anomaly record memory. The addressing control module assigns a unique integer data index to each cross-modal hot anomaly node and maps it to an independent network vertex in logical memory. The coordinate resolution circuit reads the position components of this network vertex in the bimodal pixel coordinate system, obtaining the corresponding horizontal pixel column coordinates and vertical pixel row coordinates. The coordinate resolution circuit combines and concatenates the horizontal pixel column coordinates and vertical pixel row coordinates to accurately extract the two-dimensional spatial coordinates of the corresponding network vertex.

[0072] After obtaining the spatial coordinates of the network vertices, the topology graph processing chip calls its internal feature tree acceleration engine to search for neighboring vertices. The feature tree acceleration engine uses the spatial coordinates of the current network vertex as the center and reads a preset connected boundary value from its internal non-volatile memory as the search radius. The feature tree acceleration engine traverses other network vertices within the neighborhood and extracts neighboring vertices whose distance is below the connected boundary. The connection generation logic unit extracts the spatial coordinates of the current network vertex and the selected neighboring vertices, and performs neighborhood connection operations between them to generate virtual spatial connected edges. The data aggregation bus of the topology graph processing chip concatenates and reassembles all the generated spatial connected edges with the corresponding network vertices into a graph data structure, establishing a spatial topology graph containing neighboring vertices in the high-speed static random access memory.

[0073] After the spatial topology graph is constructed, the floating-point coprocessor begins calculating the geometric connectivity weights between adjacent vertices. The floating-point coprocessor reads the spatial coordinates of adjacent vertices at both ends of a spatially connected edge and extracts the Euclidean distance between adjacent vertices using the square root of the sum of the squares of the coordinate differences. The direction resolution module extracts the spatial vector corresponding to the Euclidean distance as the connection direction vector and reads the corresponding region fiber direction vector from the data cache array. The floating-point coprocessor calculates the angle between this connection direction vector and the region fiber direction vector and performs a sine mapping operation on this angle to obtain the sine value of the angle between the connection direction and the fiber direction. The floating-point coprocessor extracts this sine value as the numerator and adds the connection Euclidean distance to a preset constant in the preset storage area to obtain the denominator. The divider logic unit performs the arithmetic operation of dividing the numerator by the denominator to generate the geometric connectivity weights between the current adjacent vertices. The specific calculation logic of the geometric connectivity weights satisfies the set topology weight mathematical formula: ; where, in the formula Indicates the generated geometric connectivity weights. It represents the angle in radians between the direction of the line and the direction of the fiber. This represents the Euclidean distance between adjacent vertices. This represents a preset constant introduced to ensure numerical stability.

[0074] After acquiring all geometric connectivity weights, the node filtering logic circuit begins filtering out normal fiber interference nodes. The node filtering logic circuit reads the set retention threshold from the flash memory and transmits it to a parallel numerical comparator. The parallel numerical comparator reads the geometric connectivity weights between each adjacent vertex and calculates the difference between this geometric connectivity weight and the set retention threshold. The status bit detection register reads the sign bit of this difference data and extracts adjacent vertices with negative differences. A negative difference indicates that the spatial distribution characteristics of this vertex tend to extend along the normal fiber physical direction, and the node filtering logic circuit determines that the adjacent vertex is a normal fiber interference node based on this logic. After receiving the interference determination command, the graph theory pruning accelerator directly erases the normal fiber interference node and all its connected spatial connectivity edges from the memory address. The graph theory pruning accelerator thoroughly filters out normal fiber interference nodes in the spatial topology graph, retaining only the effective microstructure mesh data with lateral tearing characteristics.

[0075] Example 6

[0076] Based on the above embodiments, this embodiment addresses step S5, where the process of reconstructing the boundary contour of the micro-cracked region along the connected paths between the remaining network vertices and outputting the physical coordinate sequence of the asbestos cement structure defects is as follows:

[0077] After filtering out normal fiber interference nodes, the specific execution process for reconstructing the boundary contour of the micro-cracked region relies on a dedicated contour reconstruction coprocessor. This coprocessor reads the connected path data between the remaining network vertices stored in the static random access memory via a high-speed internal data bus. The contour tracing logic unit uses the geometric extremum nodes on the connected path as starting points and performs Moiré neighborhood tracing operations along the adjacent pixels outside the connected path. The tracing operation strictly slides along the outer connected edge of the pixel node, sequentially extracting the discrete node array that constitutes the outermost contour line. A floating-point calculator reads the spatial coordinates of the first and last nodes at both ends of the outermost contour line. The floating-point calculator calculates the linear Euclidean distance between the first and last nodes using the square root of the sum of the squares of the coordinate differences. The calculation process of this first and last node distance satisfies the node distance evaluation formula: ; where, in the formula Indicates the distance between the first and last nodes. This represents the horizontal coordinate component of the first node in the dual-modal pixel coordinate system. This represents the vertical coordinate component of the first node in the dual-modal pixel coordinate system. This represents the horizontal coordinate component of the tail node in the dual-modal pixel coordinate system. This represents the vertical coordinate component of the tail node in the dual-modal pixel coordinate system.

[0078] After obtaining the distance between the first and last nodes, the parallel logic comparison gate array reads the closure threshold stored in the read-only memory. The parallel logic comparison gate array performs a parallel comparison operation between the calculated distance between the first and last nodes and the closure threshold. When the comparison result shows that the distance between the first and last nodes is less than or equal to the closure threshold, it indicates that the two ends of the cracked region exhibit a physical trend of microscopic fusion. The morphological closure circuit is activated and directly generates a virtual connection between the first and last nodes, thereby forcibly closing the entire outermost contour line. When the comparison result shows that the distance between the first and last nodes is greater than the closure threshold, it indicates that the cracked region belongs to a typical deep linear extended tear crack. The morphological closure circuit remains silent and directly preserves the original extended shape. Based on the above comparison judgment result of adaptive closure or preservation of extended shape, the dedicated contour reconstruction coprocessor directly reconstructs the complete boundary contour representing the microscopic cracked region in the video memory space.

[0079] The specific calculation process for extracting the physical coordinate sequence relies on the physical coordinate transformation engine. The coordinate extraction microcontroller traverses the boundary contour data within the video memory space, extracting the pixel coordinates of the corresponding pixel nodes one by one. The physical coordinate transformation engine obtains the spatial transformation matrix of the dual-modal pixel coordinate system from the calibration parameter storage area through the direct memory access controller. This spatial transformation matrix is ​​obtained through physical perspective transformation during the initial calibration phase of the multi-source sensing device and contains the scaling factor and translation parameters of the physical dimension. The matrix multiplication accelerator converts the extracted pixel coordinates into a homogeneous coordinate column vector and performs matrix multiplication on the spatial transformation matrix and the homogeneous coordinate column vector. The process of the matrix multiplication accelerator using the spatial transformation matrix to calculate the pixel coordinates satisfies the spatial coordinate mapping formula: ; where, in the formula This represents the horizontal physical coordinates of the structural defects generated by the solution in the real physical environment. This represents the vertical physical coordinates of the structural defects generated by the solution in the real physical environment. This represents the spatial transformation matrix containing calibration parameters. This represents the horizontal pixel column coordinates of the pixel node corresponding to the boundary contour. This represents the vertical pixel row coordinates of the pixel nodes corresponding to the boundary contour. The physical coordinate transformation engine concatenates and packages all the calculated horizontal and vertical physical coordinates according to the node tracking order, and finally outputs the physical coordinate sequence of the asbestos cement structure defects directly.

[0080] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for quality testing of asbestos cement based on multi-source sensing, characterized in that, Includes the following steps: S1: Acquire high-resolution morphological and thermodynamic images of asbestos cement specimens, and construct an initial structural distribution matrix by aligning the dual-modal pixel coordinate system; S2: Extract the local pixel orientation gradient of the topography image in the initial structure distribution matrix and construct a tensor feature map representing the interlacing direction of asbestos fibers; S3: Project the thermodynamic image in the matrix onto the tensor feature map, calculate the local heat conduction attenuation and the offset of fiber orientation, and locate the cross-modal thermal anomaly node; S4: Establish a spatial topology graph with cross-modal thermal anomaly nodes as network vertices, calculate the geometric connectivity weights between adjacent vertices, and filter out normal fiber interference nodes; S5: Reconstruct the boundary contour of the micro-cracked region along the connected paths between the remaining network vertices, and output the physical coordinate sequence of the defects in the asbestos cement structure.

2. The method for quality detection of asbestos cement based on multi-source sensing according to claim 1, characterized in that, In step S1, the process of aligning the dual-modal pixel coordinate system to construct the initial structure distribution matrix is ​​as follows: The structured light camera and infrared thermal imager are used to simultaneously scan the surface of the asbestos cement specimen, and high-resolution morphological and thermodynamic images of the asbestos cement specimen are acquired. Extract matching feature points from high-resolution topographic images and thermodynamic images, perform spatial mapping to align the dual-modal pixel coordinate system based on the matching feature points, and merge pixel channels to construct an initial structure distribution matrix.

3. The method for quality detection of asbestos cement based on multi-source sensing according to claim 1, characterized in that, In step S2, the process of constructing the tensor feature map characterizing the interlacing orientation of asbestos fibers is as follows: Spatial difference operation is performed on the topography image in the initial structure distribution matrix to extract the gray-level derivatives of pixel nodes in the topography image on the orthogonal axis and synthesize the local pixel orientation gradient. Calculate the outer product matrix of the gradient of local pixel directions, perform spatial neighborhood integration on the outer product matrix, and aggregate the matrix elements after integration into a tensor feature map representing the interlacing direction of asbestos fibers.

4. The method for quality detection of asbestos cement based on multi-source sensing according to claim 3, characterized in that, The specific method for extracting the gray-level derivatives of pixel nodes within the topography image on orthogonal axes and synthesizing local pixel directional gradients is as follows: By performing sliding convolution on the topography image of the initial structure distribution matrix using horizontal and vertical difference operators, the gray-level derivatives of pixel nodes in the topography image on the orthogonal axis are extracted. The gradient magnitude is extracted by summing the squares of the gray-level derivatives on the orthogonal axes and taking the square root. The gradient direction is extracted by calculating the arctangent of the quotient of the gray-level derivatives on the orthogonal axes. The gradient magnitude and gradient direction are combined to synthesize the local pixel direction gradient.

5. The method for quality detection of asbestos cement based on multi-source sensing according to claim 1, characterized in that, In step S3, the process of locating the cross-modal thermal anomaly node is as follows: Based on the pixel coordinates of the initial structure distribution matrix, bilinear interpolation is performed on the thermodynamic image in the matrix to project the thermodynamic image onto the spatial grid of the tensor feature map; Calculate the temperature gradient of the thermodynamic image in the spatial grid, generate a local heat conduction attenuation vector, analyze the principal direction vector of the tensor feature map as the fiber orientation, and calculate the angle offset between the two. Calculate the offset difference between the included angle offset and the set reference offset threshold, extract the spatial grid nodes with positive offset differences, and locate them as cross-modal thermal anomaly nodes.

6. The method for quality testing of asbestos cement based on multi-source sensing according to claim 1, characterized in that, In step S4, the process of calculating the geometric connectivity weights between adjacent vertices and filtering out normal fiber interference nodes is as follows: The located cross-modal thermal anomaly nodes are mapped to network vertices, the spatial coordinates of the network vertices are extracted and neighborhood connection operations are performed to establish a spatial topology graph containing adjacent vertices. Extract the Euclidean distance between adjacent vertices in the spatial topology graph, calculate the sine value of the angle between the direction of the connection and the fiber direction, and divide the sine value by the sum of the Euclidean distance and a preset constant to generate the geometric connectivity weight. Calculate the difference between the geometric connectivity weight and the set retention threshold, extract adjacent vertices with negative differences and determine them as normal fiber interference nodes, and filter out normal fiber interference nodes in the spatial topology graph.

7. The method for quality detection of asbestos cement based on multi-source sensing according to claim 6, characterized in that, The specific content of establishing a spatial topology graph containing adjacent vertices is as follows: The data indexes of the located cross-modal thermal anomaly nodes are mapped to network vertices. The position components of the network vertices in the bimodal pixel coordinate system are read, and the spatial coordinates of the network vertices are extracted. Based on spatial coordinates, extract adjacent vertices whose distance is lower than the connected boundary, perform neighborhood connection operations on adjacent vertices to generate spatial connected edges, and aggregate the spatial connected edges and network vertices to establish a spatial topology graph containing adjacent vertices.

8. The method for quality detection of asbestos cement based on multi-source sensing according to claim 1, characterized in that, In step S5, the process of outputting the physical coordinate sequence of defects in the asbestos cement structure is as follows: Read the connected paths between the remaining network vertices, perform tracing operations along the outside of the connected paths to extract the outermost contour line, and based on the comparison results of the distance between the first and last nodes and the closure threshold, adaptively close or retain the extended shape to reconstruct the boundary contour of the micro-cracked region. Extract the pixel coordinates of the pixel nodes corresponding to the boundary contour, obtain the spatial transformation matrix of the dual-modal pixel coordinate system, use the spatial transformation matrix to solve the pixel coordinates, and output the physical coordinate sequence of the defects in the asbestos cement structure.

Images