A detection method and device for airport pavement crack segmentation

Abstract

This invention discloses a detection method and apparatus for airport pavement crack segmentation. The method includes: acquiring airport pavement crack images; training an airport pavement crack segmentation model; generating crack segmentation prediction results based on the trained model; and generating a pavement crack safety status detection report based on the prediction results. The apparatus includes a processor and a memory. This invention improves the model's transfer and generalization ability across different airport pavement scenarios by introducing a novel loss function into the airport pavement crack segmentation model, thereby enhancing the detection accuracy of airport pavement cracks.

Description

Technical Field

[0001] This invention relates to the field of airport pavement inspection, and more particularly to a method and apparatus for detecting cracks in airport pavement. Background Technology

[0002] Airport pavements, as the primary carriers for aircraft ground operations, are crucial infrastructure for ensuring safe airport operations. Cracks are the most common surface defects of pavements and also the initial form of structural damage. Therefore, the International Civil Aviation Organization (ICAO) has made crack detection a mandatory requirement to ensure the healthy operation of airports. Previously, crack detection was done manually, which was susceptible to subjective factors leading to missed or false detections, and was also extremely costly in terms of manpower. Traditional manual inspections are no longer sufficient to meet increasingly demanding requirements, so automated crack detection technology has become a mainstream research direction in recent years. Benefiting from its powerful feature extraction capabilities, deep learning-based methods have been widely applied in crack detection, achieving efficient crack detection and providing data references for pavement structural safety monitoring. Due to long-term high-load operation, pavements are severely contaminated by fuel stains, aircraft wheel marks, etc., making the background environment for cracks extremely complex. The proportion of positive sample pixels is much smaller than that of background pixels, resulting in a highly unbalanced data distribution, which poses challenges to pixel-level fine-grained crack detection.

[0003] Crack detection not only provides a scientific basis for the prevention and repair of early damage, but also reduces the risk of wider road damage caused by crack propagation, ensuring traffic safety and smooth flow. In particular, in recent years, the application of deep learning technology has significantly improved the accuracy and efficiency of crack segmentation.

[0004] However, when deployed in real-world applications, despite their outstanding performance on standardized datasets, these models often fail to achieve satisfactory generalization performance in the complex and ever-changing real-world environments. In practical applications, crack segmentation models often exhibit high false positive and false negative rates due to insufficient generalization ability, making it difficult to cope with complex and varied road surface environments and cracks with intricate topological structures. Limited by the quantity and diversity of training data, overfitting becomes a common problem, leading to a significant drop in segmentation accuracy when faced with unfamiliar new scenarios. This highlights the urgent need to improve their generalization performance through various techniques. Summary of the Invention

[0005] This invention provides a method and apparatus for detecting cracks in airport pavement. By introducing a novel loss function into the airport pavement crack segmentation model, this invention improves the model's transfer and generalization ability across different airport pavement scenarios, thereby enhancing the detection accuracy of cracks in airport pavement. Details are described below:

[0006] A first aspect is a detection method for segmenting cracks in airport pavement, the method comprising:

[0007] Obtain images of cracks in the airport pavement;

[0008] A crack segmentation model for airport pavement is trained, and crack segmentation prediction results are generated for airport pavement crack images based on the trained model.

[0009] Repair the cracked pavement based on the prediction results.

[0010] The airport pavement crack segmentation model is obtained by training a U-Net neural network model with a training dataset consisting of labeled pixel-level tunnel crack images. The U-Net neural network model consists of an encoder and a decoder, and the features of the encoder and decoder are fused through skip connections.

[0011] The training of the airport pavement crack segmentation model is as follows:

[0012] An active fault-tolerant loss calculation method is proposed to be applied to the loss calculation stage during model training, including two parts: neighborhood spatial information modeling and overall Gaussian smoothing;

[0013] By using Markov random fields to model the correlation of pixels in the neighborhood of the label center point, an active fault-tolerant loss is constructed.

[0014] The active fault-tolerant loss takes the crack segmentation prediction result output by the U-Net segmentation model and its corresponding label image as input. The neighborhood spatial information modeling part is responsible for modeling the relative offset probability in the label neighborhood space corresponding to any pixel in the crack segmentation prediction result, and the Gaussian smoothing part is responsible for smoothing the neighborhood spatial offset probability of all pixels.

[0015] Specifically, the active fault-tolerant loss is constructed by modeling the correlation of pixels in the neighborhood of the label center point using Markov random fields:

[0016] Set a neighborhood window W of size ω, where ω = 2n + 1, to constrain the scope of the Markov random field; calculate the relative offset probability based on the correlation degree within the neighborhood within the window; and finally integrate them into a metric score for loss calculation.

[0017] The potential energy function of the Markov random field is expressed as:

[0018]

[0019] Where x represents the potential for offset between the label pixel and the predicted segmentation, V represents a node, E represents an edge, s and t represent the predicted segmentation and label, respectively, and ω is the neighborhood window size; Φ(·) is the univariate potential function of a node, which calculates the offset probability using the segmentation prediction and label pixel output by the segmentation model within the neighborhood window; Ψ(·) is the point-to-point potential function of an edge, used to smooth the spatial correlation between different neighborhood windows. i The effect of offset probability.

[0020] Furthermore, the univariate potential energy function of the node is:

[0021] Φ(x i ,s i ,t |ω| )=-ln p(x i | i ,t |ω| >)

[0022] Where p(x) i | i ,t i >) indicates that when the segmentation prediction is s i The probability of a relative offset occurring in the neighborhood space of the corresponding label is modeled using the following formula:

[0023]

[0024] in, For the XOR operation, ω is the size of the neighborhood window, d(s) i ,t j ) represents s i With t j The distance between them.

[0025] The point-to-point potential energy function of the edge is:

[0026]

[0027] Where μ(·) is the difference between pairs of related points in the offset matrix x, i.e., μ(x i ,x j )=|x i -x j |,s i With s j It is the corresponding value of the associated point pair in the segmentation prediction. The exponential part uses the difference between adjacent pixel labels as a variable to increase the perception ability of local regions. σ is a hyperparameter used to adjust the influence of prediction bias within the neighborhood window on the Markov random field.

[0028] The loss function of the joint Markov random field is: ​​

[0029]

[0030] Where N is the size of the image data, s i t represents the pixel value in the segmentation prediction output by the model. i This is the pixel value in the label. μ i This is the relative offset probability corresponding to that point, which is the result calculated from the active fault tolerance loss.

[0031] A second aspect is a detection device for segmenting cracks in airport pavement, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method described in any of the first aspects.

[0032] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.

[0033] The beneficial effects of the technical solution provided by this invention are:

[0034] 1. This invention proposes an improved loss function that uses Markov random fields to model the active fault tolerance mechanism. By combining pixel similarity and spatial information, it constructs the correlation between the center point and the points in the neighborhood, thereby improving the transfer performance of the pre-trained crack segmentation model in real-world scenarios. This improved loss function—PFT Loss—not only provides a more reasonable supervision signal and alleviates the shortcomings of traditional loss functions in single-pixel sensitivity, but also enhances the model's ability to perceive contextual information.

[0035] 2. In crack detection tasks, the model needs to accurately identify and segment crack regions in images. However, traditional loss functions often only focus on the prediction results of a single pixel, which is quite different from the active fault-tolerance mechanism in the human visual system. PFT Loss uses Markov random fields to model the correlation between the segmentation and the neighborhood space of the labeled pixel, effectively simulating the active fault-tolerance characteristics of the human visual system. This enables the model to more robustly identify cracks when facing complex scenes such as noise, shadows, and changes in lighting, without being disturbed by the error of a single pixel.

[0036] 3. Furthermore, PFT Loss demonstrates high versatility across different datasets and network frameworks, enabling it to work stably under various pavement types and variable environmental conditions. In rigorous experiments targeting different pavement types and complex environmental conditions, the model trained with the improved loss function showed significantly enhanced generalization ability. Not only was the accuracy maintained or improved on multiple standard test sets, but more importantly, the model exhibited more robust detection performance in real road environments.

[0037] 4. This improved generalization ability and robustness greatly enhances the model's value in practical applications. By using PFT Loss, the model can more effectively handle various uncertainties in real-world scenarios, such as different road surface materials (e.g., asphalt, cement), weather conditions (e.g., rainy days, sunny days), and changes in lighting (e.g., daytime, nighttime). This performance improvement makes the model widely applicable and of significant practical importance in actual road monitoring, maintenance, and management. Ultimately, this improvement not only enhances the accuracy and reliability of crack detection but also helps extend the service life of infrastructure and reduce maintenance costs. Attached Figure Description

[0038] Figure 1 This is a schematic diagram illustrating the principle of active fault tolerance loss calculation in this invention.

[0039] Figure 2 This is a predicted segmentation map output by the model in a certain iteration process of an embodiment of the present invention;

[0040] Figure 3 The image shows the effect of this invention on crack segmentation using models trained with different loss functions. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0042] While numerous methods have been proposed to enhance model generalization, such as increasing data diversity, optimizing network architecture, and employing transfer learning, each strategy has its advantages and disadvantages. Increasing data diversity helps models adapt to unseen scenarios, but may still not cover all situations. Adjusting network architecture can sometimes increase model complexity and computational burden, while the effectiveness of transfer learning may be limited by differences between the source and target tasks. Overall, despite these methods' progress in some aspects, they have not yet adequately addressed the model generalization problem.

[0043] Therefore, this invention improves the model's transfer and generalization ability across different airport pavement scenarios by introducing a novel loss function into the airport pavement crack segmentation model. During the analysis, this invention found that proactive fault-tolerance mechanisms are prevalent in human perception. In the data annotation stage, due to complex backgrounds and unavoidable subjective misjudgments, annotators maintain a certain tolerance range for the segmentation of small, irregular targets. When segmenting cracks, to maximize crack segmentation, annotators will classify some background pixels as cracks, thus forming a blurred boundary around the crack body. In the model evaluation stage, considering the relatively small width of the cracks, many previous works generally set a tolerance of 2 pixels for the segmentation results, and this fault-tolerance mechanism during evaluation has been used in recent works.

[0044] Further investigation revealed that the active fault-tolerance mechanism in artificial crack segmentation plays a role highly similar to the loss function used in deep learning-based crack segmentation models. This prompts the present invention to consider two questions: Does overly stringent supervision signal limit the learning of fine crack segmentation models? Can the active fault-tolerance mechanism from human vision be explicitly injected into the model learning process to enhance the model's generalization ability?

[0045] Further in-depth analysis of the loss function revealed its significant impact on generalization performance during model training. In crack segmentation, cross-entropy, Dice Loss, or Focal Loss are commonly used as loss functions, and previous work has demonstrated their effectiveness. However, cracks are typically irregular, small targets with randomly extending topological structures and very sparse internal features. Furthermore, analysis of the loss function calculation process in this invention revealed that this method, which considers only a single pixel for loss calculation, is suboptimal for small crack segmentation. Traditional single-point-based loss functions naturally filter out the active tolerance interval, ignoring the contextual information of the label pixel in its neighborhood space. This leads to the crack segmentation model easily overfitting the label distribution, thus affecting the model's generalization ability in real-world scenarios.

[0046] Example 1

[0047] A detection method for segmenting cracks in airport pavement, see [link to relevant documentation]. Figure 1 The method includes the following steps:

[0048] 101: Obtain images of cracks in the airport pavement;

[0049] Images of cracks in the airport pavement are acquired using RGB cameras mounted on a high-speed inspection robot. The image data needs to be as clear as possible for subsequent crack segmentation processing.

[0050] 102: Using a trained airport pavement crack segmentation model, generate prediction results for crack segmentation of airport pavement crack images.

[0051] The airport pavement crack segmentation model is obtained by training a U-Net neural network model using a training dataset consisting of labeled pixel-level tunnel crack images. The U-Net neural network model consists of an encoder (downsampling path) and a decoder (upsampling path), and the features of the encoder and decoder are fused through skip connections. The proposed active fault-tolerant loss calculation method is applied to the loss calculation stage during model training, including two parts: neighborhood spatial information modeling and overall Gaussian smoothing.

[0052] This invention utilizes Markov random fields to model the correlation between pixels in the neighborhood of the label center point, constructing an active fault-tolerant loss, such as... Figure 1 As shown.

[0053] In this embodiment, the training method for the tunnel crack image crack segmentation model is as follows:

[0054] First, obtain the training dataset, which includes several sets of training data. Each set of training data includes an image of an airport pavement crack and a pixel-level label for that image.

[0055] Finally, during one iteration of training the U-Net segmentation model using the constructed dataset, the proposed active fault-tolerant loss calculation method uses the crack segmentation prediction results output by the U-Net segmentation model and their corresponding label images as inputs in the loss calculation stage. Specifically, the neighborhood spatial information modeling part is responsible for modeling the relative offset probability within the label neighborhood space corresponding to any pixel in the crack segmentation prediction result, while the Gaussian smoothing part is responsible for smoothing the neighborhood spatial offset probabilities of all pixels. The calculated value of the active fault-tolerant loss is then integrated into the overall loss function as a supervision signal for the U-Net segmentation model, used for backward gradient updates. This iteration ends, and the next iteration begins, until the termination condition is met.

[0056] For example, the training of a crack segmentation model for airport pavement crack images will be explained in more detail below:

[0057] Step 1: Establish an image dataset of airport pavement cracks, which includes the following steps:

[0058] Step 1.1: Construct an airport pavement crack image dataset using airport pavement crack images captured by the airport pavement high-speed inspection robot. The dataset contains 2,000 airport pavement crack images with a fixed resolution of 1800×900.

[0059] Step 1.2: Use the Labelme annotation tool to perform pixel-level annotation on the airport pavement crack image. Pixel-level annotation here refers to labeling the category of each pixel in the acquired airport pavement crack image. The labeled categories are two: crack pixels and non-crack pixels.

[0060] Step 1.3: In order to improve the robustness of the airport pavement crack image crack segmentation model and enable it to adapt to airport pavement crack detection in complex environments, data augmentation operations are performed on the labeled airport pavement crack image dataset, namely, randomly adjusting image brightness, contrast, image rotation, scale transformation, and adding noise.

[0061] Step 1.4: Randomly divide the airport pavement crack images into the training set and the validation set at a ratio of 8:2 to complete the construction of the airport pavement crack image dataset.

[0062] Step 2: Construct an improved U-Net neural network and train it using the aforementioned airport pavement crack image dataset. During training, the proposed active fault-tolerant loss function is used for loss calculation to supervise the gradient optimization direction of the model.

[0063] Specifically, in this embodiment of the invention, a neighborhood window W of size ω, where ω = 2n + 1, is set to constrain the scope of the Markov random field; then, the relative offset probability is calculated within the window based on the correlation degree within the neighborhood; finally, it is integrated into a metric score for loss calculation.

[0064] First, briefly describe the general process of Markov random field modeling. The label T is a binary pixel matrix containing n elements, where 0 and 1 represent the background and cracks, respectively. The segmentation prediction S output by the model is a probability matrix. X is the offset matrix between any pair of points in T and S. For a given S and T, calculate the offset of any pixel s in S. i Relative to label pixel t i The probability P(x) of an offset occurring within the neighborhood window W i |s i ,t i Then the offset probability will be integrated into the loss calculation as a reference for active fault tolerance.

[0065] It should be noted that when considering the interactions between pixels, image data can be naturally constructed as an undirected graph, which, when a neighborhood window is added, exhibits Markov properties. The potential function of a Markov random field is expressed as:

[0066]

[0067] Where x represents the potential for offset between the label pixel and the predicted segmentation, V represents a node, E represents an edge, s and t represent the predicted segmentation and label, respectively, and ω is the neighborhood window size. Φ(·) is the univariate potential function of a node, which directly calculates the offset probability using the segmentation prediction and label pixel output from the segmentation model. Ψ(·) is the point-to-point potential function of an edge, used to measure the spatial correlation between nodes within the label pixel's neighborhood window relative to s. i The impact of offset probability. The optimization objective of a Markov random field model can be defined as minimizing the segmentation prediction offset probability.

[0068] 2.1 Neighborhood Window Settings. When applying Markov Random Fields (MRFs) in image segmentation tasks, the entire image is typically constructed as an undirected graph. Considering global pixel relationships would significantly increase computational complexity. Therefore, this embodiment of the invention uses a neighborhood window to constrain the effective range of the MRF. The specific settings for the neighborhood window W are as follows:

[0069]

[0070] Where i and j represent the offsets of each point within the window relative to the center point in the row and column directions, respectively. Using relative position encoding instead of sequential position makes it easier to model the distances between points. Furthermore, relative position encoding preserves rotation and translation invariance between pixels in the neighborhood. For example, when calculating the offset of a pixel x at the top-left corner of the neighborhood... a,b With center pixel x c,d When determining the distance between two points, if sequential position encoding is used, then the distance d between the two points is equal to (ac). 2 +(bd) 2 If relative position encoding is used, then d = |i + j|. Clearly, relative position encoding preserves spatial distance measurement functionality while saving computational costs.

[0071] 2.2 Node Potential Energy Calculation. As an important component of the energy function, the univariate potential energy function is responsible for calculating the potential energy of any pixel s in the predicted segmentation of the model output. i t, a pixel within the label's neighborhood window |ω| The relationship between the two points is modeled, taking into account both spatial location and semantic features, to form the offset matrix corresponding to the point x. i Assignment. Defined as follows:

[0072] Φ(x i ,s i ,t |ω| )=-ln p(x i | i ,t |ω| >) (3)

[0073] Where p(x) i ​| i ,t i >) indicates that when the segmentation prediction is s i When, the probability of a relative offset occurring in the neighborhood space of the corresponding label. To more intuitively compare with the label, the embodiments of the present invention first use s i Binarization is performed with a threshold set to 0.8. When the probability value is greater than 0.8, the segmentation result of that point can be considered correct. The following formula is used for modeling:

[0074]

[0075] in, For the XOR operation, ω is the size of the neighborhood window. In the first term of equation (4), only the label node t is used to control the effect of the Markov random field. i With predicted segmentation s i When they are not identical, the result of the XOR operation is 1, and only then is it necessary to examine the label distribution within the neighborhood window. The numerator of the first term in equation (4) is used to integrate the spatial feature semantics within the neighborhood window. When the model misclassifies a pixel, the label neighborhood window t |ω| Intra- and predicted segmentation s i Pixels with the same value t j All of them could become s i The relative offset point. Figure 1 Taking P1 and P2 as examples, four points in P1's neighborhood window have the same label value as P1, while no points in P2's neighborhood window have the same value as P2. Therefore, P1 is more likely to be the point where the prediction shift occurs. The denominator of the first term in equation (4) is used to fuse the spatial distance information within the neighborhood window, d(s i ,t j ) represents s i With t j The distance between them can be calculated directly using the coordinates in the neighborhood window, and the offset probability is inversely proportional to the distance.

[0076] 2.3 Edge Potential Energy Calculation. When the number of offset points in the neighborhood space is large, it indicates a significant difference between the labeled area and the actual crack in that local region, thus requiring a greater penalty to promote model correction. To achieve this goal, the edge potential energy function must be able to effectively penalize such local inconsistencies, thereby improving the model's attention to ambiguous regions. Based on prior knowledge, specifically, this embodiment of the invention can define a binary potential energy function as follows:

[0077]

[0078] Where μ(·) is the difference between pairs of related points in the offset matrix x, i.e., μ(x i ,x j )=|x​i -x j |。 s i With s j It is the corresponding value of the associated point pair in the segmentation prediction. The exponential part uses the difference between adjacent pixel labels as a variable to increase the perception ability of local regions. σ is a hyperparameter used to adjust the influence of prediction bias within the neighborhood window on the Markov random field.

[0079] 2.4 Loss Function. The potential function in a Markov random field (MRF) is a posterior probability estimate of the predicted segment offset within a neighborhood window, providing local active fault tolerance for traditional loss calculations. After calculating the MRF potential function, it is incorporated into the final loss function of the model training. This embodiment of the invention improves upon Dice Loss, with the specific expression as follows:

[0080]

[0081] Where N is the size of the image data, s i t represents the pixel value in the segmentation prediction output by the model. i This is the pixel value in the label. μ i This is the relative offset probability corresponding to that point, which is the result calculated from the active fault tolerance loss.

[0082] In this embodiment of the invention, the predicted offset probability of segmented pixels in the neighborhood space is injected into the numerator of the original Dice Loss expression to smooth out the impact of such strict loss calculation. When the offset probability is large, the cost of incorrect prediction will be reduced, and the model will actively tolerate the labeling error.

[0083] Example 2

[0084] The feasibility of the scheme in Example 1 is verified below with specific experimental data, as detailed in the following description:

[0085] Table 1 shows the comparative experiments with commonly used loss functions. Four evaluation metrics were selected: precision (PR), recall (RE), F1 score, and mean intersection over union (mIoU) to comprehensively and deeply evaluate the model's performance.

[0086] Table 1

[0087]

[0088] The experimental results show that, compared with other loss functions, the U-Net model trained using PFT Loss improves in all three evaluation metrics: Recall, F1 score, and mIoU. Although the Precision metric declined, its performance difference with the Recall metric was smaller, which to some extent proves that the model has a more balanced performance.

[0089] This invention improves the model's transfer and generalization ability across different airport pavement scenarios by introducing a new loss function into the airport pavement crack segmentation model, thereby enhancing the detection accuracy of airport pavement cracks.

[0090] Example 3

[0091] A detection device for segmenting cracks in airport pavement, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the following method steps in Embodiment 1:

[0092] Obtain images of cracks in the airport pavement;

[0093] A crack segmentation model for airport pavement is trained, and crack segmentation prediction results are generated for airport pavement crack images based on the trained model.

[0094] Repair the cracked pavement based on the prediction results.

[0095] The airport pavement crack segmentation model is obtained by training a U-Net neural network model with a training dataset consisting of labeled pixel-level tunnel crack images. The U-Net neural network model consists of an encoder and a decoder, and the features of the encoder and decoder are fused through skip connections.

[0096] The training of the airport pavement crack segmentation model is as follows:

[0097] An active fault-tolerant loss calculation method is proposed to be applied to the loss calculation stage during model training, including two parts: neighborhood spatial information modeling and overall Gaussian smoothing;

[0098] By using Markov random fields to model the correlation of pixels in the neighborhood of the label center point, an active fault-tolerant loss is constructed.

[0099] The active fault-tolerant loss takes the crack segmentation prediction result output by the U-Net segmentation model and its corresponding label image as input. The neighborhood spatial information modeling part is responsible for modeling the relative offset probability in the label neighborhood space corresponding to any pixel in the crack segmentation prediction result, and the Gaussian smoothing part is responsible for smoothing the neighborhood spatial offset probability of all pixels.

[0100] Specifically, Markov random fields are used to model the correlation between pixels in the neighborhood of the label center point, and an active fault-tolerant loss is constructed as follows:

[0101] Set a neighborhood window W of size ω, where ω = 2n + 1, to constrain the scope of the Markov random field; calculate the relative offset probability based on the correlation degree within the neighborhood within the window; and finally integrate them into a metric score for loss calculation.

[0102] The potential function of the Markov random field is expressed as:

[0103]

[0104] Where x represents the potential for offset between the label pixel and the predicted segmentation, V represents a node, E represents an edge, s and t represent the predicted segmentation and label, respectively, and ω is the neighborhood window size; Φ(·) is the univariate potential function of a node, which calculates the offset probability using the segmentation prediction and label pixel output by the segmentation model within the neighborhood window; Ψ(·) is the point-to-point potential function of an edge, used to smooth the spatial correlation between different neighborhood windows. i The effect of offset probability.

[0105] Furthermore, the univariate potential energy function of the node is:

[0106] Φ(x i ,s i ,t |ω| )=-ln p(x i | i ,t |ω| .)

[0107] Where p(x) i | i ,t i >) indicates that when the segmentation prediction is s i The probability of a relative offset occurring in the neighborhood space of the corresponding label is modeled using the following formula:

[0108]

[0109] in, For the XOR operation, ω is the size of the neighborhood window, d(s) i ,t j ) represents s i With t j The distance between them.

[0110] The potential energy function for each point pair on the edge is:

[0111]

[0112] ​​Where μ(·) is the difference between pairs of related points in the offset matrix x, i.e., μ(x i ,x j )=|x i -x j |,s i With s j It is the corresponding value of the associated point pair in the segmentation prediction. The exponential part uses the difference between adjacent pixel labels as a variable to increase the perception ability of local regions. σ is a hyperparameter used to adjust the influence of prediction bias within the neighborhood window on the Markov random field.

[0113] The loss function for the joint Markov random field is:

[0114]

[0115] Where N is the size of the image data, s i t represents the pixel value in the segmentation prediction output by the model. i This is the pixel value in the label. μ i This is the relative offset probability corresponding to that point, which is the result calculated from the active fault tolerance loss.

[0116] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0117] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0118] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0119] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0120] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0121] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0122] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0123] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0124] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0125] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A detection method for segmenting cracks in airport pavement, characterized in that, The method includes: Acquire images of cracks in the airport pavement; A crack segmentation model for airport pavement is trained, and crack segmentation prediction results are generated for airport pavement crack images based on the trained model. A safety status inspection report for cracked pavement is generated based on the prediction results; The training of the airport pavement crack segmentation model is as follows: An active fault-tolerant loss calculation method is proposed to be applied to the loss calculation stage during model training, including two parts: neighborhood spatial information modeling and overall Gaussian smoothing; By using Markov random fields to model the correlation of pixels in the neighborhood of the label center point, an active fault-tolerant loss is constructed. The active fault-tolerant loss takes the crack segmentation prediction result output by the U-Net segmentation model and its corresponding label image as input. The neighborhood spatial information modeling part is responsible for modeling the relative offset probability in the label neighborhood space corresponding to any pixel in the crack segmentation prediction result, and the Gaussian smoothing part is responsible for smoothing the neighborhood spatial offset probability of all pixels. Specifically, the active fault-tolerant loss is constructed by modeling the correlation of pixels in the neighborhood of the label center point using Markov random fields: setting a value of... The neighborhood window W, The scope of the Markov random field is constrained; the relative offset probability based on the correlation degree within the neighborhood is calculated within the window; and finally, it is integrated into a metric score for loss calculation. In this context, the potential function in the Markov random field is a posterior probability estimate of the predicted segment offset within the neighborhood window. After calculating the Markov random field potential function, it is incorporated into the final loss function of the model training, as shown in the following expression: ； Where N is the size of the image data. These are the pixel values ​​in the segmentation prediction output by the model. These are the pixel values ​​in the label. This is the relative offset probability corresponding to that point, which is the result calculated from the active fault tolerance loss. The potential function of the Markov random field is expressed as: ； Where x represents the potential offset between the label pixel and the predicted segmentation, V represents a node, E represents an edge, and s and t represent the predicted segmentation and the label, respectively. The size of the neighborhood window; The unary potential function of the node is used to calculate the offset probability within the neighborhood window using the segmentation prediction output by the segmentation model and the label pixels. The point-to-point potential function for the edge is used to smooth the spatial correlation between different neighborhood windows. The effect of offset probability.

2. The detection method for segmenting airport pavement cracks according to claim 1, characterized in that, The airport pavement crack segmentation model is obtained by training a U-Net neural network model with a training dataset consisting of labeled pixel-level tunnel crack images. The U-Net neural network model consists of an encoder and a decoder, and the features of the encoder and decoder are fused through skip connections.

3. The detection method for segmenting airport pavement cracks according to claim 2, characterized in that, The univariate potential energy function of the node is: ； in, Indicates when the segmentation prediction is The probability of a relative offset occurring in the neighborhood space of the corresponding label is modeled using the following formula: ； in, For XOR operation, The size of the neighborhood window, express and The distance between them.

4. The detection method for segmenting airport pavement cracks according to claim 3, characterized in that, The point-to-point potential energy function of the edge is: ； in, This represents the difference between pairs of related points in the offset matrix x, i.e. , and It represents the corresponding value of the associated point pair in the segmentation prediction. The exponential part uses the difference between adjacent pixel labels as a variable to enhance the perception of local regions. It is a hyperparameter used to adjust the effect of prediction bias within the neighborhood window on the Markov random field.

5. A detection device for segmenting cracks in airport pavement, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1-4.

Images