A Dynamic Detection Method and System for Bridge Construction Cracks Based on AI Visual Recognition

Abstract

This invention relates to the field of bridge construction monitoring technology, and discloses a method and system for dynamic detection of cracks in bridge construction based on AI visual recognition. The method analyzes a sequence of multiple consecutive images, performing pixel-level temporal and spatial feature separation and fusion to generate a target structural feature map incorporating spatiotemporal context information, thereby locating potential crack regions. A region evolution model is constructed by analyzing the multi-directional discrete variation law of pixel intensity in candidate regions, and a crack growth prediction model is used to output a quantitative description of crack growth trends. A dynamic structural integrity score is generated by combining growth trends and multi-scale texture analysis, enabling real-time crack detection and evaluation. This method improves the accuracy and robustness of crack detection in complex construction environments and can effectively predict the dynamic development trend of cracks.

Description

Technical Field

[0001] This invention relates to the field of bridge construction monitoring technology, specifically to a method and system for dynamic detection of bridge construction cracks based on AI visual recognition. Background Technology

[0002] Currently, the detection of surface cracks in bridge structures during construction mainly relies on manual inspections or static analysis methods based on digital images. Manual inspections depend on the experience of inspectors, resulting in low efficiency, high subjectivity, and difficulty in achieving 24 / 7 coverage. Conventional digital image-based techniques, such as edge detection operators, threshold segmentation, or texture analysis, can process single images and identify obvious crack features under specific conditions.

[0003] These conventional technical solutions have shortcomings. Static image analysis methods struggle to effectively distinguish linear interference common in the construction environment that resembles crack textures, leading to a high false detection rate. Static methods are completely unable to capture the dynamic evolution of cracks under the influence of factors such as construction loads and temperature changes; they can only provide an instantaneous judgment of "presence or absence," failing to assess their development trend and thus unable to meet the early warning needs of dynamic safety monitoring during the construction process.

[0004] There is a need for a detection method that can overcome interference from complex environments and achieve dynamic behavior analysis and prediction of cracks. This requires that the technical solution not only more robustly separates the real crack features from the sequence images, but also establishes a model to describe and predict the spatiotemporal evolution of cracks. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for dynamic detection of cracks in bridge construction based on AI visual recognition, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for dynamic detection of bridge construction cracks based on AI visual recognition, the method comprising:

[0007] Acquire multiple consecutive images containing the surface of the construction structure to form an image monitoring sequence;

[0008] Pixel-level temporal and spatial feature separation is performed on the image monitoring sequence to obtain temporal separation features and spatial separation features;

[0009] The temporal and spatial separation features are input into a pre-defined spatiotemporal fusion network to obtain a target structure feature map that incorporates spatiotemporal context information.

[0010] Calculate the texture coherence and abrupt change intensity of the target structure feature map at multiple scales to generate a texture scale sequence;

[0011] Based on the texture scale sequence, the distribution area of ​​potential cracks in the image monitoring sequence is determined, forming a preliminary set of candidate regions;

[0012] For each preliminary candidate region set, the discrete variation law of pixel intensity along multiple directions within it is analyzed, and a region evolution model is constructed.

[0013] The preliminary candidate region set and its regional evolution model are input into the crack growth prediction model, and a quantitative description of the crack growth trend is output.

[0014] By combining a quantitative description of crack growth trends with a texture scale sequence, a dynamic score of structural integrity for the image monitoring sequence is calculated.

[0015] Based on the dynamic scoring of structural integrity, real-time detection results of bridge construction cracks are generated.

[0016] Preferably, the method for obtaining the temporal domain separation features and spatial domain separation features is as follows:

[0017] Extract the grayscale value changes of corresponding pixels between adjacent image frames in the image monitoring sequence, and construct a three-dimensional temporal domain change volume by arranging the grayscale value changes in time sequence.

[0018] A directional filtering process is applied to the three-dimensional time-domain variable volume, and its energy accumulation in the vertical, horizontal and diagonal directions is calculated. The energy accumulation is used as a time-domain separation feature.

[0019] For each frame of the image monitoring sequence, its gradient magnitude map and direction map are calculated. Multiple rounds of convolution and pooling operations are performed on the gradient magnitude map and direction map respectively to extract deep spatial descriptors that characterize the local structure of the image. The deep spatial descriptors are used as spatial separation features.

[0020] Preferably, the method for obtaining the target structural feature map is as follows:

[0021] A spatiotemporal fusion network is pre-constructed, which includes a feature alignment module, a bidirectional attention module, and a feature reconstruction module;

[0022] Temporal and spatial separation features are input into the feature alignment module of the spatiotemporal fusion network. Based on optical flow estimation, the temporal and spatial separation features are corrected for their spatiotemporal positions to obtain aligned spatiotemporal feature pairs. The aligned spatiotemporal feature pairs are then input into the bidirectional attention module of the spatiotemporal fusion network to calculate the attention weights of the temporal and spatial separation features on the spatial separation features and vice versa. The aligned spatiotemporal feature pairs are then weighted and fused based on these two attention weights to generate fused features. The fused features are then input into the feature reconstruction module of the spatiotemporal fusion network to restore the resolution of the fused features to the original image size through deconvolution and upsampling operations, resulting in the target structure feature map.

[0023] Preferably, the method for generating the texture scale sequence is as follows:

[0024] Gaussian pyramid decomposition is performed on the target structure feature map to obtain feature sub-maps at multiple different scales. For each feature sub-map at each scale, its local binary pattern histogram is calculated, and the local binary pattern histograms at each scale are concatenated to form a multi-scale texture histogram. The chi-square distance between adjacent scale histograms in the multi-scale texture histogram is calculated, and the chi-square distance is used as a measure of texture coherence between scales.

[0025] Simultaneously, the entropy of the gradient magnitude map of the feature sub-map at each scale is calculated, and the entropy is used as the texture mutation intensity at each scale. The texture coherence measure and texture mutation intensity at each scale are arranged in scale order to form a texture scale sequence.

[0026] Preferably, the method for forming the preliminary candidate region set is as follows:

[0027] Extract scale information where the intensity of texture mutation exceeds a preset threshold from the texture scale sequence, locate high mutation regions on the corresponding scale feature sub-map of the target structure feature map, calculate the average gradient direction consistency of the high mutation regions on the full scale of the target structure feature map, and select regions with gradient direction consistency lower than a preset consistency threshold as seed regions.

[0028] Centered on the seed region, region growth is performed on the target structure feature map. The growth rule combines the local values ​​of pixel gray-level similarity and texture coherence measurement in the texture scale sequence. The growth termination condition is when the image boundary is encountered or the gray-level gradient exceeds the preset growth threshold. All grown regions are taken as preliminary candidate regions, and the coordinate set of all preliminary candidate regions is recorded as the preliminary candidate region set.

[0029] Preferably, the method for constructing the regional evolution model is as follows:

[0030] For each preliminary candidate region in the preliminary candidate region set, the position and shape of the preliminary candidate region are tracked in multiple consecutive frames of the image monitoring sequence. The Hausdorff distance change rate of the contour point set of the preliminary candidate region in consecutive frames and the autocorrelation coefficient of the mean gray value of the pixels in the region are calculated. The area change of the preliminary candidate region, the Hausdorff distance change rate of the contour point set, and the autocorrelation coefficient of the mean gray value of the pixels are used as state variables. A Kalman filter is used to model and predict the state variables. The predicted state sequence output by the Kalman filter is the region evolution model of the preliminary candidate region.

[0031] Preferably, the method for obtaining the quantitative description of the crack growth trend is as follows:

[0032] The crack growth prediction model receives a preliminary set of candidate regions and their regional evolution models as input. The crack growth prediction model contains a combination structure of a long short-term memory network and a fully connected layer. The long short-term memory network is used to encode the time series state in the regional evolution model, and the fully connected layer is used to map the encoding results into a quantitative description of the crack growth trend. The quantitative description of the crack growth trend includes the crack length change rate, the crack width expansion probability, and the predicted value of the crack extension direction angle.

[0033] Preferably, the method for calculating the dynamic score of structural integrity is as follows:

[0034] The crack length change rate and crack width expansion probability are extracted from the quantitative description of crack growth trend. The ratio of the total area of ​​the preliminary candidate region set in the current frame image to the total area of ​​the visible area on the surface of the construction structure is calculated as the crack area proportion. The minimum value of the texture coherence metric is extracted from the texture scale sequence. The minimum values ​​of crack length change rate, crack width expansion probability, crack area proportion, and texture coherence metric are input into a preset evaluation function. The evaluation function is a weighted sum of the above four input values. The weighting coefficients are learned from historical crack-free samples. The output value of the evaluation function is the dynamic score of structural integrity.

[0035] Preferably, the method for generating the real-time detection results is as follows:

[0036] The structural integrity dynamic score is compared with multiple preset dynamic thresholds, which are preset according to the construction stage and ambient lighting conditions. Based on the comparison results, it is determined whether there are cracks exceeding the safety standard in the current image monitoring sequence. If so, the crack location, growth trend and risk level are marked on the image by combining the quantitative description of the crack growth trend and the coordinates of the preliminary candidate area set, and a real-time detection result report containing visual annotations and risk level information is generated.

[0037] Preferably, the present invention also includes a dynamic detection system for bridge construction cracks based on AI visual recognition. The system includes an image acquisition device, a computing and processing device, and a result output device. The computing and processing device is used to execute instructions to realize the steps in the above-described dynamic detection method for bridge construction cracks based on AI visual recognition.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] By performing pixel-level temporal and spatial feature separation on multiple consecutive images and integrating them using a spatiotemporal fusion network, this method can effectively remove noise caused by changes in illumination, equipment micro-movements, and non-structural linear interference. Temporal features reveal the variation patterns of pixel values ​​over time, while spatial features preserve the structural information of texture and shape. The fusion of these two features enables the generated target structural feature map to have higher recognition accuracy for static cracks and slowly developing micro-cracks in dynamic construction environments, reducing false alarms caused by environmental factors.

[0040] For the initially selected candidate regions, the discrete variation patterns of pixel intensity along multiple directions are analyzed, and a regional evolution model is constructed. This model is then processed using a crack growth prediction model. This process transforms the crack from a static "object" into a "process" with dynamic evolution patterns. The variation patterns of pixel intensity in multiple directions quantify the expansion tendency and pattern of the crack edge, and the growth prediction model outputs a quantitative description of its development trend accordingly. This enables the detection system to go beyond immediate state judgment and predict the future expansion length, direction, or morphology of cracks, providing a technical basis for proactive intervention and risk warning during construction.

[0041] By combining crack growth trend prediction with multi-scale texture analysis results, the calculated structural integrity dynamic score is a comprehensive evaluation index that integrates the current status and expected evolution. This score reflects the potential risk level of the monitored area in real time, transforming discrete crack identification information into continuous, measurable safety status parameters. This represents a leap from qualitative description to quantitative assessment of monitoring results, supporting more refined construction safety management decisions. Attached Figure Description

[0042] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based visual recognition-based dynamic detection method for bridge construction cracks as described in this invention.

[0043] Figure 2 A flowchart for obtaining the target structure feature map;

[0044] Figure 3 A flowchart for generating texture scale sequences;

[0045] Figure 4A multi-dimensional radar image showing the characteristics of the actual crack area and the noise area in bridge construction crack detection.

[0046] Figure 5 A bar chart comparing the spatiotemporal fusion performance of bridge crack detection at different feature scales. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figure 1 This invention provides a method for dynamic detection of bridge construction cracks based on AI visual recognition. The method includes: firstly, acquiring multiple consecutive images of the bridge construction structure surface using an image acquisition device to form an image monitoring sequence; then, performing pixel-level temporal and spatial feature separation operations on the image monitoring sequence to obtain temporal separation features representing temporal changes and spatial separation features representing the structure of a single frame image, respectively; next, inputting these two types of features into a preset spatiotemporal fusion network, achieving deep fusion of spatiotemporal context information through the network's internal processing, and outputting a target structure feature map that integrates this information; based on this, calculating the texture coherence and abrupt change intensity indices of the target structure feature map at multiple scales, and arranging them in scale order to generate a texture scale sequence; and then, based on the texture scale sequence, calculating the texture coherence and abrupt change intensity indices of the target structure feature map at multiple scales. Texture mutations and fracture information are used to determine the distribution area of ​​potential cracks in the image monitoring sequence, forming several preliminary candidate region sets containing regional coordinate information. For each preliminary candidate region set, the discrete variation law of its internal pixel intensity along multiple directions in continuous image frames is analyzed, and a region evolution model describing the morphology and grayscale evolution of the region is constructed. The preliminary candidate region set and its corresponding region evolution model are input into a pre-trained crack growth prediction model, which outputs a quantitative description of the future growth trend of cracks. Combining this quantitative description with the previously generated texture scale sequence, a dynamic score representing the current structural integrity is obtained through specific calculations. Finally, based on this dynamic score of structural integrity, real-time detection results of bridge construction cracks, including crack location, trend, and risk level, are generated.

[0049] In one embodiment of the present invention, the method for obtaining temporal and spatial separation features is as follows: extract the grayscale value changes of corresponding pixels between adjacent image frames in the image monitoring sequence, combine these grayscale value changes arranged in chronological order to construct a three-dimensional temporal change volume; perform directional filtering on the three-dimensional temporal change volume, calculate its energy accumulation values ​​in the vertical, horizontal and diagonal directions respectively, and use these directional energy accumulation values ​​as temporal separation features; for each frame of the image monitoring sequence, calculate its gradient magnitude map and gradient direction map respectively, and perform multiple rounds of convolution and pooling operations on the obtained gradient magnitude map and direction map respectively, thereby extracting deep spatial descriptors that can characterize the local fine structure of the image, and use these deep spatial descriptors as spatial separation features.

[0050] In practical implementation, temporal and spatial separation features are acquired. For the acquisition of temporal separation features, the gray value changes of corresponding pixels between adjacent image frames in the image monitoring sequence are extracted. In the example scenario, the image monitoring sequence comes from continuous monitoring videos of the surface of the bridge construction structure. The gray value changes reflect the slight brightness differences on the surface of the construction structure caused by the generation or expansion of cracks. The gray value changes are arranged in time sequence to construct a three-dimensional temporal change volume. The three-dimensional temporal change volume is a three-dimensional matrix, whose two spatial dimensions correspond to the image pixel coordinates and the time dimension corresponds to the image frame sequence order. In the data comparison, there are differences in the gray value change patterns between the crack area and the non-crack area. The gray value changes in the crack area often show stronger coherence and directionality in the time series. In some embodiments, directional filtering is performed on the three-dimensional temporal volume. This directional filtering uses a set of preset directional filter kernels, corresponding to the vertical, horizontal, and diagonal directions, respectively. The energy accumulation of the three-dimensional temporal volume in each direction is calculated. The energy accumulation is obtained by summing the absolute values ​​after convolution operations. These directional energy accumulation values ​​are used as temporal separation features. In the data comparison of the example scene, the vertical energy accumulation value of the crack region is usually higher than that of the non-crack region because cracks often extend longitudinally, resulting in more significant changes in grayscale values ​​in the vertical direction. Optionally, the directional filtering can adjust the size of the filter kernels to match image monitoring sequences of different resolutions. The size of the filter kernels is determined based on the performance of the image acquisition device and the required detection accuracy.

[0051] In specific implementation, for the acquisition of spatial separation features, for each frame of the image monitoring sequence, its gradient magnitude map and gradient direction map are calculated. The gradient magnitude map calculates the gradient intensity of each pixel using the Sobel operator or a similar edge detection operator, and the gradient direction map records the angle of the gradient vector of each pixel. Multiple rounds of convolution and pooling operations are performed on the gradient magnitude map and gradient direction map respectively. These multiple rounds of convolution and pooling operations use a convolutional neural network structure, including multiple alternating convolutional and pooling layers, to extract deep spatial descriptors characterizing the local structure of the image. In some embodiments, the deep spatial descriptors can be the output of high-level feature maps in the convolutional neural network. These feature maps capture the texture and edge patterns in the image. In the data comparison of the example scene, crack edges exhibit high response values ​​on the gradient magnitude map, and the deep spatial descriptors further enhance these edge features and suppress interference from uniform background areas through multiple rounds of convolution and pooling operations. Optionally, the specific parameters of the multiple rounds of convolution and pooling operations, such as the number of convolutional kernels and the pooling window size, can be optimized based on the training data to adapt to the surface texture of different bridge construction materials.

[0052] In practical implementation, the acquisition processes of temporal and spatial separation features can be executed in parallel to improve processing efficiency. Temporal separation features focus on the change patterns in the time dimension, while spatial separation features focus on structural information in the spatial dimension. The combination of both provides input for subsequent feature fusion. In the data comparison of the example scenario, using both temporal and spatial separation features simultaneously can more comprehensively describe the spatiotemporal characteristics of the crack, such as the temporal changes in grayscale values ​​during crack propagation and the edge spatial distribution in a single frame image. To quantify the energy accumulation in directional filtering, the following formula is introduced:

[0053]

[0054] in: Indicates direction The cumulative energy value on the surface, Represents a three-dimensional time-domain variable volume in spatial coordinates and time Changes in grayscale values ​​at that location Indicates direction The corresponding filter core, This represents the convolution operation. This indicates taking the absolute value and summing it to cover all spatial locations and time frames. In specific implementations, the direction... Including vertical, horizontal, and diagonal directions, filter kernel Based on the directional design, for example, the vertical direction filter kernel emphasizes the vertical edge, and the horizontal direction filter kernel emphasizes the horizontal edge. In data comparison, the distribution of energy accumulation values ​​in different directions can be used to distinguish crack orientation from non-crack noise.

[0055] In one embodiment of the present invention, see [reference] Figure 2 The acquisition of the target structure feature map relies on a pre-constructed spatiotemporal fusion network, which includes a feature alignment module, a bidirectional attention module, and a feature reconstruction module. Temporal and spatial separation features are input into the feature alignment module, which corrects the spatiotemporal positions of the two types of features based on optical flow estimation, outputting aligned spatiotemporal feature pairs. These aligned spatiotemporal feature pairs are then input into the bidirectional attention module, which calculates the attention weights of the temporal and spatial separation features on top of each other. The aligned spatiotemporal feature pairs are then weighted and fused based on these two attention weights to generate fused features. Finally, the fused features are input into the feature reconstruction module, which gradually restores the resolution of the fused features to the size of the original image through deconvolution and upsampling operations, ultimately obtaining the target structure feature map.

[0056] In practical implementation, the acquisition of target structural feature maps relies on a pre-constructed spatiotemporal fusion network. The spatiotemporal fusion network consists of three components: a feature alignment module, a bidirectional attention module, and a feature reconstruction module. In the example scenario of bridge construction crack detection, the input temporal separation features reflect the temporal change energy of the gray values ​​on the surface of the construction structure, while the spatial separation features characterize the gradient and deep texture of the structure in a single frame image. Data comparison shows that the crack area exhibits energy concentration in a specific direction in the temporal separation features, and shows significant edge response in the spatial separation features. The role of the spatiotemporal fusion network is to correct, weight, and integrate these two features in the spatiotemporal dimension. In some embodiments, temporal and spatial separation features are input into the feature alignment module of the spatiotemporal fusion network. The feature alignment module performs spatiotemporal position correction on the temporal and spatial separation features based on optical flow estimation. Optical flow estimation calculates the motion vectors of pixels between adjacent feature maps, and uses the motion vectors to perform deformation interpolation on the feature maps to achieve coordinate alignment, resulting in aligned spatiotemporal feature pairs. In the example scenario, due to slight shaking of the image acquisition device or the construction structure itself, the feature alignment module can effectively compensate for this positional offset, so that crack features from the same spatial location at different times can be accurately matched. Data comparison shows that direct fusion of unaligned features leads to blurred edge information, while the edge clarity of aligned features is maintained.

[0057] In practical implementation, the aligned spatiotemporal feature pairs are input into the bidirectional attention module of the spatiotemporal fusion network. The bidirectional attention module calculates the attention weights of temporal-separated features on spatial-separated features, and vice versa. This calculation involves converting the feature map into query vectors, key vectors, and value vectors, followed by dot product operations and normalization. Based on these two attention weights, the aligned spatiotemporal feature pairs are weighted and fused to generate fused features. In the example scenario of bridge construction cracks, spatiotemporal-separated features emphasize the crack's dynamic expansion, while spatial-separated features emphasize its static shape. The bidirectional attention mechanism allows the model to dynamically determine which information to rely on more during the fusion process. Data comparison shows that in the crack tip region, the spatiotemporal feature attention weight is higher to capture the expansion trend, while in the main crack region, the spatial feature attention weight is higher to maintain morphological integrity. Optionally, the bidirectional attention module can employ a multi-head attention mechanism, which allows the model to focus on information from different parts of the spatiotemporal features in different representation subspaces.

[0058] In specific implementation, the fused features are input into the feature reconstruction module of the spatiotemporal fusion network. The feature reconstruction module restores the resolution of the fused features to the original image size through deconvolution and upsampling operations. The deconvolution operation increases the spatial size of the feature map and learns to reconstruct details through transposed convolutional layers. The upsampling operation expands the size of the feature map through interpolation methods to obtain the target structure feature map. In some embodiments, the feature reconstruction module includes a combination of multiple deconvolutional layers and upsampling layers to gradually restore the resolution of the feature map. In the example scenario, the fused features are processed through multiple layers to finally output a target structure feature map of the same size as the input image. This feature map has a high response value in the crack region and a lower response value in the non-crack region. Data comparison shows that the target structure feature map contains richer discriminative information than a single temporal or spatial feature map. Optionally, an activation function, such as the sigmoid function, can be connected to the feature reconstruction module at the end to map the numerical range of the feature map to between 0 and 1, which is convenient for subsequent processing. It can be understood that deconvolution and upsampling operations are key steps in restoring spatial details, enabling the target structure feature map to retain fine-grained information for accurate region localization. To quantify the weighted fusion process in the bidirectional attention module, the following formula is introduced:

[0059]

[0060] in: Indicates the generated fusion features, This indicates the spatiotemporal separation characteristics after alignment. This indicates the spatial separation feature after alignment. This represents the attention weight matrix of spatiotemporal separation features to spatial separation features. This represents the attention weight matrix of spatial separation features to spatiotemporal separation features. This represents element-wise multiplication. In practice, the attention weight matrix... and The feature similarity is normalized using the Softmax function to ensure that the sum of the weights at each location is 1. In the data comparison, there is a significant difference in the weight distribution between the crack region and the background region, which directly leads to the enhancement of the fused features in the crack region.

[0061] In one embodiment of the present invention, see [reference] Figure 3 The process of generating the texture scale sequence is as follows: Gaussian pyramid decomposition is performed on the target structural feature map to obtain feature sub-maps at multiple different scales; for each feature sub-map at each scale, its local binary pattern histogram is calculated, and the local binary pattern histograms at all scales are concatenated to form a multi-scale texture histogram; the chi-square distance between adjacent scale histograms in this multi-scale texture histogram is calculated, and this chi-square distance is used as a measure of texture coherence across different scales; simultaneously, the entropy value of the gradient magnitude map corresponding to each scale feature sub-map is calculated, and this entropy value is used as the texture abrupt change intensity at each scale; the texture coherence measure and texture abrupt change intensity at each scale are arranged in scale order to form the texture scale sequence. The process of forming the preliminary candidate region set is as follows: extract scale information where the intensity of texture mutation exceeds a preset threshold from the texture scale sequence, and locate high mutation regions on the corresponding scale feature sub-map of the target structural feature map; calculate the average gradient direction consistency of these high mutation regions across the entire scale of the target structural feature map, and select regions whose gradient direction consistency is lower than a preset consistency threshold as seed regions; perform region growth operations on the target structural feature map centered on these seed regions, with the growth rules combining local values ​​of pixel gray-level similarity and texture coherence measurement in the texture scale sequence, and setting the growth termination condition as encountering an image boundary or pixel gray-level gradient exceeding a preset growth threshold; mark all grown regions as preliminary candidate regions, and record the coordinate set of all preliminary candidate regions as the preliminary candidate region set.

[0062] In practice, the generation of the texture scale sequence begins with Gaussian pyramid decomposition of the target structure feature map. Gaussian pyramid decomposition, through continuous application of Gaussian smoothing and downsampling operations, yields feature sub-maps of the target structure feature map at multiple different scales. In the example scenario of bridge construction crack detection, cracks exhibit different texture characteristics at different observation scales. Larger-scale feature sub-maps reflect the overall direction and distribution of the cracks, while smaller-scale feature sub-maps capture subtle irregularities at the crack edges. Data comparison shows that crack regions cause significant changes in texture features at multiple scales, while non-crack noise is usually more pronounced at finite scales. For each feature sub-map at each scale, its local binary pattern histogram is calculated. The local binary pattern histogram generates binary patterns by comparing the gray values ​​of pixels with their neighboring pixels and statistically analyzing the distribution of these patterns in the image, thus constituting the texture description at that scale. The local binary pattern histograms at all scales are concatenated to form a multi-scale texture histogram, which integrates the texture statistics of the target structure feature map from macroscopic to microscopic levels.

[0063] In some embodiments, the chi-square distance between adjacent scale histograms in the multi-scale texture histogram is calculated. The chi-square distance measures the difference between two probability distributions and serves as a measure of texture coherence across scales. In the example scenario, if the texture pattern of a crack remains relatively stable across different scales, the chi-square distance between adjacent scale histograms is small, indicating high texture coherence. If the texture pattern changes drastically, the chi-square distance increases, and coherence decreases. Data comparison shows that real crack regions often have texture patterns that are transmitted across scales, while texture patterns generated by random noise lack coherence across different scales. Simultaneously, the entropy of the gradient magnitude map of each scale feature sub-map is calculated. The entropy of the gradient magnitude map reflects the degree of disorder or uncertainty of the local gradient in the image. Entropy is used as the intensity of texture abrupt changes at each scale. At crack edges, the gradient direction is concentrated, resulting in a lower entropy value; in complex textures or noisy regions, the gradient direction is dispersed, resulting in a higher entropy value. The texture coherence measure and texture abrupt change intensity at each scale are arranged in scale order to form a texture scale sequence. Optionally, the number of layers in the Gaussian pyramid decomposition and the neighborhood radius parameter used in the local binary mode histogram can be preset and adjusted according to the texture complexity of the bridge structure surface.

[0064] In practice, the formation of the initial candidate region set begins with extracting scale information from the texture scale sequence where the texture abrupt change intensity exceeds a preset threshold. The preset threshold is determined based on the statistical distribution of abrupt change intensity in crack and non-crack regions in historical image data. High-abrupt-change regions are located on the corresponding scale feature sub-map of the target structure feature map. High-abrupt-change regions refer to the set of connected pixels whose texture abrupt change intensity values ​​are significantly higher than the surrounding background. The average gradient direction consistency of high-abrupt-change regions across the entire scale of the target structure feature map is calculated. The average gradient direction consistency is measured by calculating the variance or standard deviation of the gradient direction angle of each pixel in the region. The higher the consistency, the more uniform the edge direction in the region. Regions with gradient direction consistency lower than the preset consistency threshold are selected as seed regions. In the bridge crack example, although the real crack edges are uniform in direction, due to their narrow and tortuous characteristics, when calculating their gradient direction at the pixel level, they are affected by surrounding pixels and noise, and their direction consistency is lower than that of some non-crack regions with regular artificial textures.

[0065] Centered on the seed region, region growing is performed on the target structural feature map. The growth rule combines pixel gray-level similarity with local values ​​of texture coherence measurement in the texture scale sequence. Pixel gray-level similarity requires that the difference between the gray-level mean of the newly added pixel and the seed region does not exceed a tolerance. The local value of texture coherence measurement requires that the local regions traversed by the growth path exhibit a low chi-square distance in the texture scale sequence, i.e., high cross-scale coherence. The growth terminates when an image boundary is encountered or the pixel gray-level gradient exceeds a preset growth threshold. All grown regions are considered as preliminary candidate regions, and the coordinate set of all preliminary candidate regions is recorded as the preliminary candidate region set. In some embodiments, the preset growth threshold is adaptively set according to the global gradient distribution of the target structural feature map, for example, taking a certain percentile of the global gradient value. It can be understood that the region growing process utilizes the spatial continuity of cracks and the scale coherence of textures, which can connect and expand scattered, highly abrupt seed points into complete suspected crack regions. To quantify the consistency of the average gradient direction, the following formula is introduced:

[0066]

[0067] in: This indicates that the average gradient direction is consistent. This represents the total number of pixels in the high-change-rate region. Indicates the first The gradient direction angle of each pixel This represents the mean of the gradient direction angles of all pixels within this region. The absolute value is represented by the sine function, which is used to handle periodic problems of angles. The closer the value is to 1, the higher the directional consistency. In practice, the preset consistency threshold is an empirical value between 0 and 1, used to distinguish between non-crack edges with highly consistent orientations and crack candidate edges with relatively dispersed orientations. Data comparison shows that many non-crack regular texture areas have high consistency. The value, and the set of pixels selected as the seed region. The value is lower.

[0068] In one embodiment of the present invention, the method for constructing the region evolution model is as follows: for each preliminary candidate region in the preliminary candidate region set, the position and morphological changes of the region are tracked in multiple consecutive frames of an image monitoring sequence; the Hausdorff distance change rate of the contour point set of the region in consecutive frames and the autocorrelation coefficient of the mean gray value of pixels within the region are calculated; using the area change of the region, the Hausdorff distance change rate of the contour point set, and the autocorrelation coefficient of the mean gray value of pixels as state variables, a Kalman filter is used to model and predict these state variables, and the predicted state sequence output by the Kalman filter is the region evolution model of the preliminary candidate region. The quantitative description of the crack growth trend is output by the crack growth prediction model, which receives the preliminary candidate region set and its region evolution model as input; the model adopts a combination structure of long short-term memory network and fully connected layer, wherein the long short-term memory network is used to encode the time series state in the region evolution model, and the fully connected layer is used to map the encoding result to a quantitative description of the crack growth trend; the quantitative description includes the predicted values ​​of crack length change rate, crack width expansion probability, and crack extension direction angle.

[0069] In practical implementation, the method for constructing the region evolution model is applied to each preliminary candidate region in the initial candidate region set. In the example scenario of crack detection in bridge construction, the image monitoring sequence consists of continuous frame images captured by a fixed camera. The preliminary candidate regions are suspected crack connected regions identified on the target structure feature map. Tracking the position and shape of the preliminary candidate regions in multiple consecutive frames involves calculating the spatial correspondence between adjacent frames. Data comparison shows that real crack regions exhibit slow and continuous morphological changes and positional drifts in consecutive frames, while noise regions that are not cracks exhibit random jumps or instantaneous appearance and disappearance. The Hausdorff distance change rate of the contour point set of the preliminary candidate regions in consecutive frames is calculated. The Hausdorff distance change rate measures the rate of change of the difference in the region's contour shape between two frames. At the same time, the autocorrelation coefficient of the mean pixel grayscale value within the region is calculated. The autocorrelation coefficient reflects the correlation of the overall brightness of the region over time. The area change of the preliminary candidate regions, the Hausdorff distance change rate of the contour point set, and the autocorrelation coefficient of the mean pixel grayscale value are used as state variables to describe the dynamic characteristics of the regions.

[0070] In practical implementation, a Kalman filter is used to model and predict state variables. The Kalman filter optimally estimates the state variables based on the linear system state equation and observation equation. In the example scenario, the state vector composed of state variables includes area change, Hausdorff distance change rate, and autocorrelation coefficient. The prediction step of the Kalman filter estimates the state of the next frame based on a physical model of regional motion. The update step corrects the prediction using newly observed actual measurements of the region. The predicted state sequence output by the Kalman filter is the regional evolution model of the preliminary candidate region. Data comparison shows that for real crack regions, the predicted state sequence of the regional evolution model, such as area change, shows a gradually increasing trend, while the predicted state sequence of noisy regions fluctuates irregularly. In some embodiments, the state transition matrix of the Kalman filter can be assumed to be a uniformly changing model, and the process noise covariance matrix and observation noise covariance matrix are obtained through historical tracking data calibration. In the prediction step of the Kalman filter, the physical model of regional motion is specifically implemented as a uniformly changing model. This model assumes that the changes in the state variables of the preliminary candidate region follow a uniform motion law based on the dynamic characteristics of cracks in the construction environment. State variables include area change, the rate of change of Hausdorff distance of the contour point set, and the autocorrelation coefficient of the mean pixel grayscale value. These variables collectively describe the morphological evolution and grayscale consistency of the region. The uniform change model is implemented through a state transition matrix that assumes the region's movement between consecutive frames is stationary and linear, thus directly deriving the predicted state of the next frame from the state value of the current frame. This physical model closely matches the actual process of crack propagation, as real cracks tend to exhibit a slow and continuous growth trend rather than random fluctuations. In the update step, newly observed actual measurements of the region are used to correct the predicted values, and the state estimate is adjusted using Kalman gain to reduce noise interference. Data comparison shows that the uniform change model can effectively distinguish the stable evolution of crack regions from the irregular changes of noisy regions, thereby enhancing the reliability of the region evolution model.

[0071] In practical implementation, the quantitative description of crack growth trend is output by the crack growth prediction model. This model receives a preliminary set of candidate regions and their regional evolution models as input. In the example scenario, the regional evolution model provides a time-varying state sequence for each preliminary candidate region. The crack growth prediction model internally includes a combination of a Long Short-Term Memory (LSTM) network and fully connected layers. The LSTM network, a variant of a recurrent neural network, is used to process time-series data. It encodes the time-series states in the regional evolution model and captures long-term dependencies in the state sequence through its gating mechanism. The fully connected layers map the LSTM network encoding results to a quantitative description of the crack growth trend. Data comparison shows that the LSTM network encoding can effectively distinguish between stable crack expansion patterns and non-crack random fluctuation patterns. The quantitative description of crack growth trend includes the crack length change rate, the crack width expansion probability, and the predicted crack extension direction angle. The crack length change rate represents the expected percentage change in crack length per unit time, the crack width expansion probability represents the likelihood of crack width expansion in the next time step, and the predicted crack extension direction angle represents an estimate of the future extension direction angle of the crack.

[0072] In some embodiments, the number of hidden layer units in the Long Short-Term Memory network can be set according to the length and complexity of the state sequence of the region evolution model, and the output dimension of the fully connected layer corresponds to the three components of the quantized description. Optionally, the crack growth prediction model uses image sequences with labeled crack growth data for supervised learning during the training phase, and the loss function can be defined as the mean squared error between the predicted quantized description and the true value. It can be understood that the quantified description of crack growth trends transforms the dynamic evolution of the region into specific, measurable physical parameters that directly serve structural integrity assessment. To quantify the rate of change of Hausdorff distance, the following formula is introduced:

[0073]

[0074] in: Indicates time The rate of change of Hausdorff distance, Represents the current frame contour point set With the contour point set of the previous frame The bidirectional Hausdorff distance between them This represents the time interval between adjacent frames. Assuming the time interval is uniform, the formula calculates the change in Hausdorff distance per unit time. In practical implementation, bidirectional Hausdorff distance... Defined as a set to the set Maximum and minimum Euclidean distances and sets to the set The maximum of the maximum and minimum Euclidean distances, in the data comparison of the actual crack area. The values ​​are usually small and change steadily, reflecting the gradual change in crack morphology, while the noise region... The value may change abruptly or fluctuate significantly.

[0075] In practice, the quantitative description of crack growth trend can be used for further analysis. See Table 1 for the differences in the quantitative description of predictions for different types of preliminary candidate regions:

[0076] Table 1: Comparison of Quantitative Descriptions of Crack Growth Trends in Preliminary Candidate Regions

[0077]

[0078] The table presents example data output from the crack growth prediction model. Regions A and C are identified as real crack areas; a positive crack length change rate indicates predicted crack extension, a high crack width expansion probability suggests a high likelihood of width expansion, and a small predicted crack extension direction angle indicates a relatively stable extension direction. Region B is identified as a noise area; its crack length change rate is close to zero, its crack width expansion probability is low, and its predicted crack extension direction angle is large but lacks clear directional significance. This demonstrates the model's ability to distinguish between different types of regions. Optionally, the data in the table comes from records of model output under specific bridge construction scenarios; actual values ​​may vary depending on the specific image sequence and model parameters.

[0079] See Figure 4 This is a multi-dimensional feature radar chart of real crack areas and noise areas in bridge construction crack detection, showcasing the differences between the two types of areas in five key features. The "coverage area" of the radar chart visually reflects the feature differences between the two types of areas: the feature coverage area of ​​real cracks is much larger than that of noise areas. This type of chart is used in the crack detection result judgment stage to help quantitatively distinguish between real cracks and noise interference, supporting the accuracy verification of AI visual recognition models. The chart reflects the "multi-dimensional recognizability" of crack features: real cracks have obvious characteristics in growth trend (length, width), morphology (area), and grayscale consistency, while noise areas have no stable characteristics. This provides a visual basis for the classification logic of the AI ​​model.

[0080] In one embodiment of the present invention, the calculation process of the dynamic structural integrity score is as follows: Two parameters, the crack length change rate and the crack width expansion probability, are extracted from the quantitative description of the crack growth trend; the ratio of the total area of ​​the preliminary candidate region set in the current frame image to the total area of ​​the visible area on the surface of the construction structure is calculated as the crack area proportion; the minimum value of the texture coherence metric is extracted from the texture scale sequence; and the minimum values ​​of the crack length change rate, crack width expansion probability, crack area proportion, and texture coherence metric are input into a preset evaluation function. This evaluation function is a weighted sum of the above four input values, and its weighting coefficients are learned through historical crack-free samples. The output value of this evaluation function is the dynamic structural integrity score. The method for generating real-time detection results involves comparing the calculated dynamic score of structural integrity with multiple preset dynamic thresholds, which are pre-set based on different construction stages and ambient lighting conditions. Based on the comparison results, it is determined whether there are cracks exceeding safety standards in the current image monitoring sequence. If so, the crack growth trend is quantitatively described and the coordinate information of the preliminary candidate region set is combined to perform visual annotation on the original image. The annotation content includes the crack location, growth trend arrow, and risk level. Finally, a real-time detection result report containing these visual annotations and risk level information is generated.

[0081] In practical implementation, the calculation process of the dynamic score of structural integrity extracts the crack length change rate and crack width expansion probability from the quantitative description of crack growth trend. The crack length change rate represents the expected change ratio of crack length per unit time, and the crack width expansion probability represents the possibility of crack width expansion. In the example scenario of crack detection in bridge construction, the value of crack length change rate and crack width expansion probability directly reflects the crack activity level. Data comparison shows that cracks that are expanding usually have a higher crack length change rate and crack width expansion probability, while stable historical cracks or false detection noise areas correspond to lower values.

[0082] The ratio of the total area of ​​the preliminary candidate region set in the current frame image to the total area of ​​the visible area on the surface of the construction structure is calculated as the crack area proportion. The total area of ​​the visible area on the surface of the construction structure is obtained by segmenting the current frame image into structural regions, such as extracting the pixel regions of bridge components using a mask. The crack area proportion reflects the spatial proportion of cracks within the current field of view. The minimum value of the texture coherence metric is extracted from the texture scale sequence. The minimum value of the texture coherence metric represents the weakest link in the texture pattern coherence of the target structure feature map across all analysis scales. A lower minimum value of the texture coherence metric indicates that the structural texture has breaks or discontinuities. The crack length change rate, crack width expansion probability, crack area proportion, and the minimum value of the texture coherence metric are input into a preset evaluation function, which is a weighted sum of the above four input values.

[0083] In some embodiments, the weighted summation expression includes four weighting coefficients, corresponding to the minimum values ​​of crack length change rate, crack width propagation probability, crack area proportion, and texture coherence metric. These weighting coefficients are learned from historical crack-free samples. The learning process uses a dataset containing image sequences of normal construction structure surfaces. An optimization algorithm adjusts the weighting coefficients so that the evaluation function output value for normal samples is below a predetermined safety baseline. The output value of the evaluation function is the dynamic structural integrity score. In the example scenario, the dynamic structural integrity score is a comprehensive numerical value; a higher score indicates a higher risk to structural integrity.

[0084] In practice, the real-time detection results are generated by comparing the calculated dynamic score of structural integrity with multiple preset dynamic thresholds. These dynamic thresholds are pre-set based on the construction stage and ambient lighting conditions. For example, different dynamic thresholds are set for the initial concrete curing stage and the later loading stage, as well as for strong daylight and auxiliary nighttime lighting conditions. Based on the comparison results, it is determined whether there are cracks exceeding safety standards in the current image monitoring sequence. If the dynamic score of structural integrity exceeds the lowest level of dynamic threshold, a crack risk is identified.

[0085] If a crack is determined to exist, the crack location, growth trend, and risk level are marked on the original image by combining a quantitative description of the crack growth trend with the coordinates of the preliminary candidate region set. The crack location is marked by drawing a bounding box or highlighting a region on the image using the coordinates of the preliminary candidate region set. The growth trend is indicated by an arrow symbol indicating the predicted direction of the crack's extension angle. The risk level is classified according to the dynamic threshold level exceeded by the structural integrity dynamic score. A real-time detection result report containing visual annotations and risk level information is generated. Optionally, the real-time detection result report can be output in the form of image overlay annotations, structured text files, or database records.

[0086] In some embodiments, a plurality of preset dynamic thresholds form a hierarchical threshold list, such as [Th1, Th2, Th3], where Th1 < Th2 < Th3. If the dynamic score of structural integrity is lower than Th1, it is regarded as safe; if it is between Th1 and Th2, it is regarded as low risk; if it is between Th2 and Th3, it is regarded as medium risk; if it is higher than Th3, it is regarded as high risk. Different risk levels correspond to different alarm and disposal suggestions. To quantify the calculation of the evaluation function, the formula is introduced:

[0087]

[0088] Where: represents the dynamic score of structural integrity, represents the crack length change rate, represents the crack width expansion probability, represents the proportion of the crack area, represents the minimum value of the texture coherence metric, , , , respectively represent the weighting coefficients corresponding to , , , . The weighting coefficients are obtained through learning from historical crack-free samples and satisfy the relationship . In specific implementation, the crack length change rate and the crack width expansion probability are obtained from the quantitative description of the crack growth trend, the proportion of the crack area is obtained through area calculation, and the minimum value of the texture coherence metric is extracted from the texture scale sequence. Data comparison shows that for real and active crack areas, their , , values are higher while the value is lower, resulting in an increase in the value of the dynamic score of structural integrity ; for non-crack areas or stationary cracks, their values are lower.

[0089] Refer to Figure 5This is a bar chart comparing the spatiotemporal fusion performance of bridge crack detection at different feature scales, primarily demonstrating the impact of feature scale on detection results. Through the comparison of two indicators, it showcases the synergistic effect of "multi-scale feature decomposition + spatiotemporal fusion." This type of chart is used in the target structure feature map generation stage to help determine the optimal feature scale and improve crack detection accuracy. The chart reflects the "effectiveness logic" of multi-scale features: mid-to-high scale features (such as scale 6) are better able to capture the texture and morphological information of cracks, and combined with spatiotemporal fusion, can maximize detection accuracy, providing a high-quality feature foundation for subsequent candidate region localization.

[0090] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic detection of bridge construction cracks based on AI visual recognition, characterized in that, The method includes the following steps: Acquire multiple consecutive images containing the surface of the construction structure to form an image monitoring sequence; Pixel-level temporal and spatial feature separation is performed on the image monitoring sequence to obtain temporal separation features and spatial separation features; The temporal and spatial separation features are input into a pre-defined spatiotemporal fusion network to obtain a target structure feature map that incorporates spatiotemporal context information. Calculate the texture coherence and abrupt change intensity of the target structure feature map at multiple scales to generate a texture scale sequence; Based on the texture scale sequence, the distribution area of ​​potential cracks in the image monitoring sequence is determined, forming a preliminary set of candidate regions; For each preliminary candidate region set, the discrete variation law of pixel intensity along multiple directions within it is analyzed, and a region evolution model is constructed. The preliminary candidate region set and its regional evolution model are input into the crack growth prediction model, and a quantitative description of the crack growth trend is output. By combining a quantitative description of crack growth trends with a texture scale sequence, a dynamic score of structural integrity for the image monitoring sequence is calculated. Based on the dynamic scoring of structural integrity, real-time detection results of bridge construction cracks are generated.

2. The method for dynamic detection of bridge construction cracks based on AI visual recognition according to claim 1, characterized in that, The methods for obtaining the temporal domain separation features and spatial domain separation features are as follows: Extract the grayscale value changes of corresponding pixels between adjacent image frames in the image monitoring sequence, and construct a three-dimensional temporal domain change volume by arranging the grayscale value changes in time sequence. A directional filtering process is applied to the three-dimensional time-domain variable volume, and its energy accumulation in the vertical, horizontal and diagonal directions is calculated. The energy accumulation is used as a time-domain separation feature. For each frame of the image monitoring sequence, its gradient magnitude map and direction map are calculated. Multiple rounds of convolution and pooling operations are performed on the gradient magnitude map and direction map respectively to extract deep spatial descriptors that characterize the local structure of the image. The deep spatial descriptors are used as spatial separation features.

3. The method for dynamic detection of bridge construction cracks based on AI visual recognition according to claim 2, characterized in that, The method for obtaining the target structure feature map is as follows: A spatiotemporal fusion network is pre-constructed, which includes a feature alignment module, a bidirectional attention module, and a feature reconstruction module; Temporal and spatial separation features are input into the feature alignment module of the spatiotemporal fusion network. Based on optical flow estimation, the temporal and spatial separation features are corrected for their spatiotemporal positions to obtain aligned spatiotemporal feature pairs. The aligned spatiotemporal feature pairs are then input into the bidirectional attention module of the spatiotemporal fusion network to calculate the attention weights of the temporal and spatial separation features on the spatial separation features and vice versa. The aligned spatiotemporal feature pairs are then weighted and fused based on these two attention weights to generate fused features. The fused features are then input into the feature reconstruction module of the spatiotemporal fusion network to restore the resolution of the fused features to the original image size through deconvolution and upsampling operations, resulting in the target structure feature map.

4. The method for dynamic detection of bridge construction cracks based on AI visual recognition according to claim 1, characterized in that, The method for generating the texture scale sequence is as follows: Gaussian pyramid decomposition is performed on the target structure feature map to obtain feature sub-maps at multiple different scales. For each feature sub-map at each scale, its local binary pattern histogram is calculated, and the local binary pattern histograms at each scale are concatenated to form a multi-scale texture histogram. The chi-square distance between adjacent scale histograms in the multi-scale texture histogram is calculated, and the chi-square distance is used as a measure of texture coherence between scales. Simultaneously, the entropy of the gradient magnitude map of the feature sub-map at each scale is calculated, and the entropy is used as the texture mutation intensity at each scale. The texture coherence measure and texture mutation intensity at each scale are arranged in scale order to form a texture scale sequence.

5. The method for dynamic detection of bridge construction cracks based on AI visual recognition according to claim 4, characterized in that, The method for forming the preliminary candidate region set is as follows: Extract scale information where the intensity of texture mutation exceeds a preset threshold from the texture scale sequence, locate high mutation regions on the corresponding scale feature sub-map of the target structure feature map, calculate the average gradient direction consistency of the high mutation regions on the full scale of the target structure feature map, and select regions with gradient direction consistency lower than a preset consistency threshold as seed regions. Centered on the seed region, region growth is performed on the target structure feature map. The growth rule combines the local values ​​of pixel gray-level similarity and texture coherence measurement in the texture scale sequence. The growth termination condition is when the image boundary is encountered or the gray-level gradient exceeds the preset growth threshold. All grown regions are taken as preliminary candidate regions, and the coordinate set of all preliminary candidate regions is recorded as the preliminary candidate region set.

6. The method for dynamic detection of bridge construction cracks based on AI visual recognition according to claim 5, characterized in that, The method for constructing the regional evolution model is as follows: For each preliminary candidate region in the preliminary candidate region set, the position and shape of the preliminary candidate region are tracked in multiple consecutive frames of the image monitoring sequence. The Hausdorff distance change rate of the contour point set of the preliminary candidate region in consecutive frames and the autocorrelation coefficient of the mean gray value of the pixels in the region are calculated. The area change of the preliminary candidate region, the Hausdorff distance change rate of the contour point set, and the autocorrelation coefficient of the mean gray value of the pixels are used as state variables. A Kalman filter is used to model and predict the state variables. The predicted state sequence output by the Kalman filter is the region evolution model of the preliminary candidate region.

7. The method for dynamic detection of bridge construction cracks based on AI visual recognition according to claim 6, characterized in that, The method for obtaining the quantitative description of the crack growth trend is as follows: The crack growth prediction model receives a preliminary set of candidate regions and their regional evolution models as input. The crack growth prediction model contains a combination structure of a long short-term memory network and a fully connected layer. The long short-term memory network is used to encode the time series state in the regional evolution model, and the fully connected layer is used to map the encoding results into a quantitative description of the crack growth trend. The quantitative description of the crack growth trend includes the crack length change rate, the crack width expansion probability, and the predicted value of the crack extension direction angle.

8. The method for dynamic detection of bridge construction cracks based on AI visual recognition according to claim 7, characterized in that, The calculation method for the dynamic score of structural integrity is as follows: The crack length change rate and crack width expansion probability are extracted from the quantitative description of crack growth trend. The ratio of the total area of ​​the preliminary candidate region set in the current frame image to the total area of ​​the visible area on the surface of the construction structure is calculated as the crack area proportion. The minimum value of the texture coherence metric is extracted from the texture scale sequence. The minimum values ​​of crack length change rate, crack width expansion probability, crack area proportion, and texture coherence metric are input into a preset evaluation function. The evaluation function is a weighted sum of the above four input values. The weighting coefficients are learned from historical crack-free samples. The output value of the evaluation function is the dynamic score of structural integrity.

9. The method for dynamic detection of bridge construction cracks based on AI visual recognition according to claim 1, characterized in that, The method for generating the real-time detection results is as follows: The structural integrity dynamic score is compared with multiple preset dynamic thresholds, which are preset according to the construction stage and ambient lighting conditions. Based on the comparison results, it is determined whether there are cracks exceeding the safety standard in the current image monitoring sequence. If so, the crack location, growth trend and risk level are marked on the image by combining the quantitative description of the crack growth trend and the coordinates of the preliminary candidate area set, and a real-time detection result report containing visual annotations and risk level information is generated.

10. A dynamic detection system for bridge construction cracks based on AI visual recognition, comprising an image acquisition device, a computing and processing device, and a result output device, characterized in that, The computing processing device is used to execute instructions to implement each step in the dynamic detection method for bridge construction cracks based on AI vision recognition as described in any one of claims 1 to 9.

Images