Damage detection method and system based on machine vision

Abstract

The invention discloses a damage detection method and system based on machine vision, and relates to the technical field of machine vision detection. A polarization structured light-thermal imaging three-mode acquisition framework of a customized machine vision array is adopted, the problems of composite noise superposition and unstable internal thermal signal diffusion caused by dynamic heterogeneity of a target surface are solved, the adaptability of damage detection to a complex environment is greatly improved, and by means of a space attention mechanism and a cascade analysis network, the damage detection accuracy is improved. According to the method, geometric measurement and heat conduction simulation technologies are combined, the problems that damage space distribution is rough in description, and there is no stable corresponding relation between surface and deep abnormity are solved, accurate space mapping and multi-dimensional quantification of damage from the surface to the interior are achieved, and through linkage of a dual optimization mechanism and an edge end improved MobileViT lightweight model, the problem that damage space distribution is rough in description is solved. Pain points where real defects and interference artifacts are difficult to distinguish are broken, stable correspondence between surface signals and deep damage is enhanced, and accuracy and practicability of damage detection in an industrial scene are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection technology, specifically a damage detection method and system based on machine vision. Background Technology

[0002] Damage detection is a fundamental engineering practice for ensuring the safe operation of large structures. Its importance lies in its direct relation to the long-term stability and accident prevention of bridge construction and aviation equipment. Although current damage detection methods have been widely applied to various imaging techniques, they generally suffer from insufficient adaptability to complex environments and coarse characterization of damage spatial distribution, making it difficult to reliably distinguish between real defects and interference artifacts in actual industrial scenarios.

[0003] The core technical challenge in this field lies in the dynamic heterogeneity of the target surface state and the non-uniform diffusion characteristics of the internal thermal signal analysis. The dynamic heterogeneity of the surface state means that changes in illumination and vibration interference will continuously superimpose on the original signal, forming a complex noise layer that is difficult to peel off. This noise layer further amplifies the instability of the non-uniform diffusion of the internal thermal signal, because the heat conduction process is inherently affected by the internal microstructure of the material, exhibiting local gradient differences, which makes it impossible to establish a stable correspondence between the surface observation signal and deep anomalies. For example, in the inspection of bridge steel beams, surface temperature fluctuations under direct sunlight can mask weak heat leakage signals at cracks, making it impossible for operators to accurately determine on-site whether the crack has penetrated more than 5 mm into the material. Therefore, how to simultaneously acquire surface differences and internal thermal signals from multiple perspectives while achieving dynamic noise filtering and accurate spatial mapping of deep anomalies has become a key issue in improving the reliability of damage detection. Summary of the Invention

[0004] The purpose of this invention is to provide a damage detection method and system based on machine vision, which solves the problems of noise interference in complex environments, rough damage characterization, and difficulty in distinguishing defects from artifacts in damage detection, and significantly improves the environmental adaptability, accuracy and industrial reliability of the detection.

[0005] The objective of this invention can be achieved through the following technical solutions: This application provides a machine vision-based damage detection method, comprising the following steps: S1. By using a customized machine vision array unit and a three-modal synchronous acquisition architecture of polarization structured light and thermal imaging, the target surface polarization texture feature image, structured light encoded 3D contour point cloud data, and internal thermal conduction visual difference signal are acquired simultaneously to obtain a multi-dimensional raw dataset. S2. Based on a multi-dimensional original dataset, the optical flow algorithm and the damage-oriented visual mask reconstruction network are integrated to track pixel-level three-modal dynamic changes. At the same time, the invalid background information is filtered out through self-supervised visual feature distillation technology to capture local gradient differences related to damage and generate a surface-interior integrated heterogeneous feature map. S3. Calculate the gradient difference variance of the heterogeneous feature map. If the variance exceeds the preset threshold, perform dynamic noise filtering through the damage-guided visual mask reconstruction network to output a denoised multimodal fusion image with a high signal-to-noise ratio. S4. Based on the denoised multimodal fusion image, the suspected damage area is located through the spatial attention mechanism, and then the damage area is quantitatively represented through the cascaded damage analysis network. S5. Based on the quantization results of the damaged area, the machine vision geometric measurement module is called. The structured light encoding 3D contour point cloud data is used to calculate the core quantization parameters of the damage. At the same time, the internal heat conduction path of the damaged area is simulated by combining the thermal conduction visual difference signal. The mapping relationship between the polarization 3D thermal damage triad and the internal damage is established to obtain the deep anomaly distribution estimation results. S6. If the correspondence between the polarization 3D thermal damage triad estimated by the mid-to-deep anomaly distribution and the internal damage is lower than the preset threshold, a dual optimization mechanism is activated. At the same time, the improved lightweight MobileViT model at the edge is linked to strengthen the stable correspondence between the surface observation signal and the deep damage under the constraint of millisecond-level inference speed, and output the damage space mapping result.

[0006] This invention provides a machine vision-based damage detection system for implementing a machine vision-based damage detection method, comprising: The data acquisition module, through a customized machine vision array unit, adopts a three-modal synchronous acquisition architecture of polarization structured light and thermal imaging to simultaneously acquire polarization texture feature images of the target surface, structured light encoded 3D contour point cloud data, and internal thermal conduction visual difference signals to obtain a multi-dimensional raw dataset. The feature extraction and fusion module, based on a multi-dimensional original dataset, integrates optical flow algorithm and damage-guided visual mask reconstruction network to track pixel-level three-modal dynamic changes. At the same time, it filters invalid background information through self-supervised visual feature distillation technology, captures local gradient differences related to damage, and generates a surface-interior integrated heterogeneous feature map. The noise filtering module calculates the gradient difference variance of the heterogeneous feature map. If the variance exceeds the preset threshold, dynamic noise filtering is performed through the damage-guided visual mask reconstruction network. Based on the feature priority of self-supervised distillation, the damage-related polarization texture, 3D contour, and thermal conduction signal are preserved, and a denoised multimodal fusion image with high signal-to-noise ratio is output. The damage localization and analysis module, based on the denoised multimodal fusion image, locates the suspected damage area through a spatial attention mechanism, and then performs quantitative characterization of the damage area through a cascaded damage analysis network. The damage assessment and report generation module, based on the quantification results of the damage area, calls the machine vision geometric measurement module, uses structured light to encode 3D contour point cloud data, calculates the core quantification parameters of the damage, and combines the thermal conduction visual difference signal to simulate the internal thermal conduction path of the damage area, establishes the mapping relationship between the polarization 3D thermal damage triad and the internal damage, and obtains the deep anomaly distribution estimation results. The dual optimization module activates a dual optimization mechanism if the correspondence between the polarization 3D thermal damage triad estimated by the mid-to-deep anomaly distribution and the internal damage is lower than a preset threshold. Simultaneously, it links with the improved lightweight MobileViT model at the edge to enhance the stable correspondence between surface observation signals and deep damage under millisecond-level inference speed constraints, and outputs damage space mapping results.

[0007] The beneficial effects of this invention are as follows: By adopting a customized machine vision array with a polarization structured light-thermal imaging trimodal synchronous acquisition architecture, and integrating optical flow algorithm, damage-guided visual mask reconstruction network and self-supervised visual feature distillation technology, the problem of superposition of illumination and vibration composite noise caused by the dynamic heterogeneity of the target surface state in the background technology, as well as the problem of non-uniform diffusion instability of internal thermal signal amplified by this noise, is effectively solved. Dynamic noise filtering in complex environments is achieved, local gradient differences related to damage are successfully captured, and surface-internal integrated heterogeneity feature map is generated, which significantly improves the adaptability of damage detection to complex environments. By leveraging spatial attention mechanisms to locate suspected damage areas, and employing a cascaded damage analysis network to achieve pixel-level precise localization and quantification of the damage area, combined with machine vision geometric measurement modules and thermal conduction path simulation technology, a mapping relationship between the polarization 3D thermal damage triad and internal damage is established. This solves the problems of coarse damage spatial distribution characterization and lack of stable correspondence between surface observation signals and deep anomalies in the background technology, accurately calculates the core quantification parameters of the damage, and obtains the estimation results of the deep anomaly distribution, realizing accurate spatial mapping and multi-dimensional quantification of damage from the surface to the interior. By activating a dual optimization mechanism and linking the improved MobileViT lightweight model at the edge, the stable correspondence between surface observation signals and deep damage is enhanced under the constraint of millisecond-level inference speed. This solves the problems of difficulty in reliably distinguishing between real defects and interference artifacts and insufficient reliability of damage detection in the background technology. It effectively improves the correspondence strength between the polarization 3D thermal damage triad and internal damage, and finally outputs highly robust damage space mapping results, ensuring the accuracy and reliability of damage detection in industrial scenarios. Attached Figure Description

[0008] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0009] Figure 1 A schematic flowchart of a machine vision-based damage detection method provided in Embodiment 1 of this application; Figure 2 This is a flowchart illustrating step S2 in a machine vision-based damage detection method provided in Embodiment 1 of this application. Figure 3 This is a flowchart illustrating step S4 in a machine vision-based damage detection method provided in Embodiment 1 of this application. Figure 4 This is a schematic diagram of a machine vision-based damage detection system provided in Embodiment 2 of this application. Detailed Implementation

[0010] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0011] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0012] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0013] Example 1 Please see Figures 1-3 This embodiment provides a machine vision-based damage detection method, including the following steps: S1. By using a customized machine vision array unit and a three-modal synchronous acquisition architecture of polarization structured light and thermal imaging, the target surface polarization texture feature image, structured light encoded 3D contour point cloud data, and internal thermal conduction visual difference signal are simultaneously acquired for the detection target (such as bridge components, mechanical parts, etc.) to obtain a multi-dimensional raw dataset. The structured light module can dynamically adapt to different sizes of detection objects without additional adjustment of visual acquisition parameters. The dataset naturally contains superimposed signals common in engineering scenarios such as sudden changes in illumination, vibration interference, and surface stains, which meet the needs of machine vision engineering applications. Furthermore, step S1 specifically includes: By using a customized machine vision array and a three-modal synchronous acquisition architecture of polarization structured light and thermal imaging, polarization texture images, structured light encoded three-dimensional contour point clouds and thermal conduction difference signals are acquired to generate multi-dimensional raw datasets. The polarization texture image is processed by a preset polarization filtering algorithm to extract the polarization features of the target surface and generate the first feature dataset. The structured light encoded data is then processed by a structured light decoding algorithm to generate a three-dimensional contour point cloud and obtain a spatial geometry dataset. Image signal analysis methods are used to extract heat conduction features and generate a second feature dataset. Through modal fusion processing, the first feature dataset, spatial geometry dataset, and second feature dataset are integrated into a multimodal dataset to generate a comprehensive feature dataset.

[0014] Among them, the preset polarization filtering algorithm adopts an adaptive polarization direction calibration strategy. First, the intensity normalization and phase correction are performed on the texture images collected at different polarization angles. Then, the polarization signal noise caused by sudden changes in illumination and surface stains is suppressed by combining Gaussian kernel and directional filtering. The inherent polarization characteristics of the target surface and environmental interference components are accurately separated. Finally, the core polarization features such as texture distortion and refractive index difference related to material damage are extracted to form the first feature dataset. The structured light decoding algorithm first dynamically matches the size of the object to be detected without adjusting the acquisition parameters. Then, it performs grayscale threshold segmentation and sub-pixel edge extraction on the acquired coded pattern. The spatial phase distribution of the coded pattern is restored by the phase unwrapping algorithm. Combined with the calibration parameters (intrinsic and extrinsic parameters) of the machine vision array, the mapping relationship between the coded points and the three-dimensional spatial coordinates is established. At the same time, the encoding offset error caused by vibration interference is corrected. Finally, a three-dimensional contour point cloud that can accurately restore the surface undulations and geometric shape of the target is generated, forming a spatial geometric dataset. When using image signal analysis methods, the heat conduction difference signals are first denoised in the time domain and smoothed in the spatial domain to suppress pseudo-thermal signals caused by sudden changes in illumination and vibration interference. Then, through heat diffusion rate calculation and temperature gradient tensor decomposition, core features such as heat conduction uniformity, abnormal hot spot area / intensity, and heat diffusion path curvature are extracted. At the same time, the relative differences in heat conduction coefficients in different regions are calibrated to accurately capture internal heat conduction anomalies caused by damage. Finally, these quantitative features are organized into a second feature dataset.

[0015] Specifically, by using a customized three-modal synchronous acquisition architecture, coupled with adaptive polarization filtering, dynamically adaptable size structured light decoding, and thermal conduction signal analysis algorithms, the system effectively solves the problems of noise interference such as sudden changes in illumination, vibration interference, and surface stains in engineering scenarios, as well as the adaptation of different sized detection objects. At the same time, it integrates scattered multimodal features to generate a comprehensive feature dataset containing core information on polarization, geometry, and thermal conduction, which meets the needs of engineering applications and provides high-quality data support for subsequent damage detection.

[0016] S2. Based on a multi-dimensional original dataset, the optical flow algorithm and the damage-oriented visual mask reconstruction network are integrated to track pixel-level three-modal dynamic changes. At the same time, the invalid background information is filtered out through self-supervised visual feature distillation technology to capture local gradient differences related to damage and generate a surface-interior integrated heterogeneous feature map. Furthermore, step S2 specifically includes: S21. Obtain a multi-dimensional raw dataset, calculate pixel-level motion vectors using an optical flow algorithm to obtain three-modal dynamic change information, and use a visual mask reconstruction network to process the three-modal dynamic change information to generate a preliminary surface feature map; the three-modal dynamic change information includes polarization texture differences, 3D contour undulations, and thermal conduction gradients; Specifically, optical flow algorithms are used to process polarization texture image sequences, 3D contour point cloud frame-to-frame data, and thermal conduction visual difference signal sequences from the multi-dimensional original dataset. For polarization texture images, dense optical flow is calculated to capture pixel-level texture motion trajectories to reflect dynamic changes in polarization characteristics. For 3D contour point clouds, inter-frame corresponding point optical flow vector matching is used to quantify the undulation displacement and morphological dynamics of the contour surface. For the thermal conduction signal sequence, optical flow analysis is used to analyze the spatiotemporal movement trend of temperature pixels and extract the gradient change law of heat diffusion. Finally, the pixel-level motion vectors of the three modalities are integrated to form complete three-modal dynamic change information including polarization texture differences, 3D contour undulations, and thermal conduction gradients, providing accurate dynamic feature support for subsequent feature map reconstruction. When using a visual mask reconstruction network to process trimodal dynamic change information, the network's dynamic mask generation module first adaptively generates pixel-level weighted masks based on abrupt changes in polarization texture differences, abnormal inflection points of 3D contour undulations, and steep bands of thermal conduction gradients. High weights are assigned to dynamic feature regions potentially related to damage (such as texture distortion, contour break points, and hot spot diffusion fronts), while low weights are assigned to stable background regions. Then, the cross-modal feature reconstruction module spatially aligns and fuses the weighted trimodal dynamic information, using residual convolutional layers to enhance the spatial representation of local dynamic differences. Finally, a preliminary surface feature map integrating polarization, geometry, and thermal conduction dynamic features is output, clearly presenting the dynamic changes of the target surface and the regions associated with potential damage.

[0017] S22. By using self-supervised visual feature distillation technology, invalid background information is filtered out from the preliminary surface feature map to obtain a refined feature map. By constructing a multimodal feature fusion mechanism, local gradient differences related to damage are captured from the refined feature map, including surface texture distortion, 3D contour breakage and heat conduction anomaly, to generate a gradient feature map containing damage-related information. S23. If there are significant local gradient differences in the gradient feature map, the damage area is determined by pixel-level analysis, a damage location map is generated, and then the damage location map and the refined feature map are fused to generate a heterogeneous feature map that integrates the surface and the interior.

[0018] The surface texture distortion includes: using a group of convolutional kernels enhanced with histogram of oriented gradients to extract micron-level texture features from the shallow network of the feature map; then adaptively tracking the texture break lines through a deformable convolutional module to calculate the gradient direction dispersion of adjacent pixels; and establishing a Markov random field model to analyze statistical anomalies in the texture pattern and identify early material degradation that is difficult to detect with the naked eye. The 3D contour breakage includes: embedding a depth-sensing attention module in the middle layer network of the feature map to align the two-dimensional features with the point cloud data collected by the time-of-flight camera; calculating the second-order differential features of the surface curvature through the normal vector field to detect sub-millimeter-level contour discontinuities; and using a graph convolutional network to construct a three-dimensional topological relationship to analyze the mesh distortion pattern in the stress concentration region. The heat conduction anomalies include: fusing temperature distribution data from an infrared thermal imager in the high-dimensional space of the feature map; constructing a lightweight solver for the Fourier heat conduction equation using prior thermodynamic knowledge; and using a spatiotemporal graph neural network to model the dynamic propagation process of the temperature gradient and identify anomalous heat accumulation regions.

[0019] Furthermore, a gradient feature map containing damage-related information is generated, specifically including: Using refined feature maps as input, the pixel-level texture distortion features of the texture flow are captured by deformable convolution adaptively, the three-dimensional geometric fracture features of the contour flow are extracted by the depth-aware attention module, and the abnormal thermal diffusion features of the thermal conduction flow are captured by the thermodynamic graph network to obtain multimodal features. Multimodal features are fed into a cascaded interaction module for deep processing. First, a spatial transformer network is used to achieve sub-pixel-level feature alignment to ensure spatial consistency of different modal features. Then, a gated attention mechanism is used for multiple rounds of cross-validation to ensure that the features of each modality corroborate each other. Finally, a feature enhancement network is used to spatially enhance the validated damage areas to obtain validated high-quality features. Based on the verified high-quality features, a dynamic weighted fusion and reconstruction process is initiated. The modality confidence evaluation network is used to calculate the weight coefficients of each modality to determine the adaptive weight allocation. Then, a multi-scale feature pyramid is used to fuse cross-resolution features, retaining damage information at different scales. The transposed convolutional layer is used to reconstruct the spatial dimension of the feature map, restore the spatial details of the feature map, and generate a fused feature map that preserves details and semantic integrity. After obtaining the fused feature map, guided gradient filtering is applied to enhance the continuity of the damage boundary and improve the quality of edge features. A nonlinear activation function is used to highlight significant gradient differences and strengthen the response of key features. Then, a channel attention mechanism is used to optimize the feature expression and improve the feature discrimination. The output is a gradient feature map containing rich damage information, providing high-quality input for subsequent damage recognition tasks.

[0020] Specifically, by integrating optical flow algorithm, damage-guided visual mask reconstruction network and self-supervised feature distillation technology, combined with multimodal feature fusion mechanism and refined gradient feature extraction method, the problem of difficulty in tracking three-modal dynamic features, many invalid background interferences, fragmentation of surface and internal damage features and difficulty in accurately capturing damage-related local gradient differences is effectively solved. The resulting heterogeneous feature map highlights key damage features and improves feature discrimination, providing high-quality feature support for subsequent damage localization and analysis.

[0021] S3. Calculate the gradient difference variance of the heterogeneous feature map. If the variance exceeds the preset threshold (i.e., noise interference affects the damage recognition accuracy), perform dynamic noise filtering through the damage-guided visual mask reconstruction network. Based on the feature priority of self-supervised distillation, selectively remove artifact components such as light abrupt changes, vibration interference, and surface stain superposition, while retaining the damage-related polarization texture, 3D contour, and core heat conduction signal, and output a denoised multimodal fusion image with a high signal-to-noise ratio. Furthermore, step S3 specifically includes: Heterogeneous features are extracted from the input image. The gradient operator is used to calculate the gradient difference of each feature map to obtain the gradient difference matrix. Then, the statistical variance of the gradient difference matrix is ​​calculated, and it is determined whether the variance value exceeds the preset threshold. If it does, it is determined that noise filtering needs to be performed. The process employs gradient operators (such as the Sobel operator) to calculate pixel-level gradients for each modal feature component (polarization texture, 3D contour, and thermal conduction features) in the heterogeneous feature map. Convolution operations are used to extract the gradient magnitude and direction of each pixel in the horizontal and vertical directions. This gradient information from different directions is then integrated into a gradient difference matrix. The element values ​​in this matrix directly reflect the drastic changes in local areas of the feature map. Damaged areas are typically accompanied by high gradient values, and noise interference can also cause abnormal gradient fluctuations. Subsequently, the statistical variance of this matrix is ​​calculated to quantify the dispersion of the overall gradient values. A larger variance indicates a more uneven gradient distribution in the feature map, suggesting potentially strong noise interference or significant damage signals, providing a quantitative basis for subsequent decisions on whether to activate dynamic noise filtering.

[0022] Based on the variance value, a damage-oriented model is obtained, a visual mask is generated, and the feature map set is filtered using the mask to obtain a filtered feature set. The filtered feature set is then processed by a reconstruction network, and a convolutional neural network is used to perform dynamic noise filtering to output a denoised multimodal fusion image with a high signal-to-noise ratio.

[0023] Based on the judgment result that the variance value exceeds the preset threshold, the pre-trained damage-oriented model is first invoked. Based on the feature priority of self-supervised distillation, the distribution law of damage-related core signals (polarization texture, 3D contour, thermal conduction features) is learned. Then, a pixel-level visual mask is adaptively generated. High-gradient feature regions that are highly correlated with damage are given high activation weights, while low-weight or even suppression weights are given to artifact regions caused by sudden illumination, vibration interference, and surface stains. Then, the mask is used to screen the heterogeneous feature map pixel by pixel, retaining the high-weight damage core feature components and removing low-weight noise artifacts to obtain a screened feature set that focuses on damage information. Finally, the screened feature set is input into the reconstruction network for feature completion and spatial alignment.

[0024] Specifically, the degree of noise interference is accurately determined by gradient difference variance calculation. With the help of damage-guided visual mask reconstruction network, self-supervised distillation feature priority screening and convolutional neural network dynamic filtering, the artifact noise problems of sudden illumination changes, vibration interference and surface stain superposition in heterogeneous feature maps are effectively solved. Invalid interference components are successfully removed, damage-related core signals are completely preserved, and a multimodal fusion image with high signal-to-noise ratio is output, clearing the noise obstacle for subsequent accurate damage identification.

[0025] S4. Based on the denoised multimodal fusion image, the suspected damage area is located through the spatial attention mechanism, and then the damage area is quantitatively represented through the cascaded damage analysis network. Furthermore, step S4 specifically includes: S41. Construction of a spatial attention-guided damage localization network: A dual-path network architecture based on coordinate attention mechanism is constructed. The coordinate attention module of the spatial path captures the accurate spatial coordinates of the suspected damage area. At the same time, the channel recalibration strategy of the channel path is used to enhance the signal-to-noise ratio of the damage features. The output provides high-quality candidate regions for subsequent analysis and avoids invalid calculations in the entire map. S42, Cascaded Damage Analysis Architecture Implementation: Based on the localized candidate region, a two-level cascaded damage analysis network is established. The first-level network achieves pixel-level precise localization of the damage boundary within the candidate region through deformable convolution and feature pyramid fusion. The second-level network, based on localization, introduces a multi-scale receptive field module and combines it with 3D point cloud registration technology to extract the depth distribution and 3D morphological features of the damage region from coarse to fine. The two-level network ensures the continuity of feature transfer through residual connections. S43. Multi-dimensional quantitative analysis model execution: Based on the depth distribution and three-dimensional morphological features of the extracted damage area, implement physical feature-based quantitative analysis. Calculate the surface curvature change through the normal vector field, and combine the thermal conduction anomaly area segmentation results to construct a complete descriptor containing geometric parameters and thermophysical parameters. Transform the morphological features extracted in the previous step into quantifiable engineering parameters to provide data support for damage assessment. S44. Comprehensive Damage Assessment and Report Generation: Based on quantitative analysis results, a graph neural network is used to aggregate the correlation between geometric deformation and thermodynamic characteristics. By establishing matching rules between damage type and morphological features, and combining multi-dimensional quantitative parameters, a structured assessment report containing spatial coordinates, three-dimensional contour parameters, damage type classification, and severity level is output, realizing a complete closed loop from feature extraction to engineering decision-making.

[0026] Among them, the coordinate attention dual-path network can efficiently screen out high-value candidate regions and greatly improve the efficiency of subsequent analysis by accurately locking the coordinates of suspected damage through spatial paths and enhancing the signal-to-noise ratio of features through channel paths. In the cascaded damage analysis network, deformable convolution adapts to irregular damage boundaries, multi-scale receptive field modules cover the needs of damage feature extraction from micrometer to millimeter level, and three-dimensional point cloud registration technology and residual connection together ensure the accuracy and continuity of damage three-dimensional morphological feature transmission. The geometric parameters of multidimensional quantitative analysis include damage area, depth, aspect ratio, etc., and the thermophysical parameters include differences in thermal conductivity, duration of abnormal hot spots, etc. The complete descriptor can comprehensively characterize the physical properties of the damage. Graph neural networks aggregate features from multiple dimensions, including spatial correlation, morphological similarity, and thermophysical consistency. The matching rules between damage type and morphological features are generated based on a large number of actual damage samples such as bridge components and mechanical parts, ensuring the engineering practicality and decision-making reliability of the structured assessment report, and adapting to the damage detection needs of different sizes and complex environments in industrial scenarios throughout the entire process.

[0027] Specifically, by employing a coordinate attention dual-path network, a two-level cascaded damage analysis architecture, a multi-dimensional quantitative analysis model, and graph neural network evaluation technology, the system effectively addresses the problems of low damage localization efficiency, inaccurate boundary and 3D morphology extraction, single quantification dimension, and lack of engineering evaluation closed loop. It achieves efficient screening of suspected damage areas, pixel-level accurate localization of damage boundaries, and multi-dimensional quantification of 3D morphology and thermophysical parameters. Finally, it outputs a structured report containing spatial coordinates, damage type, and severity, forming a complete closed loop from feature extraction to engineering decision-making, providing reliable data support for damage assessment.

[0028] S5. Based on the quantization results of the damaged area, the machine vision geometric measurement module is invoked. Using structured light to encode 3D contour point cloud data, the core quantization parameters of the damage are calculated, including depth, area, extension angle, and volume. At the same time, the internal heat conduction path of the damaged area is simulated by combining the thermal conduction visual difference signal. The mapping relationship between the polarization 3D thermal damage triad and the internal damage (such as deep corrosion and hidden crack propagation) is established to obtain the deep anomaly distribution estimation results, realizing the integrated output from damage identification to accurate measurement. The polarization 3D thermal damage triad includes polarization texture, 3D structured light, and thermal imaging data. Furthermore, step S5 specifically includes: S51. Based on the damage area localization results, call the machine vision geometric measurement module to perform refined three-dimensional reconstruction of the identified damage area. Through the point cloud data encoded by structured light, establish a digital geometric model of the damage area to provide accurate spatial reference and geometric constraints for subsequent heat conduction analysis and multimodal fusion. S52. Based on the digital geometric model of the damaged area, the thermophysical parameters of the implanted material are used to construct a numerical model of heat conduction in the damaged area. By comparing the differences between the theoretical heat conduction path and the measured thermal imaging data, the abnormal heat conduction area caused by internal damage is identified, providing thermodynamic evidence to support multimodal correlation. S53. Based on geometric and thermal conduction features, a deep correlation model of polarization-3D-thermal damage triad is established. Through multimodal data registration and feature alignment, the statistical correlation between surface geometric deformation, polarization texture anomaly and thermal conduction anomaly is explored, forming a mapping rule for inferring internal damage from surface features. S54. Construct a comprehensive assessment system for damage status. Based on geometric parameters, thermal conduction anomalies, and multimodal correlation features, use machine learning models to quantitatively assess internal damage and generate deep anomaly distribution estimation results that include damage geometric parameters, internal damage probability, and evolution trend, thus completing a complete analysis chain from surface identification to internal assessment.

[0029] Among them, the machine vision geometric measurement module completes the fine three-dimensional reconstruction of the damaged area through noise reduction, registration and meshing of point cloud data. The digital geometric model accurately restores the undulation and spatial shape of the damaged surface. The implanted material thermal property parameters include key parameters that match the material of the detection target (such as bridge components and mechanical parts), such as thermal conductivity, specific heat capacity and density. Multimodal data registration uses the Iterative Closest Point (ICP) algorithm based on feature points to achieve spatial alignment of polarization texture, 3D point cloud and thermal imaging data. Feature alignment focuses on core related features through attention mechanism, such as the correspondence between polarization distortion at surface cracks, 3D depression and heat conduction blockage. The statistical correlation of the deep correlation model is specifically reflected in the quantitative mapping of surface geometric deformation degree, polarization texture anomaly coefficient and heat conduction path offset. The machine learning model uses a lightweight neural network (such as a CNN-LSTM hybrid model) trained on a large number of damage samples. It can accurately learn the correspondence between surface features and internal damage (deep corrosion, latent cracks). In the final deep anomaly distribution estimation results, the internal damage probability is calculated based on the collaborative confidence of multimodal features. The evolution trend is combined with the material fatigue characteristics and damage geometric parameters for prediction. The entire process uses a digital geometric model as the spatial benchmark and a thermal conduction numerical model as the thermodynamic support to ensure that the analysis chain from surface identification to internal assessment is accurate and reliable, and adapts to the needs of comprehensive damage assessment in engineering scenarios.

[0030] Specifically, by accurately calculating the core geometric parameters of damage through the machine vision geometric measurement module, and combining the thermal conduction numerical model with multimodal deep correlation, it effectively solves the problems of fuzzy correlation between surface damage and internal damage (deep corrosion, hidden cracks, etc.), difficulty in quantifying and assessing internal damage, and lack of a complete chain from surface identification to internal assessment. It generates deep anomaly distribution estimation results that include damage geometric parameters, internal damage probability, and evolution trend, realizing an integrated output from damage identification to accurate measurement and internal assessment, providing a reliable basis for comprehensive damage assessment.

[0031] S6. If the correspondence between the polarization 3D thermal damage triad estimated by the mid-to-deep anomaly distribution and the internal damage is lower than the preset threshold, a dual optimization mechanism is activated. At the same time, the improved lightweight MobileViT model at the edge is linked to strengthen the stable correspondence between the surface observation signal and the deep damage under the constraint of millisecond-level inference speed. Finally, a high-precision and highly robust damage space mapping result is output, providing data support for subsequent visual twin mapping and intelligent treatment.

[0032] The dual optimization mechanism includes: introducing a dynamic visual posture self-calibration mechanism to correct the installation deviation of the machine vision acquisition unit and the visual distortion caused by environmental vibration in real time, so as to ensure the consistency of multimodal data acquisition. The parameters of the multi-scale visual attention fusion segmentation network are iteratively optimized, and the noise dynamic filtering logic and attention weight allocation are optimized by combining the feature commonality results of semi-supervised contrastive learning. Furthermore, step S6 specifically includes: If the intensity of the polarization three-dimensional features of the mid-to-deep layer abnormal distribution corresponding to the internal damage is lower than the preset threshold, then by comparing the corresponding intensity with the preset threshold, it is determined whether to activate the dual optimization mechanism, obtain the trigger signal, and then use the dual optimization mechanism to adjust the parameter configuration of the polarization three-dimensional features, enhance the correlation between the thermal damage triad and the internal damage, and obtain the optimized feature set. By using the optimized feature set, the improved lightweight MobileViT model deployed at the edge is linked to perform millisecond-level inference to generate a preliminary mapping result between surface observation signals and deep damage. If the stability of the preliminary mapping result is lower than a preset threshold, the weights of the MobileViT model are iteratively adjusted to enhance the correspondence between surface observation signals and deep damage, thereby obtaining a stable mapping result. Based on the stable mapping results, the spatial distribution characteristics of the three-dimensional polarization features are extracted to generate a high-precision damage spatial distribution. Combined with the multidimensional characterization of the thermal damage triad, the matching degree between the deep and intermediate anomalies and the internal damage is verified to obtain the final highly robust damage spatial mapping results. Then, three-dimensional visualization distribution data is output to determine the precise location and range of the damage space.

[0033] Among them, the strength of the correspondence between the polarization 3D thermal damage triad estimated by the mid-to-deep anomaly distribution and the internal damage is obtained by calculating the feature similarity and correlation confidence between the triad features and the internal damage sample library. The preset threshold is calibrated based on a large amount of engineering measured data. The iterative adjustment of the improved lightweight MobileViT model at the edge is based on stability indicators of the initial mapping results, such as inter-frame mapping error and feature association variance, and online fine-tuning of the model attention head and feature fusion layer weights.

[0034] Specifically, addressing the issues of weak correspondence between deep and intermediate anomalies and internal damage, and insufficient mapping stability, a dual optimization mechanism of dynamic visual pose self-calibration and iterative optimization of network parameters is employed. This mechanism, coupled with an improved lightweight MobileViT model at the edge, enables millisecond-level inference, effectively correcting acquisition biases, optimizing feature associations and model weights, and strengthening the stable correspondence between surface observation signals and deep damage. Ultimately, this results in high-precision and robust damage space mapping results, providing reliable data support for subsequent visual twin mapping and intelligent treatment.

[0035] Example 2 Please see Figure 4 This embodiment provides a machine vision-based damage detection system for implementing a machine vision-based damage detection method, including: The data acquisition module, through a customized machine vision array unit, adopts a three-modal synchronous acquisition architecture of polarization structured light and thermal imaging. It simultaneously acquires polarization texture feature images of the target surface, structured light encoded 3D contour point cloud data, and internal thermal conduction visual difference signals for the detection target (such as bridge components, mechanical parts, etc.), to obtain a multi-dimensional raw dataset. It can dynamically adapt to detection objects of different sizes without the need for additional adjustment of vision acquisition parameters, ensuring the efficiency and adaptability of data acquisition. The feature extraction and fusion module, based on a multi-dimensional original dataset, integrates optical flow algorithm and damage-guided visual mask reconstruction network to track pixel-level three-modal dynamic changes. At the same time, it filters invalid background information through self-supervised visual feature distillation technology, captures local gradient differences related to damage, and generates a surface-interior integrated heterogeneous feature map. Through multi-modal feature fusion, it ensures the comprehensiveness and accuracy of damage features. The noise filtering module calculates the gradient difference variance of the heterogeneous feature map. If the variance exceeds a preset threshold (i.e., noise interference affects the accuracy of damage recognition), dynamic noise filtering is performed through the damage-guided visual mask reconstruction network. Based on the feature priority of self-supervised distillation, artifact components such as sudden changes in illumination, vibration interference, and surface stains are selectively removed, while retaining the damage-related polarization texture, 3D contour, and core heat conduction signals. The module outputs a denoised multimodal fusion image with a high signal-to-noise ratio, ensuring the high quality of the feature map and providing a reliable data foundation for subsequent damage recognition. The damage localization and analysis module, based on the denoised multimodal fusion image, locates the suspected damage area through a spatial attention mechanism, and then performs quantitative characterization of the damage area through a cascaded damage analysis network. Through a multi-level network structure, it achieves damage feature extraction and quantitative analysis from coarse to fine, ensuring accurate localization and detailed characterization of the damage area. The damage assessment and report generation module, based on the quantitative results of the damaged area, calls the machine vision geometric measurement module and uses structured light-encoded 3D contour point cloud data to calculate the core quantitative parameters of the damage, including depth, area, extension angle, and volume. At the same time, it combines the visual difference signal of thermal conduction to simulate the internal thermal conduction path of the damaged area, establishes the mapping relationship between the polarization 3D thermal damage triad and internal damage (such as deep corrosion and hidden crack propagation), and obtains the deep anomaly distribution estimation results. It realizes the integrated output from damage identification to accurate measurement, and provides comprehensive data support for engineering decision-making through comprehensive assessment and report generation. The dual optimization module activates a dual optimization mechanism if the correlation strength between the estimated polarization 3D thermal damage triad and internal damage in the mid-to-deep anomaly distribution is lower than a preset threshold. Simultaneously, it links with the improved lightweight MobileViT model at the edge, strengthening the stable correspondence between surface observation signals and deep damage under millisecond-level inference speed constraints. This ultimately outputs high-precision, highly robust damage spatial mapping results, providing data support for subsequent visual twin mapping and intelligent treatment. This module ensures the system's high accuracy and robustness through dynamic optimization and model improvement.

[0036] This embodiment automates the entire process from data acquisition to damage identification, quantitative analysis, and comprehensive assessment, significantly improving the efficiency and accuracy of damage detection.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A damage detection method based on machine vision, characterized in that: Includes the following steps: S1. By using a customized machine vision array unit and a three-modal synchronous acquisition architecture of polarization structured light and thermal imaging, the target surface polarization texture feature image, structured light encoded 3D contour point cloud data, and internal thermal conduction visual difference signal are acquired simultaneously to obtain a multi-dimensional raw dataset. S2. Based on a multi-dimensional original dataset, the optical flow algorithm and the damage-oriented visual mask reconstruction network are integrated to track pixel-level three-modal dynamic changes. At the same time, the invalid background information is filtered out through self-supervised visual feature distillation technology to capture local gradient differences related to damage and generate a surface-interior integrated heterogeneous feature map. S3. Calculate the gradient difference variance of the heterogeneous feature map. If the variance exceeds the preset threshold, perform dynamic noise filtering through the damage-guided visual mask reconstruction network to output a denoised multimodal fusion image with a high signal-to-noise ratio. S4. Based on the denoised multimodal fusion image, the suspected damage area is located through the spatial attention mechanism, and then the damage area is quantitatively represented through the cascaded damage analysis network. S5. Based on the quantization results of the damaged area, the machine vision geometric measurement module is called. The structured light encoding 3D contour point cloud data is used to calculate the core quantization parameters of the damage. At the same time, the internal heat conduction path of the damaged area is simulated by combining the thermal conduction visual difference signal. The mapping relationship between the polarization 3D thermal damage triad and the internal damage is established to obtain the deep anomaly distribution estimation results. S6. If the correspondence between the polarization 3D thermal damage triad estimated by the mid-to-deep anomaly distribution and the internal damage is lower than the preset threshold, a dual optimization mechanism is activated. At the same time, the improved lightweight MobileViT model at the edge is linked to strengthen the stable correspondence between the surface observation signal and the deep damage under the constraint of millisecond-level inference speed, and output the damage space mapping result.

2. The damage detection method based on machine vision according to claim 1, characterized in that: Step S1 specifically includes: By using a customized machine vision array and a three-modal synchronous acquisition architecture of polarization structured light and thermal imaging, polarization texture images, structured light encoded three-dimensional contour point clouds and thermal conduction difference signals are acquired to generate multi-dimensional raw datasets. The polarization texture image is processed by a preset polarization filtering algorithm to extract the polarization features of the target surface and generate the first feature dataset. The structured light encoded data is then processed by a structured light decoding algorithm to generate a three-dimensional contour point cloud and obtain a spatial geometry dataset. Image signal analysis methods are used to extract heat conduction features and generate a second feature dataset. Through modal fusion processing, the first feature dataset, spatial geometry dataset, and second feature dataset are integrated into a multimodal dataset to generate a comprehensive feature dataset.

3. The damage detection method based on machine vision according to claim 1, characterized in that: Step S2 specifically includes: S21. Obtain a multi-dimensional raw dataset, calculate pixel-level motion vectors using an optical flow algorithm to obtain three-modal dynamic change information, and use a visual mask reconstruction network to process the three-modal dynamic change information to generate a preliminary surface feature map; the three-modal dynamic change information includes polarization texture differences, 3D contour undulations, and thermal conduction gradients; S22. By using self-supervised visual feature distillation technology, invalid background information is filtered out from the preliminary surface feature map to obtain a refined feature map. By constructing a multimodal feature fusion mechanism, damage-related local gradient differences are captured from the refined feature map to generate a gradient feature map containing damage-related information. S23. If there are significant local gradient differences in the gradient feature map, the damage area is determined by pixel-level analysis, a damage location map is generated, and then the damage location map and the refined feature map are fused to generate a heterogeneous feature map that integrates the surface and the interior.

4. The damage detection method based on machine vision according to claim 3, characterized in that: Generate gradient feature maps containing damage-related information, specifically including: Using refined feature maps as input, the pixel-level texture distortion features of the texture flow are captured by deformable convolution adaptively, the three-dimensional geometric fracture features of the contour flow are extracted by the depth-aware attention module, and the abnormal thermal diffusion features of the thermal conduction flow are captured by the thermodynamic graph network to obtain multimodal features. Multimodal features are fed into a cascaded interaction module for deep processing. First, a spatial transformer network is used to achieve sub-pixel-level feature alignment to ensure spatial consistency of different modal features. Then, a gated attention mechanism is used for multiple rounds of cross-validation to ensure that the features of each modality corroborate each other. Finally, a feature enhancement network is used to spatially enhance the validated damage areas to obtain validated high-quality features. Based on the verified high-quality features, a dynamic weighted fusion and reconstruction process is initiated. The modality confidence evaluation network is used to calculate the weight coefficients of each modality to determine the adaptive weight allocation. Then, a multi-scale feature pyramid is used to fuse cross-resolution features, retaining damage information at different scales. The transposed convolutional layer is used to reconstruct the spatial dimension of the feature map, restore the spatial details of the feature map, and generate a fused feature map that preserves details and semantic integrity. After obtaining the fused feature map, guided gradient filtering is applied to enhance the continuity of the damage boundary and improve the quality of edge features. A nonlinear activation function is used to highlight significant gradient differences and strengthen the response of key features. Then, a channel attention mechanism is used to optimize the feature representation and improve the feature discrimination, outputting a gradient feature map containing rich damage information.

5. The damage detection method based on machine vision according to claim 1, characterized in that: Step S3 specifically includes: Heterogeneous features are extracted from the input image. The gradient operator is used to calculate the gradient difference of each feature map to obtain the gradient difference matrix. Then, the statistical variance of the gradient difference matrix is ​​calculated, and it is determined whether the variance value exceeds the preset threshold. If it does, it is determined that noise filtering needs to be performed. Based on the variance value, a damage-oriented model is obtained, a visual mask is generated, and the feature map set is filtered using the mask to obtain a filtered feature set. The filtered feature set is then processed by a reconstruction network, and a convolutional neural network is used to perform dynamic noise filtering to output a denoised multimodal fusion image with a high signal-to-noise ratio.

6. The damage detection method based on machine vision according to claim 1, characterized in that: Step S4 specifically includes: S41. Construction of a spatial attention-guided damage localization network: A dual-path network architecture based on coordinate attention mechanism is constructed. The coordinate attention module of the spatial path captures the precise spatial coordinates of the suspected damage area, while the channel recalibration strategy of the channel path is used to enhance the signal-to-noise ratio of the damage features and output candidate regions. S42, Cascaded damage analysis architecture implementation: Based on the localized candidate region, a two-level cascaded damage analysis network is established. The first-level network achieves pixel-level accurate localization of the damage boundary within the candidate region by fusing deformable convolution with feature pyramid. The second-level network, based on localization, introduces a multi-scale receptive field module and combines it with 3D point cloud registration technology to extract the depth distribution and 3D morphological features of the damage region from coarse to fine. S43. Multi-dimensional quantitative analysis model execution: Based on the depth distribution and three-dimensional morphological features of the extracted damage area, implement physical feature-based quantitative analysis, calculate the surface curvature change through the normal vector field, and combine the thermal conduction abnormal area segmentation results to construct a complete descriptor containing geometric parameters and thermophysical parameters, transforming the morphological features extracted in the previous step into quantifiable engineering parameters. S44. Comprehensive Damage Assessment and Report Generation: Based on quantitative analysis results, a graph neural network is used to aggregate the correlation between geometric deformation and thermodynamic characteristics. By establishing matching rules between damage type and morphological characteristics and combining multi-dimensional quantitative parameters, a structured assessment report containing spatial coordinates, three-dimensional contour parameters, damage type classification, and severity level is output.

7. The damage detection method based on machine vision according to claim 1, characterized in that: Step S5 specifically includes: S51. Based on the damage area localization results, call the machine vision geometric measurement module to perform refined three-dimensional reconstruction of the identified damage area, and establish a digital geometric model of the damage area through the point cloud data encoded by structured light. S52. Based on the digital geometric model of the damaged area, implant the thermal property parameters of the material, construct a numerical model of heat conduction in the damaged area, and identify the abnormal heat conduction area caused by internal damage by comparing the difference between the theoretical heat conduction path and the measured thermal imaging data. S53. Based on geometric and thermal conduction features, a deep correlation model of polarization-3D-thermal damage triad is established. Through multimodal data registration and feature alignment, the statistical correlation between surface geometric deformation, polarization texture anomaly and thermal conduction anomaly is explored, forming a mapping rule for inferring internal damage from surface features. S54. Construct a comprehensive assessment system for damage status. Based on geometric parameters, thermal conduction anomalies, and multimodal correlation characteristics, use machine learning models to quantitatively assess internal damage and generate deep anomaly distribution estimation results that include damage geometric parameters, internal damage probability, and evolution trend.

8. The damage detection method based on machine vision according to claim 1, characterized in that: Step S6 specifically includes: If the intensity of the polarization three-dimensional features of the mid-to-deep layer abnormal distribution corresponding to the internal damage is lower than the preset threshold, then by comparing the corresponding intensity with the preset threshold, it is determined whether to activate the dual optimization mechanism, obtain the trigger signal, and then use the dual optimization mechanism to adjust the parameter configuration of the polarization three-dimensional features, enhance the correlation between the thermal damage triad and the internal damage, and obtain the optimized feature set. By using the optimized feature set, the improved lightweight MobileViT model deployed at the edge is linked to perform millisecond-level inference to generate a preliminary mapping result between surface observation signals and deep damage. If the stability of the preliminary mapping result is lower than a preset threshold, the weights of the MobileViT model are iteratively adjusted to enhance the correspondence between surface observation signals and deep damage, thereby obtaining a stable mapping result. Based on the stable mapping results, the spatial distribution characteristics of the three-dimensional polarization features are extracted to generate a high-precision damage spatial distribution. Combined with the multidimensional characterization of the thermal damage triad, the matching degree between the deep and intermediate anomalies and the internal damage is verified to obtain the final highly robust damage spatial mapping results. Then, three-dimensional visualization distribution data is output to determine the precise location and range of the damage space.

9. A machine vision-based damage detection system, applied to a machine vision-based damage detection method as described in any one of claims 1-8, characterized in that: include: The data acquisition module, through a customized machine vision array unit, adopts a three-modal synchronous acquisition architecture of polarization structured light and thermal imaging to simultaneously acquire polarization texture feature images of the target surface, structured light encoded 3D contour point cloud data, and internal thermal conduction visual difference signals to obtain a multi-dimensional raw dataset. The feature extraction and fusion module, based on a multi-dimensional original dataset, integrates optical flow algorithm and damage-guided visual mask reconstruction network to track pixel-level three-modal dynamic changes. At the same time, it filters invalid background information through self-supervised visual feature distillation technology, captures local gradient differences related to damage, and generates a surface-interior integrated heterogeneous feature map. The noise filtering module calculates the gradient difference variance of the heterogeneous feature map. If the variance exceeds the preset threshold, dynamic noise filtering is performed through the damage-guided visual mask reconstruction network. Based on the feature priority of self-supervised distillation, the damage-related polarization texture, 3D contour, and thermal conduction signal are preserved, and a denoised multimodal fusion image with high signal-to-noise ratio is output. The damage localization and analysis module, based on the denoised multimodal fusion image, locates the suspected damage area through a spatial attention mechanism, and then performs quantitative characterization of the damage area through a cascaded damage analysis network. The damage assessment and report generation module, based on the quantification results of the damage area, calls the machine vision geometric measurement module, uses structured light to encode 3D contour point cloud data, calculates the core quantification parameters of the damage, and combines the thermal conduction visual difference signal to simulate the internal thermal conduction path of the damage area, establishes the mapping relationship between the polarization 3D thermal damage triad and the internal damage, and obtains the deep anomaly distribution estimation results. The dual optimization module activates a dual optimization mechanism if the correspondence between the polarization 3D thermal damage triad estimated by the mid-to-deep anomaly distribution and the internal damage is lower than a preset threshold. Simultaneously, it links with the improved lightweight MobileViT model at the edge to enhance the stable correspondence between surface observation signals and deep damage under millisecond-level inference speed constraints, and outputs damage space mapping results.

Images