A method, apparatus and device for intelligent detection of cracks

Abstract

The present disclosure relates to a method, device and equipment for intelligent detection of cracks. The method comprises: obtaining an original image of a crack; determining multi-level features of the crack in a corrected image of the crack; dynamically weighting and fusing the multi-level features of the crack to obtain fused features of the crack; and performing risk early warning on the crack by using a Prophet time series decomposition algorithm and an empirical model of crack propagation rate according to parameters of the crack and the fused features. The technical solution of the present application fuses the multi-level features of the crack to obtain fused features, and performs risk early warning by using the Prophet time series decomposition algorithm and the empirical model of crack propagation rate based on the fused features, which can significantly improve the accuracy of risk early warning.

Description

Technical Field

[0001] This disclosure relates to the field of crack detection technology, and in particular to an intelligent crack detection method, apparatus and device. Background Technology

[0002] With the rapid advancement of urbanization, a large number of existing building structures are entering their aging period. Cracks in buildings, as the most direct manifestation of structural damage, require accurate identification and quantitative assessment as a core basis for judging the safety status of buildings. Current crack detection mainly relies on manual visual inspection combined with tools such as crack width gauges and feeler gauges for measurement, resulting in a relatively high error rate in the crack detection results. Summary of the Invention

[0003] To overcome the problems existing in related technologies, this disclosure provides an intelligent crack detection method, apparatus, and equipment to solve the problem of high error rate caused by manual crack detection in related technologies.

[0004] According to a first aspect of the present disclosure, a smart crack detection method includes:

[0005] Obtain the original image of the crack;

[0006] Based on the confidence level of the reference object corner detection, the consistency of the checkerboard corner spacing, and the gray-scale uniformity of the reference object area, the quality of the original image of the crack is determined; in response to the qualified quality of the original image of the crack, the original image of the crack is corrected to obtain the corrected image of the crack.

[0007] The multi-level features of the crack in the corrected image of the crack are determined. The multi-level features of the crack include local binary pattern texture features, edge gradient response features and local contrast enhancement features.

[0008] The multi-level features of the crack are dynamically weighted and fused to obtain the fused features of the crack;

[0009] Based on the parameters and fusion characteristics of the cracks, the Prophet time series decomposition algorithm and the crack propagation rate empirical model are used to provide risk warnings for the cracks.

[0010] The parameters include: maximum crack width, total length, number of nodes, and topological complexity index.

[0011] Secondly, this application proposes an intelligent crack detection device, comprising:

[0012] The acquisition module is used to acquire the original image of the crack;

[0013] The image processing module is used to determine the quality of the original image of the crack based on the confidence level of the corner detection of the reference object, the consistency of the spacing between the corner points of the checkerboard, and the gray-level uniformity of the reference object area; in response to the fact that the quality of the original image of the crack is qualified, the original image of the crack is corrected to obtain the corrected image of the crack.

[0014] The feature determination module is used to determine the multi-level features of the crack in the corrected image of the crack. The multi-level features of the crack include local binary pattern texture features, edge gradient response features, and local contrast enhancement features.

[0015] The feature fusion module is used to dynamically weight and fuse the multi-level features of the crack to obtain the fused features of the crack.

[0016] The hybrid prediction module uses the Prophet time series decomposition algorithm and the empirical model of crack propagation rate to provide risk warnings for cracks based on crack parameters and fusion characteristics.

[0017] Thirdly, an electronic device includes: at least one processor and at least one memory; the memory is used to store one or more program instructions; the processor is used to execute one or more program instructions to perform the intelligent crack detection method as described above.

[0018] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0019] Compared with the prior art, the technical solution of this application obtains fused features from the multi-level features of cracks. Based on the fused features, the Prophet time series decomposition algorithm and the crack propagation rate empirical model are used for risk warning, which can significantly improve the accuracy of risk warning.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0022] Figure 1 This is a flowchart illustrating an intelligent crack detection method according to an exemplary embodiment;

[0023] Figure 2 This is a structural block diagram illustrating an intelligent crack detection system according to an exemplary embodiment;

[0024] Figure 3 This is a structural block diagram of a feature-gated fusion network according to an exemplary embodiment;

[0025] Figure 4 This is a flowchart illustrating another intelligent crack detection method according to an exemplary embodiment;

[0026] Figure 5 This is a structural block diagram of a backbone network according to an exemplary embodiment;

[0027] Figure 6 This is a structural block diagram of an intelligent crack detection device according to an exemplary embodiment;

[0028] Figure 7 This is a structural block diagram of an electronic device according to an exemplary embodiment. Detailed Implementation

[0029] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0030] This application proposes an intelligent crack detection method, see appendix. Figure 1 The above method includes the following steps:

[0031] In step S102, the original image of the crack is obtained.

[0032] In this embodiment, a camera device can be used to photograph the cracks in the building to obtain the original image of the cracks. These cracks can be cracks in the building's walls, roof, or other similar structures.

[0033] In step S104, the quality of the original image of the crack is determined based on the confidence level of the corner detection of the reference object, the consistency of the corner spacing of the checkerboard, and the gray-scale uniformity of the reference object area; in response to the fact that the quality of the original image of the crack is qualified, the original image of the crack is corrected to obtain the corrected image of the crack.

[0034] In this embodiment, the quality of the original image of the crack is first determined based on the confidence level of the reference object corner detection, the consistency of the checkerboard corner spacing, and the grayscale uniformity of the reference object area. If the quality is deemed acceptable, the original image of the crack is corrected to obtain a corrected image of the crack. The purpose of correction is to eliminate distortion caused by images taken from different angles. Correction improves the crack detection accuracy.

[0035] Before correction, preprocessing can be performed, including but not limited to noise removal to make the image clearer.

[0036] In step S106, the multi-level features of the crack in the corrected image of the crack are determined. The multi-level features of the crack include local binary pattern texture features, edge gradient response features, and local contrast enhancement features.

[0037] In this embodiment, three different auxiliary feature branch networks can be used to process the corrected image of the crack to obtain local binary pattern texture features, edge gradient response features, and local contrast enhancement features.

[0038] The aforementioned auxiliary feature branch network can also be referred to as an auxiliary network.

[0039] See appendix Figure 2 A first auxiliary network is used to generate a local binary pattern texture feature map based on the input corrected image of the crack, the local binary pattern texture feature map carrying the aforementioned local binary pattern texture features.

[0040] The second auxiliary network is used to generate an edge gradient response feature map based on the input corrected image of the crack, the edge gradient response feature map carrying the aforementioned edge gradient response features;

[0041] A third auxiliary network is used to generate a local contrast enhancement feature map based on the input corrected image of the crack, the local contrast enhancement feature map carrying the aforementioned local contrast enhancement features.

[0042] A backbone network is used to generate a deep semantic feature map based on the input corrected image of the crack.

[0043] In step S108, the multi-level features of the crack are dynamically weighted and fused to obtain the fused features of the crack.

[0044] In this embodiment, a multi-level feature fusion network can be used to achieve the dynamic weighted fusion of the multi-level features of the crack as described above. This multi-level feature fusion network can also be called a feature-gated fusion unit.

[0045] The overall fusion process is as follows: the deep semantic feature map output by the backbone network and the three auxiliary feature branch maps are dynamically weighted and fused in the feature gating fusion unit. The fused feature map is then input into the detection head to complete the crack detection and segmentation.

[0046] The aforementioned detection head can also be called a detection head network or a detection head model.

[0047] The three auxiliary feature branch maps mentioned above are the edge gradient response map, the LBP texture feature map, and the local contrast enhancement map.

[0048] See appendix Figure 3 The feature gating fusion unit includes a feature alignment module, a gating weight generation module, and a dynamic weighted fusion module.

[0049] Regarding the feature alignment module, since the spatial resolution of the feature maps of the three auxiliary feature branches may differ from the resolution of the deep semantic feature map output by the backbone network, the three auxiliary feature branches are uniformly adjusted to have the same spatial size and number of channels as the deep semantic feature map output by the backbone network through bilinear interpolation upsampling or 3×3 convolution with a stride of 1. The four aligned feature maps are denoted as follows:

[0050] The adjusted backbone feature map F_backbone, the adjusted edge gradient feature map F_edge, the adjusted texture feature map F_lbp, and the adjusted local contrast enhancement feature map F_contrast.

[0051] The gated weight generation module is used to generate dynamic fusion weights for each feature channel or each spatial location. Specifically, it includes a splicing submodule, a global information extraction submodule, and a weight generation submodule.

[0052] The above-mentioned stitching submodule is used to stitch the four aligned feature maps, namely the adjusted backbone feature map F_backbone, the adjusted edge gradient feature map F_edge, the adjusted texture feature map F_lbp, and the adjusted local contrast enhancement feature map F_contrast, together in the channel dimension to obtain the stitched feature map F_concat, which has a shape of H×W×4C; where C is the number of channels in a single branch.

[0053] The global information extraction module described above can be implemented using a global average pooling layer to perform global average pooling on the concatenated feature map F_concat, compressing the spatial information into a 1×1×4C global description vector.

[0054] The aforementioned weight generation submodule includes a first fully connected layer, a ReLU layer, a second fully connected layer, and a Sigmoid layer. The first fully connected layer performs linear weighted fusion and dimensionality transformation on the input global description vector, resulting in four one-dimensional vectors representing the weights of the feature maps. These four weight vectors are w_backbone, w_edge, w_lbp, and w_contrast, each with a dimension of 1×1×C. The ReLU layer sets the less-than-zero elements in the four one-dimensional vectors representing the feature map weights to zero, achieving non-linear processing. The second fully connected layer transforms the four one-dimensional vectors representing the feature map weights into four weight values. The Sigmoid layer transforms the four weight values ​​into probability values ​​between 0 and 1. Each weight value between 0 and 1 represents the importance of the corresponding feature map at the current spatial location or channel. Higher weights indicate a greater contribution of the feature map to the fusion process.

[0055] In some embodiments, the dynamic update mechanism for weights is as follows: all parameters of the weight generation submodule are learnable parameters; these parameters include the weights and biases of the fully connected layers. During model training, these parameters are automatically optimized through end-to-end backpropagation. After training, the gated weight generation module automatically adjusts the contribution weights of each branch according to the crack features of different input images: for low-contrast cracks, the weights of the edge gradient branch and the contrast branch are automatically increased; for complex texture backgrounds, the weights of the LBP branch are automatically increased. Weighted fusion does not occur only once in the gated weight generation module, but is performed at multiple feature levels: the outputs at different stages of the backbone network (such as C2, C3, C4, C5) are fused with the corresponding auxiliary features. The fused feature maps of each level are then input into the Feature Pyramid Network (FPN) for multi-scale feature aggregation. This multi-level, multi-scale fusion strategy ensures that the model can capture both the local details of subtle cracks, i.e., shallow features, and understand the overall structural semantics of cracks, i.e., deep features.

[0056] In step S110, based on the parameters of the crack and the fusion characteristics, the Prophet time series decomposition algorithm and the crack propagation rate empirical model are used to provide a risk warning for the crack.

[0057] The parameters include, but are not limited to, the maximum width, total length, number of nodes, and topological complexity index of the crack. For any one of the parameters, a hybrid prediction strategy combining the Prophet time series decomposition algorithm and an empirical model of crack propagation rate is used to provide risk warnings.

[0058] In this embodiment, the Prophet time series decomposition algorithm described above uses the Prophet model, which decomposes a complex time series into several simpler components, and then predicts and combines them separately. For crack detection, the Prophet time series decomposition algorithm processes key parameter sequences recorded at fixed time intervals (e.g., weekly, monthly), such as the maximum crack width.

[0059] The decomposition process mainly includes the following parts, which are explained below: Trend Term: Models the long-term direction of change in crack parameters, whether it is continuous growth, stabilization, or other patterns. The Prophet model uses a linear or logistic growth function to fit this non-periodic change. Seasonality Term: Captures periodic fluctuations caused by environmental factors (such as temperature and humidity). For example, wall materials may cause crack widths to be slightly wider in summer and slightly narrower in winter due to thermal expansion and contraction. The Prophet model uses Fourier series to fit this periodic pattern. Holidays / Events Term: Handles the impact of occasional, irregular events, such as extreme weather or vibrations from nearby construction. Error Term: Represents random fluctuations that the Prophet model cannot explain. The workflow is as follows: Historical data is input into the Prophet model, which automatically learns and decomposes the above terms, then makes predictions for each term separately, and finally sums the prediction results of each part to obtain a future prediction value based on historical statistical regularities.

[0060] In this embodiment, the empirical model for crack propagation rate described above is the Paris formula. The Paris formula is a classic fracture mechanics model used to describe the rate of crack propagation in materials under cyclic stress. This prediction result is based on physical and mechanical principles and does not rely on the statistical regularities of historical data.

[0061] This application refines the granularity of the early warning to each parameter, enabling risk warnings for each parameter, thus significantly improving the accuracy of early warnings for cracks.

[0062] See appendix Figure 4The historical detection data and environmental temperature data of the aforementioned cracks are stored in the crack time-series database. Using the Prophet time-series decomposition model and the Paris empirical crack propagation rate model, combined with the current crack parameters, a hybrid prediction strategy is employed to obtain prediction results, which represent the predicted development trend. Based on the prediction results, the structural safety risk level is determined; this level is divided into four categories: A, B, C, and D. A represents safe, B represents basically safe, C represents restricted use, and D represents dangerous. Furthermore, based on the structural safety risk level, differentiated inspection recommendations and early warning information are provided.

[0063] The technical solution of this application dynamically weights and fuses the multi-level features of the crack to obtain the fused features of the crack. Based on the fused features of the crack, a hybrid prediction strategy combining the Prophet time series decomposition algorithm and the crack propagation rate empirical model is used for risk warning, which can obtain a more accurate risk warning.

[0064] In some embodiments, the structure of the backbone network described above is as follows: the backbone network adopts the ResNet-50-FPN architecture.

[0065] The ResNet-50 network serves as the backbone for feature extraction, while FPN is used for multi-scale feature fusion.

[0066] The ResNet-50 network consists of five stages (Stages 0-4), each containing several residual blocks (Bottleneck Blocks).

[0067] Table 1. ResNet-50 Network Parameter Table

[0068]

[0069] Each residual block, Bottleneck, consists of a structure consisting of sequentially connected 1×1 convolutions (dimensionality reduction), 3×3 convolutions, and 1×1 convolutions (dimensionality increase), and includes a skip connection.

[0070] See appendix Figure 5 Feature Pyramid Networks (FPN) build upon ResNet-50 to construct multi-scale feature fusion paths with top-down and lateral connections.

[0071] Bottom-up path: The outputs of each stage of ResNet-50 (C2, C3, C4, C5) serve as feature maps at different scales. As shown in the figure, from bottom to top, the sequence is: input layer, first stage layer, with the output of the first stage layer being C2; the output of the second stage layer being C3; the output of the third stage layer being C4; and the output of the fourth stage layer being C5. Top-down path: Starting from the highest-level feature map, it is fused with lower-level feature maps step by step through upsampling. Lateral connection: The top-down features of each layer are fused with the corresponding bottom-up features through 1×1 convolution. For example, feature map C5 is first convolved with 1×1; then upsampled by 2 times to enlarge its size, resulting in feature map C51, which has the same size as feature map C4; feature map C4 is first convolved with 1×1; then pixel values ​​are added to feature map C51 pixel by pixel, and then convolved again to obtain feature map P4. It can be seen that feature map P4 is obtained by fusing feature maps C5 and C4. Similarly, we can obtain P3, P2, and P1. Output feature maps at four scales: P2, P3, P4, and P5, each corresponding to a different receptive field and resolution.

[0072] The improvements to the backbone network are explained below. Compared to the standard ResNet-50-FPN, the backbone network of this invention has the following adaptive improvements:

[0073] Dilated convolution is introduced into the residual block to expand the receptive field without increasing the number of parameters, thereby improving the ability to detect large-scale cracks.

[0074] An attention module is added to the lateral connections of the FPN to enhance the feature response of the crack region and suppress background interference.

[0075] In some embodiments, the auxiliary branches operate in parallel, extracting different visual features from the original input image. The output of each branch is a feature map with the same spatial size as the feature map of a certain stage of the backbone network, which serves as the input to the gated fusion unit.

[0076] In some embodiments, the edge gradient response branch is described below. The function of the edge gradient response branch is to extract gradient intensity information of the crack edges in the image. Cracks appear as regions of drastic pixel value changes in the image, and the edge gradient response feature map can highlight these changes, especially enhancing low-contrast cracks.

[0077] Feature extraction includes the following steps:

[0078] Step 1: Convert the RGB image of the input crack correction image above into a grayscale image.

[0079] Step 2: Use the Sobel operator to calculate the horizontal gradient (Gx) and vertical gradient (Gy) respectively; the Sobel operator is a 3×3 convolution kernel, which has a smoothing and suppression effect on noise.

[0080] Step 3: Calculate the gradient magnitude G = sqrt(Gx² + Gy²).

[0081] Step 4: Perform nonlinear enhancement on the gradient magnitude (such as exponential mapping or adaptive threshold stretching) to further amplify the weak edge response;

[0082] Step 5: Map the single-channel gradient map to a C-channel feature map through a learnable 1×1 convolutional layer, aligning it with the number of backbone feature channels.

[0083] The characteristics of the edge gradient response feature map are as follows: the crack edge has a high response value, and the response value in flat areas is close to zero; it is relatively insensitive to changes in illumination, the gradient is a local pixel difference and is not affected by global brightness; for fine cracks with a width of less than 0.1 mm, the gradient response can still be detected.

[0084] In some embodiments, a Local Binary Pattern (LBP) texture branch is used to extract local texture structure information of the corrected image of the crack. The texture patterns of the crack region differ significantly from those of the intact background region, and LBP can quantify this difference.

[0085] (2) Feature extraction process:

[0086] Step 1: Convert the RGB image of the input crack correction image above into a grayscale image;

[0087] Step 2: In the grayscale image above, for each pixel, take 8 neighboring pixels within its 3×3 neighborhood;

[0088] Step 3: Use the pixel value of the center pixel as a threshold, and compare the pixel values ​​of the 8 neighboring pixels with this threshold respectively;

[0089] For any one of the eight neighboring pixels mentioned above, if the pixel value of the neighboring pixel is greater than or equal to the threshold mentioned above, then the encoding of the neighboring pixel is set to 1; if the pixel value of the neighboring pixel is less than the threshold mentioned above, then the encoding of the neighboring pixel is set to 0. This results in an 8-bit binary code.

[0090] Step 4: Convert the above 8-bit binary code into decimal values ​​from 0 to 255 in sequence, and use them as the LBP value of the center pixel.

[0091] Step 5: Perform histogram equalization or normalization on the LBP values ​​of the center pixels to enhance texture contrast.

[0092] Step 6: Map the single-channel LBP map to a C-channel feature map using a learnable 1×1 convolutional layer.

[0093] The output feature map exhibits higher LBP values ​​in areas of abrupt texture change (crack edges) and lower LBP values ​​in uniform areas. LBP features are inherently robust to changes in illumination, based on local relative comparisons rather than absolute grayscale values. They can effectively distinguish crack edges from pseudo-crack interference such as ordinary wall textures and stains.

[0094] In some embodiments, a local contrast enhancement branch is used to enhance the visual contrast of low-contrast areas in an image. Cracks typically appear as dark lines embedded in a light background, and local contrast enhancement can amplify this contrast.

[0095] Feature extraction process: Step 1: Convert the input RGB image to grayscale;

[0096] Step 2: In the grayscale image above, for each pixel, calculate the mean and standard deviation of the grayscale value within its local neighboring pixels (such as an 11×11 window).

[0097] Step 3: Calculate the local contrast value: C=(I-μ) / (σ+ε), where I is the gray value of the current pixel, μ is the mean of the neighborhood, σ is the standard deviation of the neighborhood, and ε is a small constant used to prevent division by zero.

[0098] Step 4: Apply a non-linear mapping to the above contrast values, such as using the Sigmoid function or the ReLU function, to enhance the positive contrast area; where the crack is dark and C is a negative value, the absolute value is taken for enhancement.

[0099] Step 5: Map the single-channel contrast map to a C-channel feature map using a learnable 1×1 convolutional layer to obtain the aforementioned local contrast enhancement feature map.

[0100] The characteristics of the local contrast enhancement feature map are as follows: the contrast value of the crack region (dark lines) is significantly higher than that of the background region. This effectively overcomes the problem of crack visibility under adverse conditions such as low light and shadow occlusion. Unlike global contrast enhancement, local contrast enhancement preserves the ability to perceive fine structures.

[0101] In some embodiments, the standard Mask R-CNN detection head comprises three parallel branches: a Classification Head, which predicts the probability that each candidate region belongs to the crack category; a Bounding Box Regression Head, which refines the position and size of the candidate boxes; and a Mask Segmentation Head, which generates a pixel-level segmentation mask for each candidate region.

[0102] This application makes the following improvements to the detection head: Dilated convolution is introduced into the mask segmentation branch. In existing technologies, the standard Mask R-CNN segmentation head uses a continuous fully convolutional network (FCN) to generate the mask, resulting in a limited receptive field and insufficient ability to segment long, narrow cracks. The technical solution of this application replaces some standard 3×3 convolutions with dilated convolutions in the convolutional layers of the mask segmentation branch, setting different dilation rates (e.g., dilation rate = 2, 4, 6) to construct a multi-scale dilated convolution pyramid. The convolutional outputs with different dilation rates are concatenated along the channel dimension and then fused using 1×1 convolutions. The resulting technical effect is a significant expansion of the receptive field of the segmentation head without increasing the number of parameters, improving the ability to completely segment long cracks and mesh-like cracks.

[0103] This paper introduces deformable convolution. In existing technologies, cracks exhibit diverse morphologies (straight lines, curves, bifurcations, mesh-like structures), and the fixed sampling grid of standard convolution is ill-suited to adapting to the arbitrary orientation of cracks. The technical solution of this application introduces deformable convolution within the 3×3 convolution of the mask segmentation branch. The deformable convolution learns a spatial offset for each sampling point, enabling the sampling grid to adaptively conform to the actual orientation of the crack. The offset is learned from the input feature map through an additional convolutional layer. The achieved technical effect is that the segmentation mask can more accurately conform to the true boundary of the crack, especially significantly improving the segmentation accuracy for curved and bifurcated cracks.

[0104] Gradient-based Loss Balancing. In existing technologies, crack pixels account for a very small proportion of the entire image (extreme imbalance between positive and negative samples), and standard cross-entropy loss can cause the model to be biased towards background prediction. The technical solution in this application introduces an edge gradient weighting term into the mask loss function to calculate the difference between the gradient map of the segmentation mask and the gradient map of the real mask, as an additional loss term.

[0105] The gradient loss term forces the model to focus on the edge region of the crack, rather than just the interior. The resulting improvement is enhanced segmentation accuracy of the crack boundary, and a reduction in jagged edges and fractures.

[0106] Feature sharing optimization for classification and segmentation. In existing technologies, the standard Mask R-CNN uses the features after ROI pooling independently for the classification and segmentation branches, resulting in insufficient information sharing. The technical solution in this application adds a feature interaction module before feeding the feature maps into the classification and segmentation branches respectively after ROI pooling. This feature interaction module uses a cross-attention mechanism to allow classification and segmentation features to provide complementary information to each other. The resulting technical effects are as follows: improved classification accuracy indirectly promotes improved segmentation quality, and the overall mAP is improved.

[0107] In some embodiments, step S110, based on the parameters of the crack and the fusion characteristics, uses the Prophet time-series decomposition algorithm and an empirical model of crack propagation rate to provide a risk warning for the crack, which may further include the following steps:

[0108] In step S204, for any sampling time point, the first dynamic weight parameter of the Prophet time series decomposition algorithm and the second dynamic weight parameter of the crack propagation rate empirical model are determined respectively. The sum of the first dynamic weight parameter and the second dynamic weight parameter is 1.

[0109] In step S206, the Prophet time-series decomposition algorithm is used to determine a first prediction result based on the historical data, current data, prediction days, and environmental data of the parameters of the crack. The environmental data includes temperature and humidity.

[0110] In step S208, the product of the first dynamic weight parameter and the first prediction result is calculated to obtain the first product.

[0111] In step S210, the second prediction result is calculated using an empirical model of crack propagation rate.

[0112] In step S212, the product of the second dynamic weight parameter and the second prediction result is calculated to obtain the second product.

[0113] In step S214, the first product and the second product are summed to obtain the mixed prediction result of the parameters at the sampling time point.

[0114] In this embodiment, the following calculation formula is used:

[0115] Where, hybrid_values ​​is the hybrid prediction result of the parameters; weights is the first dynamic weight parameter; prophet_values ​​is the first prediction result; This is the second prediction result.

[0116] In some embodiments, the first prediction result and the second prediction result described above can be aligned to obtain an aligned first prediction result and second prediction result.

[0117] The alignment process refers to aligning the first and second prediction results according to their sampling time points. In other words, at each sampling time point, there is a corresponding set of first and second prediction results.

[0118] In some embodiments, the method may further include the following steps:

[0119] For any sampling time point and any parameter, the first dynamic weight parameter of the Prophet time series decomposition algorithm is determined, including:

[0120] Determine the basic weight parameters and normalized uncertain weight parameters of the parameters of the crack at the sampling time point;

[0121] The first dynamic weight parameter of the Prophet time series decomposition algorithm is determined based on the basic weight parameter and the normalized uncertain weight parameter.

[0122] The first dynamic weight parameters of the Prophet temporal decomposition algorithm are calculated using the following formula. :

[0123]

[0124] threshold)));

[0125] in, Basic weight parameters; The minimum value of the weight parameters; np.exp represents the exponential function; This represents the normalized uncertain weight parameter.

[0126] threshold represents the threshold value of the uncertainty weighting parameter, which can be flexibly set.

[0127] In some embodiments, the default configurations of the aforementioned weight parameters are as follows: the aforementioned basic weight parameters The threshold for the uncertainty weighting parameter is 0.5; the threshold value is 0.3; the minimum value of the weighting parameter is... The value is 0.2; the maximum value of the weight parameter is 0.8.

[0128] In some embodiments, determining the normalized uncertainty weight parameter of the parameter of the crack at the sampling time point specifically includes:

[0129] The parameter value at the sampling time point, the historical mean and historical standard deviation of the parameter at historical sampling time points are determined respectively.

[0130] Based on the parameter value of the parameter at the sampling time point, the historical mean and historical root mean square deviation of the parameter at historical sampling time points, the normalized uncertain weight parameter of the parameter at the sampling time point of the crack is determined.

[0131] In some embodiments, the above-described normalized uncertain weight parameters are determined. The method includes the following steps:

[0132] Obtain historical statistical information of the above parameters within a predetermined prediction interval;

[0133] The aforementioned historical statistical information includes: the width of the prediction interval within the predetermined prediction interval, and the sampling sequence of the parameter at historical time sampling points within the predetermined prediction interval;

[0134] The prediction interval width is the difference between the upper and lower bounds of the prediction interval for the parameter output by the Prophet model at the current sampling time point to be predicted. Its unit is exactly the same as that of the parameter being predicted.

[0135] The aforementioned statistical information includes: the historical prediction interval width of the parameter; the historical average and standard deviation of the parameter. The historical average can be obtained by calculating all historical data for the parameter and then dividing by the number of occurrences. The standard deviation can be calculated using the standard deviation formula. For example, if the parameter is the maximum width of a crack, historical data for that maximum width can be collected. If the historical data includes, for example, 100 maximum width data points, then the average and standard deviation of these 100 maximum widths can be calculated.

[0136] When the mean squared error is greater than zero,

[0137] normalized_uncertainty=(width-width_mean) / width_std.

[0138] Where width_std is the mean square error of the parameter in the predetermined prediction interval; width_mean is the historical average value of the parameter in the predetermined prediction interval; and width is the parameter value at the current sampling time point.

[0139] The technical solution of this application fully considers the influence of historical statistical data in the process of determining the normalized uncertain weight parameters, thus significantly improving the calculation accuracy of the normalized uncertain weight parameters. This is beneficial for obtaining more accurate normalized uncertain weight parameters.

[0140] In some embodiments, the above method may further include the following steps:

[0141] Determine the prediction interval width of the Prophet time series decomposition algorithm;

[0142] Based on the prediction interval width of the Prophet time series decomposition algorithm, the first dynamic weight parameter of the Prophet time series decomposition algorithm and the second dynamic weight parameter of the crack propagation rate empirical model are adjusted respectively, specifically including:

[0143] The width of the Prophet prediction interval decreases, the first dynamic weight parameter of the Prophet time series decomposition algorithm increases, and the second dynamic weight parameter of the crack propagation rate empirical model decreases.

[0144] The width of the Prophet prediction interval increases, the first dynamic weight parameter of the Prophet time series decomposition algorithm decreases, and the second dynamic weight parameter of the crack propagation rate empirical model increases.

[0145] In this embodiment, the predicted interval width is the interval width in the selected historical data. The larger the predicted interval width, the greater the uncertainty, and the smaller the first dynamic weight parameter.

[0146] The smaller the prediction interval width, the less uncertainty there is, and the larger the first dynamic weight parameter should be.

[0147] In some embodiments, determining the type of the crack based on the fusion features of the crack may further include the following steps: determining the topology map of the crack according to the fusion features of the crack, specifically, 1. Skeleton extraction: the fusion feature map of the crack is binarized to obtain a binary segmentation mask map of the crack, and the binary segmentation mask map of the crack is thinned to reduce the crack from a multi-pixel width to a single-pixel width skeleton line.

[0148] 2. Node detection: Scan the 3×3 neighborhood of each pixel in the skeleton image, and determine the node type of the pixel based on the number of skeleton pixels in the neighborhood.

[0149] If the number of pixels in the neighboring skeleton is 1, the node type of this pixel is determined to be an endpoint (Terminal Node).

[0150] The number of pixels in the neighborhood skeleton is 2, which determines that the node type of this pixel is a normal path point (non-node).

[0151] The number of pixels in the neighborhood skeleton is 3, which determines that the node type of this pixel is a T-shaped branching node.

[0152] If the number of neighboring skeleton pixels is greater than or equal to 4, the node type of that pixel is determined to be an X-shaped intersection node.

[0153] 3. Edge construction: Tracing along the skeleton line, recording the coordinates of all path points between two nodes;

[0154] 4. Graph Construction: Construct networkx.Graph objects with nodes and edges to store the topology.

[0155] The crack type determination parameters are determined based on the crack topology map.

[0156] The type of crack is determined based on the crack type judgment parameters.

[0157] The types of cracks include: linear cracks, radial cracks, and network cracks. Radial cracks are characterized by a high-central convergence node connecting multiple outward-radiating crack branches. Radial cracks have a tree-like structure radiating outwards from the center and do not contain closed loops, meaning the number of loops is zero. Network cracks, on the other hand, have a number of loops greater than one.

[0158] The parameters for determining the crack type include, but are not limited to: node degree distribution, connected components, number of cycles, and topological complexity index.

[0159] In this embodiment, the crack type determination parameters described above satisfy the following relationship:

[0160] Cycles = E - V + C; where E is the number of edges, V is the number of nodes, C is the number of connected components, and Cycles is the number of cycles.

[0161] One rule related to the number of rings mentioned above is: if the number of rings is greater than zero, then the type of crack is determined to be a network crack. Regarding the connected component C mentioned above, graph theory connected component algorithms (such as BFS / DFS) can be used to calculate the number of connected components in the topology graph; the number of nodes contained in each connected component is counted; the relevant rule for determining the connected components mentioned above is: if the number of connected components is 1, then the entire crack network is completely connected; and the number of nodes contained in the largest connected component is greater than or equal to 90% of the total number of nodes. If there are multiple independent connected components, then it does not meet the characteristics of a radial crack and may be multiple independent cracks.

[0162] In some embodiments, determining the type of crack based on the crack type determination parameter may further include the following steps: in response to the ring number being greater than zero and the topology complexity index being greater than a predetermined first topology complexity index threshold, the type of crack is a mesh crack.

[0163] In this embodiment, the threshold for the first topological complexity index can be flexibly set, for example, to 0.4.

[0164] In response to the node degree distribution being greater than a predetermined node degree distribution threshold, the number of rings is equal to zero, the topology complexity index is less than or equal to the first topology complexity index threshold, and greater than or equal to the second topology complexity index threshold, and the type of crack is radial crack.

[0165] In this embodiment, the threshold for the second topological complexity index can be flexibly set, for example, to 0.15.

[0166] Table 2. Crack Type Determination Table Based on Crack Type Judgment Parameters

[0167]

[0168] In some embodiments, if the above crack type determination parameters meet the following conditions:

[0169] A radial crack is identified if the following conditions are met: maximum degree k_max is greater than or equal to 4; degree_variance is greater than a predetermined degree_variance threshold; number of cycles is zero; number of connected components C is 1; and TCI is greater than or equal to 0.15 and less than or equal to 0.4. If the following conditions are met: number of cycles is greater than zero; TCI is greater than 0.40; it is identified as a network crack. If none of the above conditions are met, it is identified as a linear crack.

[0170] In some embodiments, determining the topology complexity index includes calculating the topology complexity index using the following formula:

[0171] TCI=α×(k_max / V)+β×(Cycles / E)+γ×(1 / C);

[0172] Where TCI is the topological complexity index; α is the first weight coefficient; β is the second weight coefficient; γ is the third weight coefficient; the above weight coefficients can be calibrated based on experimental data. k_max is the maximum node degree; the maximum degree among all nodes; V is the total number of nodes; Cycles is the number of cycles; C is the number of connected components; E is the total number of edges.

[0173] For example, one formula for calculating the Topological Complexity Index (TCI) is as follows;

[0174] TCI=0.5×(k_max / len(V))+0.3×(Cycles / max(len(E),1))+0.2×(1 / C);

[0175] Where len(E) represents the number of edges in edge set E;

[0176] len(V) represents the number of vertices in the vertex set V.

[0177] In some embodiments, step S104, correcting the original image of the crack to obtain a corrected image of the crack, may further include the following steps:

[0178] Determine the coordinates of each pixel in the original image, and denot them as the first coordinate set;

[0179] Determine the homography transformation matrix, which represents the point-to-point mapping relationship between the original image plane and the target correction plane;

[0180] Based on the homography transformation matrix and the first coordinate set, determine the second coordinate set in the target correction plane;

[0181] The second coordinate set is adjusted to obtain the third coordinate set, which specifically includes:

[0182] Based on the homography transformation matrix, determine the projection points of the four vertices of the original image onto the target correction plane;

[0183] In the target correction plane, determine the coordinates of each projection point of the four vertices;

[0184] Determine the minimum x-coordinate and minimum y-coordinate of the coordinates of each projection point of the four vertices;

[0185] Based on the minimum values ​​of the horizontal and vertical coordinates, the coordinates of each pixel in the second coordinate set are translated to obtain the third coordinate set.

[0186] In this embodiment, the minimum x-coordinate and the minimum y-coordinate of each pixel can be subtracted to obtain the translated coordinates. This yields the third coordinate set.

[0187] In this embodiment, this is a crucial but easily overlooked aspect in engineering implementation. If the original corner point is directly mapped to an ideal rectangle starting at (0,0), due to the non-linear characteristics of perspective transformation, the corrected wall edge is easily clipped outside the canvas, resulting in the loss of information about the corner or crack end.

[0188] To solve this problem, the system adopts a four-corner projection prediction strategy:

[0189] 1. Calculate the preliminary projection: Using the preliminary homography matrix generated in the previous step, calculate the projection positions of the four vertices of the original image (the entire frame) in the target space.

[0190] 2. Determine the circumscribed boundary: Calculate the range of x and y coordinates of the four projection points and calculate the minimum circumscribed rectangle (min_x, min_y, max_x, max_y) that contains all projection points.

[0191] 3. Calculate translation: The system automatically calculates the translation offset (-min_x, -min_y) in the horizontal and vertical directions to ensure that all projected coordinates are translated into the positive range.

[0192] 4. Update the target point set: Overlay the offset onto the entire ideal reference target rectangle to form the final target coordinates. In this way, the minimum coordinates after the four corners of the original image are projected fall exactly near the upper left corner (0,0) of the output canvas, thus preserving the entire effective area of ​​the wall without loss and avoiding the reduction of the field of view caused by correction.

[0193] Based on the grayscale value of each pixel in the original image and the third coordinate set, a bilinear interpolation algorithm is used to determine the grayscale value of each pixel in the third coordinate set.

[0194] In this embodiment, during the pixel value mapping process, the bilinear interpolation algorithm is used to calculate the grayscale value of each pixel in the target correction plane. This interpolation method effectively suppresses the jagged edges commonly found in the target correction plane by calculating the weighted average of the nearest neighbor 2×2 regions, while preserving the sharp transition features of the crack edges. This is highly beneficial for subsequent gradient-based crack fine segmentation. For areas not covered after transformation (such as areas outside the distorted corners), they are uniformly filled with a pure black background.

[0195] The corrected image is generated based on the grayscale value of each pixel and the third coordinate set.

[0196] In some embodiments, step S104, determining the quality of the original image of the crack based on the confidence level of the reference object corner detection, the consistency of the checkerboard corner spacing, and the gray-level uniformity of the reference object area, may further include the following steps:

[0197] Corner Detection Confidence. Based on AprilTag detection (recommended for encoding references), evaluate the confidence of AprilTag detection. The AprilTag detector outputs a quality score during decoding, reflecting the reliability of tag detection. A higher quality score indicates a clearer tag image, less distortion, and more reliable decoding. The method includes: inputting the image and determining the confidence score; converting to grayscale, as AprilTag detection is typically performed on grayscale images; creating the AprilTag detector; setting detection parameters to improve accuracy; detection parameters include: using AprilTag subpixel thinning, corner thinning window size, maximum number of iterations, and minimum precision threshold. Performing detection; calculating the confidence score for each detected tag.

[0198] Assess the consistency of corner spacing on a chessboard. Ideally, the spacing between all adjacent corner points should be equal. Camera angle distortion can cause variations in corner spacing at different locations; the larger the coefficient of variation, the more severe the distortion. A smaller coefficient of variation is better. Determine the Euclidean distance between all adjacent corner points; the horizontal and vertical spacing between adjacent corner points; calculate the statistical measures, including the standard deviation and mean. The coefficient of variation equals the standard deviation divided by the mean; a smaller coefficient of variation indicates more consistent spacing.

[0199] Gray-scale uniformity of the reference area. Gray-scale uniformity reflects whether the lighting conditions are ideal. Uneven lighting will affect the reference area. Excessive brightness differences between different areas will affect corner detection and subsequent measurement accuracy. The ratio of gray-scale standard deviation to mean (coefficient of variation) is used as a measure of uniformity.

[0200] In some embodiments, the loss function of the above model is a total loss function comprising: a classification loss function, a bounding box regression loss function, and a combined mask loss function. The combined mask loss function is determined based on the gradient loss; specifically, the gradients of the predicted mask and the ground truth mask are calculated, and the gradient loss is determined based on these gradients.

[0201] In some embodiments, during training, the optimizer is trained using SGD with a momentum of 0.9 and a weight decay of 0.0005.

[0202] This invention utilizes a multi-level feature fusion network to dynamically gate and fuse edge gradient response maps, Local Binary Pattern (LBP) texture feature maps, and local contrast enhancement maps from traditional image processing with deep semantic features extracted by the backbone network. This effectively enhances the model's ability to detect minute cracks under low contrast and low illumination conditions. Experiments show that on the test dataset, the method of this invention improves the recall rate of minute cracks with a width of less than 0.1 mm by 12.3 percentage points compared to the baseline Mask R-CNN model, achieving an overall mAP@0.5 of 92.4% and a segmentation IoU of 88.7%.

[0203] This invention transforms crack segmentation results into a graph structure representation. Through skeleton extraction, node detection (endpoint and intersection identification), and edge connection reconstruction, a crack topology network is constructed. Based on the topology network characteristics, the system can automatically identify linear cracks, mesh cracks, radial cracks, and other patterns, achieving an accuracy rate of 91.5%. The topology analysis results can directly support structural engineers in making preliminary judgments about the causes of cracks (e.g., mesh cracks often indicate material aging, while radial cracks often indicate localized stress concentration).

[0204] Based on historical time-series data and environmental temperature and humidity data, it can predict the development trend of cracks over the next 30, 90, and 180 days. The mean absolute percentage error (MAPE) of the prediction results is less than 15%. Based on the prediction results and the "Standard for Reliability Appraisal of Civil Buildings" (GB 50292-2015), it automatically determines the structural safety risk level (Level A: Safe / Level B: Basic Safe / Level C: Restricted Use / Level D: Dangerous), and generates differentiated inspection suggestions and early warning information, realizing a shift from post-detection to pre-detection.

[0205] The technical solution of this application addresses the problem of insufficient detection accuracy for minute cracks under low contrast and low illumination conditions. Existing deep learning models rely solely on high-level semantic features of RGB images, failing to adequately utilize low-level visual features such as local gradient information and texture differences at crack edges. This results in missed detections of minute cracks less than 0.1 mm in width or cracks with low contrast. The technical solution of this application constructs a multi-level feature fusion network, dynamically weighting and fusing traditional image processing features (edge ​​gradient response, local binary pattern texture features, and local contrast) with deep learning features, thereby enhancing the model's ability to perceive minute cracks.

[0206] This invention addresses the issues of missing crack topology information and reliance on manual interpretation for crack pattern recognition. Existing methods, after outputting pixel-level segmentation masks, only calculate geometric parameters such as crack length and width through simple morphological operations, lacking a semantic understanding of the overall crack structure and failing to automatically distinguish topological patterns such as linear cracks, mesh cracks, and radial cracks. This invention constructs a crack topology network model, transforming the crack skeleton into a graph structure, and extracts topological features such as node degree distribution, number of connected components, and number of ring structures to achieve automatic crack pattern recognition and intelligent assessment of structural safety levels.

[0207] This invention addresses the problems of passive reference object calibration processes, which are susceptible to interference from shooting conditions. Existing calibration methods require manual measurement of the reference object's dimensions and manual input into the system, and cannot automatically assess factors affecting calibration accuracy such as shooting angle distortion and illumination uniformity. This invention designs an adaptive reference object with coded information. By recognizing the coded pattern on the reference object, the system automatically extracts the reference object's specifications. Simultaneously, based on indicators such as the quality of reference object corner detection, checkerboard integrity, and grayscale uniformity between corner points, the system automatically determines the image shooting quality and adaptively corrects perspective distortion, achieving intelligent calibration and ensuring the reliability of calibration results.

[0208] This invention addresses the issues of incomplete crack development trend prediction and delayed structural safety early warning. It establishes a time-series model of key crack parameters (maximum width, total length, number of nodes, and topological complexity index), combines this model with environmental temperature and humidity data, predicts crack development trends over a future period, and automatically outputs safety risk levels and inspection recommendations based on structural engineering design specifications.

[0209] Compared with existing technologies, the intelligent crack detection method and system based on adaptive coding reference calibration and topology analysis proposed in this invention have the following significant advantages:

[0210] The detection accuracy is significantly improved. By employing a multi-level feature fusion network, edge gradient response maps, local binary pattern texture feature maps, and local contrast enhancement maps from traditional image processing are used as auxiliary feature channels. These are dynamically gated and fused with deep semantic features extracted by the backbone network, effectively enhancing the model's ability to detect minute cracks under low contrast and low illumination conditions. Experiments show that on the test dataset, the method of this invention improves the recall rate for minute cracks with a width less than 0.1 mm by 12.3 percentage points compared to the baseline Mask R-CNN model, achieving an overall mAP@0.5 of 92.4% and a segmentation IoU of 88.7%.

[0211] Automatic analysis of crack topology. This invention innovatively transforms crack segmentation results into a graph structure representation. Through skeleton extraction, node detection (endpoint and intersection identification), and edge connection reconstruction, a crack topology network is constructed. Based on the topology network characteristics, it can automatically identify linear cracks, mesh cracks, radial cracks, and other patterns, with an accuracy rate of 91.5%. The topology analysis results can directly support structural engineers in making preliminary judgments about the causes of cracks; for example, mesh cracks often indicate material aging, while radial cracks often indicate local stress concentration, significantly improving the engineering application value of the detection results.

[0212] The calibration process is intelligent, and the measurement results are more reliable. The adaptive coded reference object designed in this invention integrates digital coded identifiers (such as AprilTag or QR codes). The system can automatically obtain the reference object's specifications (side length, single-cell size) by recognizing the coded pattern, eliminating the need for manual input. Simultaneously, based on indicators such as the reference object's corner detection confidence level, the consistency of the checkerboard grid corner spacing, and the grayscale uniformity of the reference object area, the system automatically evaluates the image capture quality: when the detection quality does not meet a preset threshold, the user is prompted to retake the image; when the quality meets the standard, the system automatically calculates the perspective transformation matrix based on the reference object's four corner coordinates, adaptively correcting the shooting angle distortion and eliminating measurement errors introduced by non-orthogonal shooting. Experiments show that after adopting adaptive calibration, the crack width measurement error decreased from ±0.15mm to ±0.08mm, and the length measurement error decreased from ±3.2% to ±1.5%.

[0213] Crack Development Trend Prediction and Risk Warning. This invention establishes a time-series prediction model for key crack parameters. When historical inspection data exists for the same wall area, the development trend of cracks can be predicted over the next 30, 90, and 180 days based on historical time-series data and environmental temperature and humidity data. The prediction model employs a hybrid prediction strategy combining the Prophet time-series decomposition algorithm with an empirical crack propagation rate model (Paris formula) from the field of structural health monitoring. The mean absolute percentage error (MAPE) of the prediction results is less than 15%. Based on the prediction results and the "Standard for Reliability Appraisal of Civil Buildings" (GB 50292-2015), the structural safety risk level is automatically determined. The structural safety risk levels include: Level A (Safe); Level B (Basic Safe); Level C (Restricted Use); and Level D (Dangerous). Differentiated inspection suggestions and early warning information are generated, realizing a shift from post-detection to pre-detection.

[0214] The system boasts high integration and strong engineering applicability. This invention provides a complete end-to-end solution, covering the entire process of image acquisition quality assessment, adaptive calibration, crack detection and segmentation, parameter quantization measurement, topological pattern recognition, cause-assisted determination, time-series trend prediction, and risk warning output. The system provides services through a RESTful API interface, supports Docker containerized deployment, and has an average single-instance response time of less than 1.5 seconds (1920×1080 resolution image), meeting the engineering application needs of large-scale building inspection.

[0215] Secondly, this application proposes an intelligent crack detection device 600, see appendix. Figure 6 ,include:

[0216] Acquisition module 61 is used to acquire the original image of the crack;

[0217] Image processing module 62 is used to determine the quality of the original image of the crack based on the confidence level of the corner detection of the reference object, the consistency of the spacing between the corner points of the checkerboard, and the gray-scale uniformity of the reference object area; in response to the fact that the quality of the original image of the crack is qualified, the original image of the crack is corrected to obtain the corrected image of the crack.

[0218] The feature determination module 63 is used to determine the multi-level features of the crack in the corrected image of the crack. The multi-level features of the crack include local binary pattern texture features, edge gradient response features and local contrast enhancement features.

[0219] The feature fusion module 64 is used to dynamically weight and fuse the multi-level features of the crack to obtain the fused features of the crack.

[0220] The hybrid prediction module 65 uses the Prophet time series decomposition algorithm and the crack propagation rate empirical model to provide risk warnings for cracks based on crack parameters and fusion characteristics.

[0221] In some embodiments, the hybrid prediction module 65 is further configured to determine, for any sampling time point, the first dynamic weight parameter of the Prophet time series decomposition algorithm and the second dynamic weight parameter of the crack propagation rate empirical model, respectively.

[0222] Wherein, the sum of the first dynamic weight parameter and the second dynamic weight parameter is 1;

[0223] The Prophet time series decomposition algorithm is used to determine the first prediction result based on historical data, current data, prediction days, and environmental data of the crack parameters;

[0224] Calculate the product of the first dynamic weight parameter and the first prediction result to obtain the first product;

[0225] The second prediction result was obtained by calculating using an empirical model of crack propagation rate.

[0226] Calculate the product of the second dynamic weight parameter and the second prediction result to obtain the second product;

[0227] The first and second products are summed to obtain the mixed prediction results of the parameters at the sampling time points.

[0228] In some embodiments, the hybrid prediction module 65 is further configured to determine, for any sampling time point and for any parameter, the first dynamic weight parameter of the Prophet time series decomposition algorithm, including:

[0229] Determine the basic weight parameters and normalized uncertain weight parameters of the crack at the sampling time point;

[0230] Based on the basic weight parameters and the normalized uncertain weight parameters, the first dynamic weight parameters of the Prophet time series decomposition algorithm are determined.

[0231] In some embodiments, the hybrid prediction module 65 is further configured to determine the parameter value of the parameter at the sampling time point, the historical mean of the parameter at historical sampling time points, and the historical mean square deviation of the parameter, respectively.

[0232] Based on the parameter values ​​at the sampling time point, the historical mean and historical root mean square deviation of the parameter at historical sampling time points, the normalized uncertain weight parameters of the parameters at the sampling time point of the crack are determined.

[0233] In some embodiments, a type determination module is also included for determining the topology of the cracks based on the fusion characteristics of the cracks;

[0234] Determine the crack type and judgment parameters based on the crack topology diagram;

[0235] Determine the type of crack based on crack type judgment parameters;

[0236] The parameters for determining the crack type include: node degree distribution, connected components, number of cycles, and topological complexity index.

[0237] In some embodiments, the type determination module is further configured to determine the type of crack as a mesh crack in response to the ring number being greater than zero and the topology complexity index being greater than a predetermined first topology complexity index threshold.

[0238] In response to a node degree distribution greater than a predetermined node degree distribution threshold, the number of rings equals zero.

[0239] If the topological complexity index is less than or equal to the first topological complexity index threshold and greater than or equal to the second topological complexity index threshold, the crack type is radial crack.

[0240] In some embodiments, the type determination module is also used for TCI=α×(k_max / V)+β×(Cycles / E)+γ×(1 / C);

[0241] Where α is the first weight coefficient; β is the second weight coefficient; γ is the third weight coefficient; k_max is the maximum node degree; V is the total number of nodes; Cycles is the number of cycles; C is the number of connected components; and E is the total number of edges.

[0242] Thirdly, this application proposes an electronic device comprising: at least one processor and at least one memory; the memory for storing one or more program instructions; and the processor for executing one or more program instructions to perform an intelligent crack detection method as described above.

[0243] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A smart crack detection method, characterized in that, include: Obtain the original image of the crack; The quality of the original image of the crack is determined based on the confidence level of the corner detection of the reference object, the consistency of the spacing between the corner points of the checkerboard, and the gray-level uniformity of the reference object area. In response to the original image of the crack being of acceptable quality, the original image of the crack is corrected to obtain a corrected image of the crack; Determine the multi-level features of the crack in the corrected image of the crack, the multi-level features of the crack including local binary pattern texture features, edge gradient response features and local contrast enhancement features; The multi-level features of the crack are dynamically weighted and fused to obtain the fused features of the crack; Based on the parameters and fusion characteristics of the crack, the Prophet time series decomposition algorithm and the crack propagation rate empirical model are used to provide risk warning for the crack; The parameters include: the maximum width of the crack, the total length, the number of nodes, and the topological complexity index.

2. The intelligent crack detection method according to claim 1, characterized in that, The Prophet time series decomposition algorithm and an empirical model of crack propagation rate are used to provide risk warnings for the cracks, including: For any sampling time point, determine the first dynamic weight parameter of the Prophet time series decomposition algorithm and the second dynamic weight parameter of the crack propagation rate empirical model respectively; Wherein, the sum of the first dynamic weight parameter and the second dynamic weight parameter is 1; The Prophet time series decomposition algorithm is used to determine the first prediction result based on the historical data, current data, prediction days, and environmental data of the parameters of the crack. Calculate the product of the first dynamic weight parameter and the first prediction result to obtain the first product; The second prediction result was obtained by calculating using an empirical model of crack propagation rate. Calculate the product of the second dynamic weight parameter and the second prediction result to obtain the second product; The first product and the second product are summed to obtain the mixed prediction result of the parameters at the sampling time point.

3. The intelligent crack detection method according to claim 2, characterized in that, The method further includes: For any sampling time point and any parameter, the first dynamic weight parameter of the Prophet time series decomposition algorithm is determined, including: Determine the basic weight parameters and normalized uncertain weight parameters of the parameters of the crack at the sampling time point; The first dynamic weight parameter of the Prophet time series decomposition algorithm is determined based on the basic weight parameter and the normalized uncertain weight parameter.

4. The intelligent crack detection method according to claim 3, characterized in that, Determining the normalized uncertain weight parameters of the parameters of the crack at the sampling time point specifically includes: Determine the parameter value at the sampling time point, the historical mean and historical root mean square deviation of the parameter at historical sampling time points, respectively; Based on the parameter value of the parameter at the sampling time point, the historical mean and historical root mean square deviation of the parameter at historical sampling time points, the normalized uncertain weight parameter of the parameter at the sampling time point of the crack is determined.

5. The intelligent crack detection method according to claim 1, characterized in that, The method further includes: Determining the type of crack based on the fusion characteristics of the crack includes: The topology of the cracks is determined based on their fusion characteristics; The crack type determination parameters are determined based on the topology of the crack. The type of crack is determined based on the crack type determination parameters; The parameters for determining the crack type include: node degree distribution, connected components, number of cycles, and topological complexity index.

6. The intelligent crack detection method according to claim 5, characterized in that, Determining the type of crack based on the crack type determination parameters includes: In response to the ring number being greater than zero and the topology complexity index being greater than a predetermined first topology complexity index threshold, the type of the crack is a mesh crack; In response to the node degree distribution being greater than a predetermined node degree distribution threshold, the number of rings equals zero. The topological complexity index is less than or equal to the first topological complexity index threshold and greater than or equal to the second topological complexity index threshold, and the type of crack is a radial crack.

7. The intelligent crack detection method according to claim 5, characterized in that, Determining the topological complexity index includes: TCI=α×(k_max / V)+β×(Cycles / E)+γ×(1 / C); Where α is the first weight coefficient; β is the second weight coefficient; γ is the third weight coefficient; k_max is the maximum node degree; V is the total number of nodes; Cycles is the number of cycles; C is the number of connected components; and E is the total number of edges.

8. The intelligent crack detection method according to claim 1, characterized in that, Correcting the original image of the crack to obtain a corrected image of the crack includes: Determine the coordinates of each pixel in the original image, and denot them as the first coordinate set; Determine the homography transformation matrix, which represents the point-to-point mapping relationship between the original image plane and the target correction plane; Based on the homography transformation matrix and the first coordinate set, determine the second coordinate set in the target correction plane; The second coordinate set is adjusted to obtain the third coordinate set, which specifically includes: Based on the homography transformation matrix, determine the projection points of the four vertices of the original image onto the target correction plane; In the target correction plane, determine the coordinates of each projection point of the four vertices; Determine the minimum x-coordinate and minimum y-coordinate of the coordinates of each projection point of the four vertices; Based on the minimum values ​​of the horizontal and vertical coordinates, the coordinates of each pixel in the second coordinate set are translated to obtain the third coordinate set. Based on the grayscale value of each pixel in the original image and the third coordinate set, the grayscale value of each pixel in the third coordinate set is determined by a bilinear interpolation algorithm. The corrected image is generated based on the grayscale value of each pixel and the third coordinate set.

9. An intelligent crack detection device, characterized in that, include: The acquisition module is used to acquire the original image of the crack; The image processing module is used to determine the quality of the original image of the crack based on the confidence level of the corner detection of the reference object, the consistency of the spacing between the corner points of the checkerboard, and the gray-level uniformity of the reference object area; in response to the fact that the quality of the original image of the crack is qualified, the original image of the crack is corrected to obtain the corrected image of the crack. The feature determination module is used to determine the multi-level features of the crack in the corrected image of the crack, wherein the multi-level features of the crack include local binary pattern texture features, edge gradient response features and local contrast enhancement features. The feature fusion module is used to dynamically weight and fuse the multi-level features of the crack to obtain the fused features of the crack; The hybrid prediction module, based on the parameters of the crack and the fusion characteristics, uses the Prophet time series decomposition algorithm and the crack propagation rate empirical model to provide risk warnings for the crack; The parameters include: the maximum width of the crack, the total length, the number of nodes, and the topological complexity index.

10. An electronic device, characterized in that, include: At least one processor and at least one memory; The memory is used to store one or more program instructions; The processor is configured to run one or more program instructions to perform the intelligent crack detection method as described in any one of claims 1 to 8.

Images