Airport runway crack monitoring method based on multi-scale global feature fusion network

Abstract

This application provides a method for monitoring airport runway cracks based on a multi-scale global feature fusion network, applied in the field of image processing technology. It includes: acquiring an image of the airport runway surface; inputting the airport runway surface image into a multi-scale global feature fusion network to obtain a crack region segmentation result image; the multi-scale global feature fusion network includes an encoder network and a decoder network. The encoder network is used to extract features from the airport runway surface image through DOPE and PPA modules to generate a multi-granularity cascaded fused feature map; the decoder network is used to decode and optimize the feature representation extracted by the encoder network and perform global semantic fusion through an IDAF module to obtain a crack region segmentation result image corresponding to the airport runway surface image; and the quantitative index of the crack is calculated based on the crack region segmentation result image to obtain the monitoring result of the airport runway crack. This method reduces the computational complexity of the model while improving the semantic segmentation accuracy and robustness of the cracks.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method for monitoring airport runway cracks based on a multi-scale global feature fusion network. Background Technology

[0002] As a core infrastructure for aviation development, the structural integrity of airport runways directly affects the safety of aircraft takeoffs and landings. Runway cracks are a typical form of pavement structural deterioration. If not detected and addressed in a timely manner, runway cracks will continue to expand under the influence of aircraft loads, ambient temperatures, and other factors, easily inducing serious problems such as potholes and subsidence, significantly reducing the runway's service life and creating safety hazards. Therefore, achieving rapid, accurate, and automated monitoring of airport runway cracks is a core requirement for airport operation and maintenance management.

[0003] Traditional methods for detecting cracks on airport runways rely primarily on manual inspections, which suffer from low efficiency, high subjectivity, and low recognition rates for minute cracks. Furthermore, these methods are limited by environmental factors such as inclement weather and nighttime operations, making them unsuitable for routine monitoring. In addition, traditional machine vision detection methods mainly depend on manually designed shallow features such as textures and edges. When faced with the complex and interfering environment of airport runway surfaces, the algorithm's feature representation capabilities are limited, resulting in insufficient detection accuracy and robustness, and a high susceptibility to false positives and false negatives.

[0004] Therefore, how to extract high-precision features of runway cracks and improve crack detection accuracy and robustness has become an urgent problem to be solved. Summary of the Invention

[0005] This application provides a method for monitoring airport runway cracks based on a multi-scale global feature fusion network. By constructing a multi-scale global feature fusion network, the method utilizes the Deep Overlapping Block Embedding (DOPE) module to compensate for spatial overlap features, thus solving the problem of spatial continuity loss caused by block processing. The Progressive Patch Aggregation (PPA) module is used to achieve cascaded aggregation of multi-granularity features, enhancing the ability to extract crack features at different scales. Furthermore, the Information Decoding and Fusion IDAF module is used to decode and optimize features from different scale coding layers and perform global semantic fusion, effectively reducing the feature semantic hierarchy differences between the coding and decoding layers. This reduces the computational complexity of the model while improving the semantic segmentation accuracy and robustness of airport runway cracks, achieving high-precision quantitative assessment and monitoring of cracks.

[0006] This application provides a method for addressing airport runway cracks based on a multi-scale global feature fusion network, including: Acquire images of the airport runway surface; The airport runway surface image is input into a multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image. The multi-scale global feature fusion network includes an encoder network and a decoder network. The encoder network extracts features from the airport runway surface image using a Deep Overlapping Block Embedding (DOPE) module and a Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map. The decoder network decodes and optimizes the feature representation extracted from the encoder network and performs global semantic fusion using an Information Decoding and Fusion IDAF module to reduce the semantic information difference between the encoder network and the decoder network, thereby obtaining the crack region segmentation result map corresponding to the airport runway surface image. The quantitative index of the crack is calculated based on the crack area segmentation result map to obtain the monitoring results of the airport runway crack.

[0007] According to an embodiment of this application, a method for identifying airport runway cracks based on a multi-scale global feature fusion network is provided. The step of inputting the airport runway surface image into the multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image includes: preprocessing the airport runway surface image to obtain a corresponding preprocessed image to be detected; inputting the preprocessed image to be detected into the encoder network to obtain a multi-granularity cascaded fusion feature map; inputting the multi-granularity cascaded fusion feature map into the decoder network to obtain a segmentation probability map of the airport runway crack; and binarizing the segmentation probability map of the airport runway crack to obtain a crack region segmentation result map corresponding to the airport runway surface image.

[0008] According to an embodiment of this application, a method for detecting runway cracks based on a multi-scale global feature fusion network is provided. The step of inputting the preprocessed image to be detected into the encoder network to obtain a multi-granularity cascaded fusion feature map specifically includes: embedding multi-scale features into the preprocessed image to be detected through a Deep Overlapping Block Embedding (DOPE) module to obtain an initial feature map corresponding to the preprocessed image to be detected; and performing cascaded feature aggregation on the initial feature map through a Progressive Patch Aggregation (PPA) module to obtain a multi-granularity cascaded fusion feature map.

[0009] According to an embodiment of this application, a method for detecting runway cracks based on a multi-scale global feature fusion network is provided. The step of embedding multi-scale features into the preprocessed image to be detected using a Deep Overlapped Block Embedding (DOPE) module to obtain an initial feature map corresponding to the preprocessed image to be detected specifically includes: dividing the preprocessed image to be detected into a grid using a sliding window, setting the overlap ratio of the sliding window to obtain multiple overlapping blocks corresponding to the preprocessed image to be detected; and inputting the multiple overlapping blocks into the Deep Overlapped Block Embedding (DOPE) module, wherein the Deep Overlapped Block Embedding (DOPE) module includes L convolutional layers, where L is... Integers greater than 1; In the first convolutional layer of the L-layer convolutional layer, the original pixel information of each overlapping block is converted into a high-dimensional feature vector through linear mapping, and then nonlinearly normalized through an activation function to obtain preliminary embedding features; In each convolutional layer other than the first convolutional layer in the L-layer convolutional layer, based on the embedding features output by the adjacent previous convolutional layer, deep feature extraction and nonlinear activation are performed on the decoupling information of the multiple overlapping blocks in the current convolutional layer to obtain multi-level enhanced embedding features; The multi-level enhanced embedding features output by the last convolutional layer in the L-layer convolutional layer are spatially reduced to generate an initial feature map representing the spatial overlap relationship.

[0010] According to an embodiment of this application, a method for detecting airport runway cracks based on a multi-scale global feature fusion network is provided. The step of performing cascaded feature aggregation on the initial feature map using the Progressive Patch Aggregation (PPA) module to obtain a multi-granularity cascaded fused feature map specifically includes: the encoder network comprising N feature extraction layers, where N is an integer greater than 1; for the first feature extraction layer of the encoder network, multi-granularity feature extraction is performed on the initial feature map using the Progressive Patch Aggregation (PPA) module to obtain a first cascaded fused feature map; in each of the N feature extraction layers other than the first feature extraction layer, based on the cascaded fused feature map output by the adjacent previous feature extraction layer, cross-scale feature aggregation is performed on the cascaded fused feature map of the current layer using the Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map corresponding to the feature extraction layer.

[0011] According to an embodiment of this application, a method for detecting airport runway cracks based on a multi-scale global feature fusion network is provided. The step of extracting multi-granularity features from the initial feature map using the Progressive Patch Aggregation (PPA) module to obtain a first cascaded fused feature map specifically includes: for multiple parallel convolutional branches in the PPA module, performing feature mapping on the input initial feature map using grouped convolutional kernels with different receptive field sizes to obtain granular features corresponding to each parallel convolutional branch; concatenating the granular features output by each parallel convolutional branch according to the channel dimension to obtain initial fused features; and performing batch normalization and nonlinear activation on the initial fused features to obtain the first cascaded fused feature map.

[0012] According to an embodiment of this application, a method for detecting airport runway cracks based on a multi-scale global feature fusion network is provided. The information decoding and fusion IDAF module includes a feature reconstruction unit and a semantic fusion unit. The step of inputting the multi-granularity cascaded fusion feature map into the decoder network to obtain a segmentation probability map of the airport runway crack specifically includes: for each decoding layer in the decoder network, using the feature reconstruction unit of the information decoding and fusion IDAF module, spatially reconstructing the multi-granularity cascaded fusion feature map or the fusion feature corresponding to the previous decoding layer to obtain a reconstructed fusion feature map; dimensionally aligning the reconstructed fusion feature map, the output features of the feature extraction layer of the encoder network, and the output features of intermediate nodes to obtain an aligned multi-scale feature stream; The feature extraction layer of the encoder network is at the same level as the current decoding layer. The intermediate nodes are convolutional nodes in the nested skip connection path, excluding the starting node of the encoder network. The multi-scale feature stream is input to the semantic fusion unit of the information decoding and fusion IDAF module. Adaptive weight allocation and background noise suppression are performed using channel attention and spatial attention mechanisms to obtain the fusion features corresponding to each decoding layer. The fusion features of each decoding layer are progressively converged to the final decoding layer through nested skip connections to generate deep fusion features. The deep fusion features are used to characterize the pixel-level edge precision and high-level semantic logic of the airport runway surface image. The deep fusion image is subjected to convolutional mapping and nonlinear mapping processing to obtain the segmentation probability map of the airport runway crack.

[0013] This application embodiment also provides a method for training a multi-scale global feature fusion network, the method being used to train a multi-scale global feature fusion network as described in any of the above claims, the method comprising: Acquire sample images of the airport runway surface, and perform pixel-level annotation on the crack areas in the airport runway surface sample images to obtain annotated preprocessed training samples. The preprocessed training samples are input into the initialized multi-scale global feature fusion network for forward propagation to generate a sample segmentation probability map. A loss function is constructed based on the semantic deviation between the sample segmentation probability map and the corresponding pixel-level annotation. The parameters of the multi-scale global feature fusion network are adjusted using the backpropagation algorithm until the loss function converges, thus obtaining the trained multi-scale global feature fusion network.

[0014] This application embodiment also provides a multi-scale global feature fusion network training method. The step of acquiring airport runway surface sample images and performing pixel-level annotation on crack regions in the airport runway surface sample images to obtain annotated preprocessed training samples specifically includes: acquiring surface images covering different lighting conditions, weather environments, and runway areas during airport runway inspection using image acquisition equipment to obtain an original image dataset; classifying and annotating crack regions in the original image dataset using an annotation tool to generate pixel-level crack semantic segmentation label files corresponding one-to-one with the original image dataset; the crack categories include at least transverse cracks, longitudinal cracks, and mesh cracks; dividing the original image dataset and the corresponding pixel-level crack semantic segmentation label files into a training set, a validation set, and a test set according to a preset ratio; preprocessing the training set and using data augmentation strategies to expand the number of samples to obtain the annotated preprocessed training samples.

[0015] This application also provides an airport runway crack system based on a multi-scale global feature fusion network, including: Image acquisition module, used to acquire images of the airport runway surface; A feature extraction module is used to input the airport runway surface image into a multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image. The multi-scale global feature fusion network includes an encoder network and a decoder network. The encoder network extracts features from the airport runway surface image using a Deep Overlapping Block Embedding (DOPE) module and a Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map. The decoder network decodes and optimizes the feature representation extracted from the encoder network and performs global semantic fusion using an Information Decoding and Fusion IDAF module to reduce the semantic information difference between the encoder network and the decoder network, thereby obtaining the crack region segmentation result map corresponding to the airport runway surface image. The crack monitoring module is used to calculate the quantitative index of the crack based on the crack area segmentation result map, and obtain the monitoring result of the airport runway crack.

[0016] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the airport runway crack method based on a multi-scale global feature fusion network as described above.

[0017] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the airport runway crack method based on a multi-scale global feature fusion network as described above.

[0018] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the airport runway crack method based on a multi-scale global feature fusion network as described above.

[0019] The airport runway crack monitoring method based on a multi-scale global feature fusion network provided in this application involves acquiring an airport runway surface image; inputting the airport runway surface image into a multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image; wherein, the multi-scale global feature fusion network includes an encoder network and a decoder network, the encoder network is used to extract features from the airport runway surface image through a Deep Overlapping Block Embedding (DOPE) module and a Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map; the decoder network is used to decode and optimize the feature representation extracted from the encoder network and perform global semantic fusion through an Information Decoding and Fusion IDAF module to reduce the semantic information difference between the encoder network and the decoder network, thereby obtaining a crack region segmentation result map corresponding to the airport runway surface image; and calculating the crack quantification index based on the crack region segmentation result map to obtain the monitoring result of the airport runway crack. By constructing a multi-scale global feature fusion network and using the Deep Overlapping Block Embedding (DOPE) module to compensate for spatial overlap features, the problem of spatial continuity loss caused by block processing is solved. The Progressive Patch Aggregation (PPA) module is used to achieve cascaded aggregation of multi-granularity features, which enhances the ability to extract crack features at different scales. The Information Decoding and Fusion (IDAF) module is used to decode and optimize features from different scale coding layers and perform global semantic fusion, which effectively reduces the feature semantic hierarchy difference between the coding layer and the decoding layer. While reducing the computational complexity of the model, it improves the semantic segmentation accuracy and robustness of airport runway cracks, and realizes high-precision quantitative assessment and monitoring of cracks. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the airport runway crack method based on a multi-scale global feature fusion network provided in this application embodiment. Figure 2 This is a schematic diagram of the structure of the multi-scale global feature fusion network provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the DOPE module embedded in the deep overlap block provided in the embodiment of this application; Figure 4 This is a schematic diagram of the progressive patch aggregation PPA module provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the information decoding and fusion IDAF module provided in the embodiments of this application; Figure 6 This is a comparative schematic diagram of the airport runway crack area segmentation results provided in the embodiments of this application; Figure 7 This is a flowchart illustrating the multi-scale global feature fusion network training method provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of the airport runway crack system based on a multi-scale global feature fusion network provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] To better understand the embodiments of this application, the prior art will first be described in detail: With the widespread application of deep learning in semantic segmentation, U-Net (Unet), based on a symmetric encoder-decoder structure and skip connection feature fusion, and its improved nested U-Net (Unet++) network, have become the mainstream models for road crack segmentation. However, for the specific high-standard scenario of airport runways, existing Unet++-based detection models still have the following limitations: Disadvantage 1: Insufficient multi-scale feature capture capability: The encoder stage often uses a fixed convolution kernel of a single scale, which leads to a limited receptive field and makes it difficult to balance the local texture information of small cracks with the global spatial features of long-distance cracks at the same time, which can easily lead to the missed detection of small cracks.

[0024] Disadvantage 2: Semantic gap exists in cross-layer feature fusion: Traditional direct skip connections do not consider the semantic hierarchy differences between the encoding and decoding layers, and lack effective feature alignment and filtering mechanisms, resulting in low feature fusion efficiency.

[0025] Disadvantage 3: Conflict between computational redundancy and real-time performance: High parameter redundancy and large model computational complexity during feature extraction make it difficult to meet the engineering requirements of large-scale, real-time detection of airport runways.

[0026] Disadvantage 4: Lack of robustness in complex environments: In airport runway environments, the model is not robust enough to interference factors such as changes in lighting, road stains, and aviation markings, resulting in distorted segmentation results.

[0027] In summary, although existing studies have made local optimizations to the network by introducing attention mechanisms or multi-scale convolutions, they are mostly limited to single-module improvements and have failed to achieve global collaboration of multi-scale features. They still cannot effectively solve the core pain point of small target crack detection on airport runways, and it is difficult to balance the detection accuracy and computational efficiency of the model.

[0028] To address the aforementioned technical issues, this application provides a method for monitoring airport runway cracks based on a multi-scale global feature fusion network. By constructing a multi-scale global feature fusion network and utilizing the Deep Overlapping Block Embedding (DOPE) module to compensate for spatial overlap features, the method solves the problem of spatial continuity loss caused by block processing. The Progressive Patch Aggregation (PPA) module is used to achieve cascaded aggregation of multi-granularity features, enhancing the ability to extract crack features at different scales. Furthermore, the Information Decoding and Fusion IDAF module is used to decode and optimize features from different scale coding layers and perform global semantic fusion, effectively reducing the difference in feature semantic levels between the coding and decoding layers. This reduces the computational complexity of the model while improving the semantic segmentation accuracy and robustness of airport runway cracks, achieving high-precision quantitative assessment and monitoring of cracks.

[0029] It should be noted that the execution entity involved in the embodiments of this application can be an airport runway crack monitoring system based on a multi-scale global feature fusion network, or it can be an electronic device. Optionally, the electronic device may include: a computer / laptop, a mobile terminal, a server, an airborne processing unit, or a control chip with computing processing functions, etc.

[0030] The following section uses an electronic device as an example to illustrate in detail the airport runway crack monitoring method based on a multi-scale global feature fusion network provided in this application: Figure 1 This is a flowchart illustrating the airport runway crack method based on a multi-scale global feature fusion network provided in this application embodiment. Figure 1 As shown, the method includes the following steps 101-103.

[0031] Step 101: Obtain an image of the airport runway surface.

[0032] Among them, airport runway surface images refer to standardized raw image data reflecting the health status of the runway surface, which are obtained through image acquisition modules during routine airport runway inspections.

[0033] Optionally, the image acquisition module can consist of a high-definition imaging device, a motion control unit, and a data transmission unit. It is used to acquire standardized images of the runway surface during airport runway inspections, supports wired / wireless data transmission, and is compatible with various acquisition carriers such as fixed inspection vehicles, drones, and portable acquisition devices, ensuring the resolution, clarity, and spatial consistency of the acquired images.

[0034] Alternatively, the high-definition imaging device may be a high-definition scanning camera or an imaging device mounted on a drone.

[0035] Step 102: Input the airport runway surface image into a multi-scale global feature fusion network to obtain the crack region segmentation result map corresponding to the airport runway surface image. The multi-scale global feature fusion network includes an encoder network and a decoder network. The encoder network is used to extract features from the airport runway surface image through the Deep Overlapping Block Embedding (DOPE) module and the Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map. The decoder network is used to decode and optimize the feature representation extracted from the encoder network and perform global semantic fusion through the Information Decoding and Fusion IDAF module to reduce the semantic information difference between the encoder network and the decoder network, thereby obtaining the crack region segmentation result map corresponding to the airport runway surface image.

[0036] Among them, the multi-scale global feature fusion network refers to a deep neural network architecture designed for airport runway crack features. It utilizes a symmetrical encoder-decoder structure to perform cross-level cascading and global correlation modeling of spatial features under different receptive fields, thereby achieving pixel-level accurate extraction of tiny crack targets in the complex background of airport runways.

[0037] For example, Figure 2 This is a schematic diagram of the structure of the multi-scale global feature fusion network provided in the embodiments of this application. For example... Figure 2 As shown, the Multi-scale Progressive Global Fusion Network (MPGF-Net) includes an encoder network and a decoder network. The encoder network is located in... Figure 2 The left-hand main path of the encoder network consists of multiple cascaded feature extraction layers. Specifically, the feature extraction layers of this encoder network are connected through the Deep Overlapping Patch Embedding (DOPE) module and the Progressive Patch Aggregation (PPA) module to extract features from airport runway surface images. Each feature extraction layer outputs a feature map corresponding to a different resolution, forming a feature hierarchy from shallow local texture to deep global semantics, i.e., generating multi-granularity cascaded fused feature maps.

[0038] The decoder network is located at Figure 2 In the right and middle node regions of the decoder network, multiple Information Decoding and Fusion (IDAF) modules are used to restore and enhance the multi-granularity cascaded fusion feature maps extracted by the encoder network. Specifically, the decoder network adopts a nested dense connection structure similar to Unet++, with node labels as follows. ,in, This indicates the resolution level of the feature map. This indicates the depth of feature fusion. The decoder network aggregates deep feature maps with features from encoder networks at the same and different levels through upsampling operations. It then uses the Information Decoding and Fusion IDAF module to decode and optimize these heterogeneous feature representations, achieving global semantic fusion to reduce the semantic information difference between the encoder and decoder networks.

[0039] It is understandable that, such as Figure 2As shown, the encoder network and decoder network are connected by dense skip connection paths (indicated by the dashed and solid arrows in the figure). These skip connection paths directly transmit the shallow spatial information generated by the encoder network at different stages to the corresponding decoder layer nodes. This allows the decoder network to fully utilize the minute crack details preserved in the encoding stage while restoring pixel-level resolution layer by layer. Finally, the information is transmitted from the rightmost output node (i.e., node...) Output the crack region segmentation result image with the same size as the original image.

[0040] A crack region segmentation result image refers to a binary image or probability heatmap generated after pixel-level classification of the original airport runway surface image. In this crack region segmentation result image, crack regions and background regions (such as normal pavement, markings, and stains) are assigned different pixel values ​​or color labels, thereby accurately outlining the geometric contours, location distribution, and extension direction of the cracks.

[0041] The DOPE module refers to the initial feature mapping component deployed in the input layer of the encoder network. The DOPE module divides the input airport runway surface image into blocks through an overlapping sliding window mechanism and uses a depth convolution operator to map the image pixels to a high-dimensional feature space.

[0042] The PPA module refers to a downsampling and feature compression component deployed between adjacent layers of each feature extraction layer in the encoder network. This PPA module uses a progressive spatial dimensionality reduction strategy to locally aggregate and expand the channels of high-resolution patches from the previous feature extraction layer.

[0043] The IDAF module refers to the module deployed in each nested node of the decoder network (i.e., The feature processing operator at () is responsible for aligning and weighting the heterogeneous features from the previous node at the same level, the upsampled node at the next higher level, and the jump connection path across levels.

[0044] It should be noted that by constructing a multi-scale coding structure that includes DOPE and PPA modules, and combining it with the decoding fusion mechanism of the IDAF module, deep collaboration between the features of tiny cracks in airport runways and global semantic information is achieved. This effectively eliminates the semantic gap between encoding and decoding, and significantly improves the accuracy and robustness of crack detection in complex environments.

[0045] In some embodiments, the step of an electronic device inputting an airport runway surface image into a multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image may include: the electronic device preprocessing the airport runway surface image to obtain a corresponding preprocessed image to be detected; the electronic device inputting the preprocessed image to be detected into an encoder network to obtain a multi-granularity cascaded fusion feature map; the electronic device inputting the multi-granularity cascaded fusion feature map into a decoder network to obtain a segmentation probability map of the airport runway cracks; and the electronic device binarizing the segmentation probability map of the airport runway cracks to obtain a crack region segmentation result map corresponding to the airport runway surface image.

[0046] Among them, the preprocessed image to be detected refers to the standardized image data generated after preprocessing the original acquired airport runway surface image, which meets the input requirements of the multi-scale global feature fusion network.

[0047] Optionally, preprocessing operations may include image denoising, grayscale conversion, size normalization, and pixel value standardization.

[0048] A segmentation probability map is a single-channel or multi-channel floating-point matrix output by the decoder network, with the same size as the preprocessed image to be detected. The value of each pixel in this segmentation probability map represents the confidence or probability that the pixel belongs to a crack region; the closer the value is to 1, the higher the probability that the pixel is a crack.

[0049] Binarization refers to the process of discretizing a segmented probability map using a preset probability threshold.

[0050] In this embodiment, the electronic device performs preprocessing on the acquired original airport runway surface image. Median filtering is used to suppress road surface noise, and the preprocessed airport runway surface image is scaled to a preset network size to obtain a preprocessed image to be detected. Next, the electronic device inputs this preprocessed image into an encoder network. Through progressive downsampling of the image's spatial dimensions and nonlinear mapping of the channel dimensions, semantic features at different spatial resolutions are extracted. Feature information is transferred and channel dimensions are amplified between different levels to generate a multi-granularity cascaded fusion feature map representing the crack texture and spatial distribution. Then, the electronic device inputs the multi-granularity cascaded fusion feature map into a decoder network. The features are upsampled and restored using the hierarchical nodes within the decoder network, and a segmentation probability map of the airport runway cracks is obtained through cross-level feature alignment and fusion. Finally, the electronic device binarizes the segmentation probability map of the airport runway cracks to obtain a segmentation result map of the crack region corresponding to the airport runway surface image.

[0051] It should be noted that by performing standardized preprocessing on the runway images and combining the segmentation probability map generated by the encoding and decoding architecture to perform binarization segmentation, a complete closed loop from raw data augmentation to pixel-level classification is constructed, which significantly improves the extraction accuracy and decision reliability of tiny cracks in complex pavement backgrounds.

[0052] In some embodiments, the step of the electronic device inputting the preprocessed image to be detected into the encoder network to obtain a multi-granularity cascaded fused feature map may include: the electronic device performing multi-scale feature embedding on the preprocessed image to be detected through the Deep Overlapping Block Embedding (DOPE) module to obtain an initial feature map corresponding to the preprocessed image to be detected; the electronic device performing cascaded feature aggregation on the initial feature map through the Progressive Patch Aggregation (PPA) module to obtain a multi-granularity cascaded fused feature map.

[0053] Among them, multi-scale feature embedding refers to using a sliding window mechanism with overlapping regions to perform pixel-level scanning of the preprocessed image to be detected, and mapping local blocks of the image to a high-dimensional space through multiple sets of convolution operators with different receptive fields while preserving the continuity of the neighborhood space, thereby achieving simultaneous capture of microscopic details and macroscopic orientation of cracks in the initial stage.

[0054] The initial feature map refers to the feature matrix output by the first layer of the encoder network, which contains high spatial resolution information of the original image and has preliminary semantic expression capabilities.

[0055] Cascaded feature aggregation refers to the operation of connecting or fusing feature representations of different depths by progressively reducing the spatial resolution of the feature map and doubling the number of channels.

[0056] In this embodiment, the electronic device first calls the DOPE module to perform overlapping block partitioning and convolutional mapping on the preprocessed image to be detected, generating an initial feature map that preserves the continuity of crack edges. Subsequently, the electronic device uses the PPA module to perform stepwise downsampling on the initial feature map, cascading feature patches of different levels by reducing the spatial scale and increasing the channel depth, and finally generating a multi-granularity cascaded fused feature map containing multi-scale receptive field information.

[0057] It should be noted that by using the overlapping embedding mechanism of the DOPE module and the progressive cascading aggregation of the PPA module, the neighborhood spatial continuity of the fine cracks can be effectively preserved from the initial stage, and the deep cross-level penetration of multi-scale features can be achieved, which significantly enhances the encoder network's ability to express features of complex crack shapes and the completeness of extraction.

[0058] In some embodiments, the step of the electronic device performing multi-scale feature embedding on the preprocessed image to be detected using the Deep Overlapped Block Embedding (DOPE) module to obtain an initial feature map corresponding to the preprocessed image to be detected may include: the electronic device performing gridded segmentation of the preprocessed image to be detected using a sliding window, setting the overlap ratio between the sliding windows to obtain multiple overlapping blocks corresponding to the preprocessed image to be detected; and inputting the multiple overlapping blocks into the Deep Overlapped Block Embedding (DOPE) module, which includes L convolutional layers, where L is an integer greater than 1; and in the first convolutional layer of the L convolutional layers... In this process, the electronic device converts the original pixel information of each overlapping block into a high-dimensional feature vector through linear mapping, and performs nonlinear normalization through an activation function to obtain preliminary embedding features. In each convolutional layer other than the first convolutional layer in the L-layer convolutional layer, the electronic device performs deep feature extraction and nonlinear activation on the decoupling information of multiple overlapping blocks in the current convolutional layer based on the embedding features output by the adjacent previous convolutional layer to obtain multi-level enhanced embedding features. The electronic device performs spatial dimensionality reduction on the multi-level enhanced embedding features output by the last convolutional layer in the L-layer convolutional layer to generate an initial feature map representing the spatial overlap relationship.

[0059] Optionally, the overlap ratio of the sliding window refers to the ratio of the number of pixels overlapping between adjacent sliding windows in the horizontal or vertical direction to the total number of pixels in the window. Optionally, the overlap ratio of the sliding window can be set to 30%.

[0060] Understandably, by setting a reasonable overlap ratio, adjacent blocks can share the contextual information of the edge region during the feature extraction stage, thereby effectively eliminating the boundary effect caused by traditional non-overlapping blocks and ensuring the continuity of crack edge feature extraction.

[0061] For example, Figure 3 This is a schematic diagram of the structure of the DOPE module embedded in the deep overlap block provided in an embodiment of this application. Combined with... Figure 3 The DOPE module includes a first convolutional layer, a first normalized activation layer, a second convolutional layer, a second normalized activation layer, and a downsampling layer (i.e., Downsampling, Down) connected in sequence.

[0062] Optionally, the first normalized activation layer is composed of a combination of batch normalization (BN) and a Gaussian error linear unit (GELU). It should be noted that the structure of the second normalized activation layer may be the same as or different from that of the first normalized activation layer.

[0063] Optionally, the sliding window size of the DOPE module is 64×64.

[0064] Optionally, for each convolutional layer, the kernel size is 3×3, the number of kernels is set to 64, the stride is 1, and the padding is 1.

[0065] Taking the DOPE module, which contains two convolutional layers, as an example, the process by which the electronic device obtains the initial feature map corresponding to the preprocessed image to be detected is as follows: First, the electronic device uses a sliding window with a size of 64×64 and an overlap ratio of 30% to perform grid-based segmentation of the preprocessed image to be detected, resulting in multiple overlapping blocks corresponding to the preprocessed image to be detected, so that some common pixel areas are retained between adjacent blocks.

[0066] Secondly, the electronic device inputs the acquired multiple overlapping blocks into the first convolutional layer of the DOPE module. In the first convolutional layer, the electronic device uses convolution kernels to perform convolution operations on each overlapping block, converting the original pixel information into a high-dimensional feature vector through linear mapping. Subsequently, the electronic device calls the first normalized activation layer to perform batch normalization processing and nonlinear normalization based on the GELU function on the high-dimensional feature vector to obtain preliminary embedded features.

[0067] Next, the electronic device inputs the initial embedded features into the second convolutional layer. The convolutional kernel further extracts local features from multiple overlapping blocks to obtain deeper semantic representations. A second normalized activation layer then performs non-linear activation to obtain multi-level enhanced embedded features. In this process, the cascading of two 3×3 convolutional layers significantly enhances the module's ability to perceive the micro-geometric textures of airport runways (such as fine cracks and pavement grooves).

[0068] Finally, the electronic device calls the downsampling layer to perform spatial resolution compression and channel mapping on the multi-level enhanced embedding features output by the last convolutional layer. Through spatial dimensionality reduction, the electronic device maps the high-dimensional embedding features into an initial feature map with rich semantics and spatial overlap, providing a high-quality basic representation for subsequent feature aggregation and global modeling.

[0069] It should be noted that, through the synergistic effect of the overlapping sliding window mechanism and the multi-layer nonlinear convolutional architecture in the DOPE module, the continuity of the edge of the small crack in the pavement is effectively preserved during the feature embedding stage. This eliminates the boundary fragmentation phenomenon commonly seen in block detection from the underlying logic and significantly improves the accuracy of local feature capture under complex texture interference.

[0070] In some embodiments, the step of the electronic device performing cascaded feature aggregation on the initial feature map using the Progressive Patch Aggregation (PPA) module to obtain a multi-granularity cascaded fused feature map may include: the encoder network includes N feature extraction layers, where N is an integer greater than 1; for the first feature extraction layer of the encoder network, the electronic device performs multi-granularity feature extraction on the initial feature map using the Progressive Patch Aggregation (PPA) module to obtain a first cascaded fused feature map; in each feature extraction layer other than the first feature extraction layer in the N feature extraction layers, the electronic device performs cross-scale feature aggregation on the cascaded fused feature map based on the cascaded fused feature map output by the adjacent previous feature extraction layer, using the Progressive Patch Aggregation (PPA) module of the current layer to generate a multi-granularity cascaded fused feature map corresponding to the feature extraction layer.

[0071] For example, combined Figure 2 The encoder network includes four feature extraction layers for downsampling. Taking an encoder network with four feature extraction layers as an example, the process of generating multi-granularity cascaded fusion feature maps corresponding to the feature extraction layers in an electronic device is as follows: First, for the first feature extraction layer (i.e., nodes) The electronic device calls the PPA module corresponding to the first feature extraction layer, inputs the received initial feature map into the PPA module, performs multi-granularity feature extraction on the initial feature map, initially extracts the edge pixels and texture primitives of the crack, and obtains the first cascaded fusion feature map; Secondly, regarding the second feature extraction layer (i.e., nodes) The electronic device calls the PPA module corresponding to the second feature extraction layer based on the first concatenated fusion feature map output by the first feature extraction layer. By performing strided convolution or pooling operations inside the PPA module, the spatial dimension of the first concatenated fusion feature map is halved, and the number of feature channels is increased simultaneously. While capturing a wider range of road surface semantic information, the second concatenated fusion feature map is generated.

[0072] Next, for the third feature extraction layer (i.e., nodes) The electronic device inputs the second-level cascaded fusion feature map into the corresponding PPA module. The PPA module captures complex contextual features and transforms the local fragmented crack features into a structured feature representation with a medium scale, enabling the encoder network to identify the extension trend and branching direction of the crack and generate the third-level cascaded fusion feature map.

[0073] Subsequently, for the fourth feature extraction layer (i.e., nodes) The electronic device performs deep semantic aggregation on the third-level concatenated fusion feature map. In the fourth feature extraction layer, the corresponding PPA module compresses the spatial resolution to a certain proportion of the original input, and generates the fourth-level concatenated fusion feature map by obtaining semantic information from the global perspective.

[0074] Finally, the electronic device horizontally integrates the cascaded fusion feature maps generated by the above feature extraction layers to ultimately generate a multi-granularity cascaded fusion feature map containing everything from low-level fine textures to high-level global semantic information.

[0075] It should be noted that through the progressive aggregation of multi-layer PPA modules, a smooth transition and deep fusion of feature information between different spatial scales are achieved, which effectively avoids the loss of minor crack features caused by abrupt changes in the receptive field in traditional downsampling operations. This ensures that the feature map retains the precise localization capability at the low-level pixel level while possessing high-level semantic discriminative power.

[0076] In some embodiments, the step of the electronic device extracting multi-granularity features from the initial feature map through the Progressive Patch Aggregation (PPA) module to obtain the first concatenated fused feature map may include: the electronic device performing feature mapping on the input initial feature map for multiple parallel convolutional branches in the Progressive Patch Aggregation (PPA) module using grouped convolutional kernels with different receptive field sizes to obtain granular features corresponding to each parallel convolutional branch; the electronic device concatenating the granular features output by each parallel convolutional branch according to the channel dimension to obtain the initial fused features; and the electronic device performing batch normalization and nonlinear activation on the initial fused features to obtain the first concatenated fused feature map.

[0077] For example, Figure 4 This is a schematic diagram of the progressive patch aggregation PPA module provided in an embodiment of this application. (Combined with...) Figure 4 The PPA module includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a normalized activation layer, and a downsampling layer, all in parallel.

[0078] Optionally, the structure of the normalized activation layer is the same as that of the first normalized activation layer described above.

[0079] Taking a PPA module containing three convolutional layers as an example, the process by which the electronic device obtains the first cascaded fused feature map is as follows: First, the electronic device synchronously inputs the initial feature map into three parallel convolutional layers: the first, second, and third. During this process, the device simultaneously performs feature mapping using three sets of grouped convolutional kernels with different receptive field sizes (e.g., kernel sizes of 3×3, 5×5, and 7×7, or dilated convolutions of the same size but with different dilation rates). Next, the device concatenates the fine-grained texture features output from the first convolutional layer, the medium-grained geometric features output from the second convolutional layer, and the coarse-grained semantic features output from the third convolutional layer, according to the channel dimension, to generate an initial fused feature containing multi-scale receptive field information. Subsequently, the device maps the feature distribution to a stable region through batch normalization and calls a non-linear activation function to enhance the non-linear expressive power of the features, outputting a mapped cascaded feature stream. Finally, the device calls a downsampling layer to perform spatial resolution compression on the initial fused features, reducing data dimensionality while completing deep aggregation of local patch features, ultimately obtaining a first cascaded fused feature map that matches the input requirements of the next feature extraction layer.

[0080] It should be noted that by using the parallel grouped convolutional architecture of the PPA module, multi-granularity disease features of the runway surface are extracted synchronously using multiple heterogeneous receptive fields. This significantly reduces the number of model parameters and computational complexity, while enhancing the network's accuracy in identifying micro-cracks and complex backgrounds. It achieves a balance between feature extraction efficiency and global context awareness, providing rich multi-scale information support for high-precision crack semantic segmentation.

[0081] In some embodiments, the step of the electronic device inputting a multi-granularity cascaded fused feature map into a decoder network to obtain a segmentation probability map of airport runway cracks may include: the electronic device, for each decoding layer in the decoder network, using the feature reconstruction unit of the Information Decoding and Fusion IDAF module, spatially reconstructing the multi-granularity cascaded fused feature map or the fused feature corresponding to the previous decoding layer to obtain a reconstructed fused feature map; the electronic device dimensionally aligns the reconstructed fused feature map, the output features of the feature extraction layer of the encoder network, and the output features of the intermediate nodes to obtain an aligned multi-scale feature flow; the feature extraction layer of the encoder network is at the same level as the current decoding layer, and the intermediate nodes are... The electronic device includes convolutional nodes in the nested skip connection path, excluding the encoder network's starting node. It inputs multi-scale feature streams to the semantic fusion unit of the information decoding and fusion IDAF module, using channel attention and spatial attention mechanisms for adaptive weight allocation and background noise suppression to obtain fused features corresponding to each decoding layer. The electronic device progressively converges the fused features of each decoding layer to the final decoding layer through nested skip connections, generating deep fusion features. These deep fusion features characterize the pixel-level edge precision and high-level semantic logic of the airport runway surface image. The electronic device then performs convolutional mapping and nonlinear mapping processing on the deep fusion image to obtain a segmentation probability map of airport runway cracks.

[0082] The feature reconstruction unit is used to restore the spatial dimension of the fusion features from deep, low-resolution layers by using deconvolution or interpolation upsampling operators, so that the fusion features are consistent with the features to be fused in the current layer in terms of length and width.

[0083] Optionally, the feature reconstruction unit uses 2×2 deconvolution for upsampling and combines it with 1×1 convolution to align the feature dimensions. Optionally, the features are uniformly mapped to 128 dimensions.

[0084] The semantic fusion unit is used to dynamically weight the converged multi-feature streams through parallel or serial channel attention and spatial attention mechanisms, aiming to enhance the semantic details related to cracks and suppress pavement background noise.

[0085] For example, combined Figure 2 The decoder network has 4 upsampling layers. Each upsampling layer is upsampled by deconvolution. The output layer uses 1×1 convolution to map the feature dimension to 2 (crack or background) and outputs a segmentation probability map through the Sigmoid activation function.

[0086] Figure 5 This is a schematic diagram of the information decoding and fusion IDAF module provided in an embodiment of this application. For the decoder network, taking the decoding layer node... Taking the IDAF module processing at the location as an example, the process by which the electronic device obtains the fused features is as follows: First, the electronic device obtains information from deep-level nodes (i.e., nodes). The fused feature map, due to the multiple downsampling of deep features, has a spatial resolution of only H / 4×W / 4, and cannot be directly fused with the current layer features. The electronic device calls the feature reconstruction unit in the IDAF module and uses the 2×2 deconvolution operator to upsample the fused feature map, restoring the spatial scale of the feature map from H / 4×W / 4 to H / 2×W / 2, and obtains the reconstructed fused feature map, realizing the accurate regression of deep semantics to shallow resolution in physical scale.

[0087] Secondly, the electronic device acquires multiple input feature streams. Specifically, due to the different convolution depths traversed by each path, the channel dimensions exhibit heterogeneity, allowing the electronic device to acquire the following heterogeneous features: First, features from the previous node at the same resolution level. The output features, represented by dashed lines, provide local semantic information of the current level after preliminary decoding; secondly, from lower-level nodes. The features, after 2×2 deconvolution upsampling and 1×1 convolution dimension alignment by the feature reconstruction unit, are represented by solid lines. This path provides deeper-level global semantic guidance for the current node, assisting in locating the approximate area of ​​the crack; third, features from encoder network nodes... and intermediate nodes The nested skip connection feature stream is passed across levels, and the path is represented by dots and lines. Then, the electronic device uses a 1×1 convolution kernel as a dimension alignment operator to uniformly map the number of channels of the above three feature streams to a preset value (such as 128 dimensions), eliminating the differences in numerical dimensions of the multi-scale feature streams and generating an aligned multi-scale feature stream.

[0088] Understandably, these multi-scale feature streams, through nested dense connections, directly permeate the rich spatial edge details preserved in the encoding stage to the current decoding node, effectively compensating for the minute crack information lost by deep features during multiple downsampling processes.

[0089] Subsequently, the electronic device inputs the aligned multi-scale feature stream into the semantic fusion unit of the IDAF module. The electronic device calls the built-in channel attention mechanism, models the importance of each channel through global pooling, automatically identifies and enhances high-contribution channels containing crack textures, weakens ineffective background channels, and generates channel-enhanced feature streams. At the same time, the electronic device uses a spatial attention mechanism to perform saliency analysis on the image plane of the channel-enhanced feature streams, actively suppresses background noise such as rubber wear, markings, and uneven lighting on the track surface by utilizing the correlation between pixels, so that the network attention is highly focused on the pixel-level edges of the cracks, thereby generating spatial activation feature streams.

[0090] Finally, the electronic device passed The operator performs nonlinear fusion of the weighted features (i.e., the spatial activation feature flow) to obtain the current node. The corresponding fusion feature. This fusion feature brings together collaborative information from each feature extraction layer of the encoder network, intermediate nodes at each level, and the deep decoder network. It not only has a clear expression of crack boundaries but also possesses robust semantic logic, providing a basis for the final output node. This lays the foundation for generating high-precision segmentation probability maps.

[0091] It is understandable that the above nodes The processing is performed in parallel or sequentially in each nested node of the decoder network.

[0092] For the final convergence output node of the decoder network First, the electronic device receives data from the previous node at the same level via the IDAF module. Parallel decoding features, from lower-level nodes Upsampled deep enhancement features, and features from the encoder network start node. With intermediate nodes , The system transmits ultra-high resolution spatial features. Next, the electronic device uses a 1×1 convolutional layer to perform final dimensionality compression and information integration on the converged multi-path features, generating a deep fusion feature that balances low-level edge sharpness with high-level semantic logic. Finally, the electronic device uses a 1×1 output convolutional layer to map the number of channels of the deep fusion feature to 2 (corresponding to the crack class and background class respectively), and calls the Sigmoid activation function to convert the mapped values ​​into probability values ​​in the [0,1] interval. Finally, the electronic device outputs a segmentation probability map with the same size as the original preprocessed image to be detected, completing the crack detection of the airport runway surface image.

[0093] It should be noted that by using the IDAF module for deep fusion at each level of nested nodes and the final output node, dynamic alignment of the shallow details of the encoder with the deep semantics of the decoder is achieved, ensuring that even in complex airport environments, high-precision crack segmentation results with closed boundaries and accurate semantics can still be output.

[0094] Step 103: Calculate the quantitative index of the crack based on the crack area segmentation result map to obtain the monitoring results of the airport runway cracks.

[0095] Among them, the quantitative index of cracks refers to the numerical parameters that characterize the severity of crack damage obtained by extracting geometric features and mapping physical scales of crack regions identified in the crack region segmentation result map using pixel-level analysis methods.

[0096] Optionally, the quantitative indicators of cracks include, but are not limited to, crack area, crack length, crack average width, and crack density.

[0097] Optionally, the crack area is calculated as follows: count the total number of crack pixels in the crack region segmentation result image, and calculate the actual physical area (unit: mm²) occupied by the crack region based on the preset physical area of ​​a single pixel.

[0098] Optionally, the crack length is calculated as follows: the central single-pixel connected component of the crack region is extracted using a skeleton extraction algorithm, the cumulative Euclidean distance of the skeleton pixels is calculated, and then converted into the actual length (unit: mm).

[0099] Optionally, the average crack width is calculated by dividing the calculated crack area by the crack length to obtain the equivalent average width (unit: mm) that reflects the degree of lateral crack propagation.

[0100] Optionally, the crack density is calculated as follows: within a preset unit observation area of ​​the runway, the ratio of the total length of all cracks to the observation area is calculated (unit: mm / m²).

[0101] In this embodiment, after acquiring the crack region segmentation result image generated in step 102, the electronic device first establishes a scale mapping relationship between pixel space and physical space, transforming the abstract pixel set in the crack region segmentation result image into geometric parameters with physical meaning. Secondly, the electronic device uses a connected component analysis algorithm to identify each independent crack target in the crack region segmentation result image and labels each independent crack target. Next, the electronic device calculates the crack area by statistically analyzing the total number of pixels of each crack target, uses a skeleton extraction algorithm to extract the crack center skeleton to determine the crack length, calculates the average crack width, and combines this with the unit observation area to derive the crack density. Finally, the electronic device generates the monitoring results of the airport runway cracks based on the crack region segmentation result image and the aforementioned quantitative indicators.

[0102] Optionally, the monitoring results include a crack region segmentation map and various quantitative indicators. The crack region segmentation map includes crack region labels, and each quantitative indicator records the center coordinates, geometric parameters, and corresponding damage degree score for each crack.

[0103] Optionally, based on the above monitoring results, the electronic equipment can perform further auxiliary maintenance analysis: by comparing quantitative indicators with maintenance thresholds using preset disease level evaluation standards. For example, if the average crack width is less than a first preset threshold, preventative crack filling or sealing treatment is performed to prevent moisture from seeping into the structural layer; if the calculated crack density exceeds a preset second threshold, it is recommended to perform pavement milling and repaving or structural reinforcement, thereby providing highly accurate decision support for airport runway maintenance departments.

[0104] For example, Figure 6 This is a comparative schematic diagram of the airport runway crack area segmentation results provided in the embodiments of this application. Figure 6 This visually demonstrates the comparison between the original airport runway surface image and the binary mask image (cracks are white, background is black) generated by the model prediction. This binary mask image serves as the basic data source for the extraction of the aforementioned quantitative indicators, with the set of white pixels constituting the extracted objects of physically meaningful geometric parameters. Through this combination of text and images, maintenance departments can simultaneously obtain the intuitive spatial distribution of defects and refined physical quantitative attributes, thereby achieving closed-loop monitoring of the airport runway's health status.

[0105] It should be noted that by transforming high-precision semantic segmentation results into multi-dimensional physical quantitative indicators, a leap from qualitative disease identification to quantitative severity assessment has been achieved, providing objective, accurate, and decision-guiding digital basis for automated inspection of airport runways.

[0106] In this embodiment, an airport runway surface image is acquired; the airport runway surface image is input into a multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image; wherein, the multi-scale global feature fusion network includes an encoder network and a decoder network, the encoder network is used to extract features from the airport runway surface image through a Deep Overlapping Block Embedding (DOPE) module and a Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map; the decoder network is used to decode and optimize the feature representation extracted from the encoder network and perform global semantic fusion through an Information Decoding and Fusion IDAF module to reduce the semantic information difference between the encoder network and the decoder network, thereby obtaining a crack region segmentation result map corresponding to the airport runway surface image; the crack quantification index is calculated based on the crack region segmentation result map to obtain the monitoring result of the airport runway crack. This method constructs a multi-scale global feature fusion network and uses the Deep Overlapping Block Embedding (DOPE) module to compensate for spatial overlap features, thus solving the problem of spatial continuity loss caused by block processing. It utilizes the Progressive Patch Aggregation (PPA) module to achieve cascaded aggregation of multi-granularity features, enhancing the ability to extract crack features at different scales. Furthermore, it uses the Information Decoding and Fusion IDAF module to decode and optimize features from different scale coding layers and perform global semantic fusion, effectively reducing the feature semantic hierarchy differences between the coding and decoding layers. This reduces the computational complexity of the model while improving the semantic segmentation accuracy and robustness of airport runway cracks, enabling high-precision quantitative assessment and monitoring of cracks.

[0107] The following section uses an electronic device as the execution subject to elaborate on the multi-scale global feature fusion network training method provided in the embodiments of this application: Figure 7This is a flowchart illustrating the multi-scale global feature fusion network training method provided in this application embodiment. Figure 7 As shown, the method includes the following steps 701-703.

[0108] Step 701: Obtain sample images of the airport runway surface and perform pixel-level annotation on the crack areas of the airport runway surface sample images to obtain annotated preprocessed training samples.

[0109] Cracked areas refer to a set of pixels in an airport runway surface sample image that are formed by physical damage, load pressure, or environmental factors and have a specific geometric shape (such as linear, mesh, or band-like).

[0110] Annotated preprocessed training samples refer to standardized data pairs generated by preprocessing and labeling crack and lesion pixels with category labels, which contain the original image data and the corresponding pixel-level mask labels.

[0111] In some embodiments, the step of an electronic device acquiring sample images of an airport runway surface and performing pixel-level annotation on crack regions in the airport runway surface sample images to obtain annotated preprocessed training samples may include: the electronic device acquiring surface images covering different lighting conditions, weather environments, and runway areas during airport runway inspection using an image acquisition device to obtain an original image dataset; the electronic device using an annotation tool to classify and annotate crack regions in the original image dataset to generate pixel-level crack semantic segmentation label files corresponding one-to-one with the original image dataset; the crack categories include at least transverse cracks, longitudinal cracks, and mesh cracks; the electronic device dividing the original image dataset and the corresponding pixel-level crack semantic segmentation label files into a training set, a validation set, and a test set according to a preset ratio; the electronic device preprocessing the training set and using data augmentation strategies to expand the number of samples to obtain annotated preprocessed training samples.

[0112] The image acquisition device refers to a mobile inspection terminal equipped with a high-definition camera. In this embodiment, a runway inspection vehicle equipped with a 20-megapixel high-definition line scan camera is used to capture the texture details of the runway surface during vehicle movement.

[0113] The original image dataset refers to a collection of images containing real-world defects, acquired under various variable conditions. Specifically, this original image dataset covers surface images from different time periods (e.g., morning, noon, and evening), different weather conditions (e.g., sunny and cloudy), and different functional areas (e.g., main runway and taxiway). It ensures that the original image dataset covers various crack types, such as transverse cracks, longitudinal cracks, network cracks, and micro-discontinuous cracks, and includes complex background interference such as uneven lighting, road surface stains, and runway marking interference.

[0114] For example, the original image dataset mentioned above contains 1000 images.

[0115] Annotation tools refer to semantic segmentation annotation platforms that support pixel-level fine-grained operations.

[0116] In this embodiment, the LabelStudio semantic segmentation annotation tool is used, and professional inspectors delineate the horizontal, vertical, mesh-like, and minute discontinuous cracks in the original image.

[0117] Pixel-level crack semantic segmentation tag files refer to attribute tag files that are spatially aligned with the original image.

[0118] Data augmentation strategies refer to image processing techniques that simulate sample diversity by transforming the image using algorithms. These data augmentation strategies include one or more combinations of random cropping, random rotation (0-360°), random horizontal / vertical flipping, scaling, and brightness perturbation.

[0119] In this embodiment, the electronic device first controls a runway inspection vehicle equipped with a 20-megapixel high-definition line scan camera to perform multi-dimensional image acquisition during airport inspection tasks. The electronic device captures original images with a resolution of 1920×1080, covering different lighting conditions, weather conditions, and functional areas, by setting a specific sampling frequency, thus constructing an original image dataset containing horizontal, vertical, mesh-like, and minute discontinuous cracks.

[0120] Secondly, with the assistance of professional inspectors, the electronic device uses the open-source multimodal data annotation platform (Label Studio) to delineate the crack contours in the image pixel by pixel, generating an initial annotation file in JSON format. Subsequently, the electronic device automatically converts the initial annotation file into a pixel-level crack semantic segmentation label file in 8-bit grayscale format. In this label file, pixels in the crack area are uniformly assigned a value of 255 (white), and pixels in the background area are uniformly assigned a value of 0 (black), thereby constructing binary supervision information that is strictly aligned with the spatial position of the original image.

[0121] Next, the electronic device logically partitions the labeled image-label pairs. The device divides the dataset into training, validation, and test sets according to a preset ratio of 7:2:1 to establish the benchmark set, hyperparameter tuning set, and generalization evaluation set for network training.

[0122] Finally, the electronic device performs deep data augmentation on the training set. It calls a pre-defined algorithm library and employs a combination of strategies including random cropping (e.g., 1024×1024 size), random rotation (0-360°), random flipping, scaling (0.8-1.2 times), and brightness perturbation (±20%) to expand the number of samples in the original training set (e.g., from 1000 to 5600), thereby generating labeled preprocessed training samples that can effectively represent the diverse physical environment characteristics.

[0123] It should be noted that by combining multi-dimensional image acquisition with combined data augmentation strategies, a crack sample space with extremely high generalization ability was constructed. This not only effectively alleviated the overfitting problem of deep learning models under small sample conditions, but also provided solid labeled data support for the subsequent network to extract high-precision features in complex runway backgrounds.

[0124] Step 702: Input the preprocessed training samples into the initialized multi-scale global feature fusion network for forward propagation processing to generate a sample segmentation probability map.

[0125] The initialized multi-scale global feature fusion network refers to a deep learning model built on the PyTorch 2.0 framework and Python 3.13 programming environment, with network weights and parameters already randomly initialized or pre-trained. The module design and parameter settings of this initialized multi-scale global feature fusion network are as follows: For the infrastructure layer, the initialized multi-scale global feature fusion network adopts an improved Unet++ nested dense connection architecture. The encoder network is configured with 4 downsampling layers, each consisting of two cascaded 3×3 convolutional layers, a BN layer, and a ReLU activation layer, using 2×2 max pooling to perform spatial dimensionality reduction; the decoder network is configured with 4 upsampling layers, using deconvolution operators to restore spatial resolution, and achieving multi-depth feature penetration through nested skip connection paths.

[0126] For the parameter configuration of the DOPE module, the DOPE module is deployed in the input layer of the encoder network. A sliding window with a size of 64×64 and an overlap ratio of 30% is used to perform gridding on the input sample image. For each overlapping block, feature embedding is performed using a 3×3 convolution with 64 kernels, a stride of 1, and padding of 1 to generate an initial feature map with spatial overlap logic.

[0127] For the parameter configuration of the PPA module, the PPA module is embedded in all four feature extraction layers of the encoder network. Each PPA module uses three parallel 3×3 convolutional kernels for multi-granularity feature extraction, with each kernel having 64 kernels. To optimize computational efficiency, feature transformation is performed using grouped convolutions with 8 groups. Subsequently, multi-scale features from different receptive fields are cascaded and fused along the channel dimension to achieve cross-scale feature information interaction.

[0128] For the parameter configuration of the IDAF module, the IDAF module is deployed at the intersection of skip connections in each layer of the decoder network. The IDAF module includes a feature reconstruction unit and a semantic fusion unit. The feature reconstruction unit uses 2×2 deconvolution for spatial recovery, and works with 1×1 convolution to uniformly align and map the dimensions of the heterogeneous features to 128 dimensions. The semantic fusion unit integrates the squeeze-and-excitation (SE) channel attention mechanism and the attention mechanism in the Convolutional Block Attention Module (CBAM).

[0129] In step 702, firstly, the electronic device utilizes the DOPE module in the encoder network to perform overlapping sliding window segmentation and high-dimensional feature embedding on the labeled preprocessed training samples. The overlapping window mechanism captures the microscopic geometric continuity of the cracks, outputting an initial feature map. Subsequently, the electronic device inputs the initial feature map into the backbone path of the encoder network (i.e., each feature extraction layer). At each layer, the PPA module is called to cascade multi-granularity feature flows generated by grouped convolutions, generating a more representative multi-granularity cascaded fused feature map. Next, these feature maps are passed to the IDAF module of the decoder network. The IDAF module performs spatial reconstruction and dimensional alignment on the feature representations from different layers of the encoder network, and introduces channel attention and spatial attention mechanisms to adaptively weight and fuse heterogeneous feature flows, obtaining the decoded deep fused features, providing a predictive basis for the calculation of the loss function. Finally, the electronic device maps the decoded deep fused features to the classification space, generating a sample segmentation probability map representing the confidence level of each pixel belonging to the crack region through a nonlinear mapping function.

[0130] It should be noted that by building a multi-scale network model integrating DOPE overlapping embedding, PPA parallel aggregation, and IDAF attention fusion under the PyTorch framework, efficient learning of the complex texture of runway cracks is achieved from the bottom layer of the algorithm, which significantly improves the convergence speed and feature discrimination accuracy of the model during the training phase.

[0131] Step 703: Construct a loss function based on the semantic deviation between the sample segmentation probability map and the corresponding pixel-level annotation, and use the backpropagation algorithm to adjust the parameters of the multi-scale global feature fusion network until the loss function converges, thus obtaining the trained multi-scale global feature fusion network.

[0132] Optionally, the loss function adopts a hybrid loss of cross-entropy loss and Dice loss, with the weight ratio of the two set to 1:1.

[0133] Understandably, cross-entropy loss focuses on pixel-level classification accuracy, while Dice loss focuses on solving the problem of extreme imbalance between crack target and background region samples (i.e., crack pixels are far fewer than background pixels). By using hybrid loss calculation, the network can ensure the stability of global segmentation while capturing tiny cracks.

[0134] Optionally, the backpropagation algorithm employs the Adaptive Moment Estimation (Adam) optimizer, with parameters set as follows: , .

[0135] For example, the training environment of this multi-scale global feature fusion network is an Intel Xeon Gold 6230 CPU for the central processing unit (CPU) and an NVIDIA ARTX 3090 GPU (24GB VRAM) for the graphics processing unit (GPU). GPU-accelerated computing is implemented based on Compute Unified Device Architecture (CUDA) 11.8 and CUDA Deep Neural Network library (cuDNN) 8.4.

[0136] In step 703, the electronic device initializes the network weights using the He initialization method to adapt to the ReLU or GELU activation function and avoid gradient vanishing in the early stages of training. Simultaneously, the initial learning rate is set to 0.001, and the batch size is 8. Next, the electronic device calculates the semantic deviation between the sample segmentation probability map and the pixel-level annotations, constructing a hybrid loss function using cross-entropy loss and Dice loss in a 1:1 weight ratio. Then, the electronic device calls the Adam optimizer, adjusting the weight parameters in the DOPE, PPA, and IDAF modules layer by layer from the decoder network to the encoder network based on the gradient signal generated by the loss function, achieving progressive optimization of the network's ability to express crack features. Finally, after each training epoch, the electronic device inputs validation set samples into the current model and calculates semantic segmentation evaluation metrics including the loss function value, Intersection over Union (IoU), Mean Intersection over Union (mIoU), and Dice coefficient. Electronic devices introduce an early stopping mechanism: if the loss function value on the validation set no longer decreases in 20 consecutive training rounds, the network is considered to have reached convergence and training is stopped.

[0137] It should be noted that in this embodiment, the multi-scale global feature fusion network achieves optimal performance in the 65th round. At this point, the validation set loss reaches its minimum, IoU reaches 92.5%, mIoU reaches 93.1%, and the Dice coefficient reaches 94.2%. Compared to the original Unet++ network, all indicators are improved by more than 5%, the number of model parameters is reduced by 32%, and the inference speed is improved by 40%. The electronic device saves the network parameters at this point as the final trained multi-scale global feature fusion network.

[0138] In the embodiments of this application, the technical solutions of steps 701-703 above significantly improve the feature discrimination power and generalization performance of the network model in complex backgrounds by constructing a multi-dimensional enhanced crack dataset, cooperating with an innovative network architecture integrating DOPE, PPA and IDAF modules, and using a hybrid loss function for collaborative optimization, thereby achieving a simultaneous leap in model lightweighting and detection accuracy and inference speed.

[0139] The airport runway crack system based on a multi-scale global feature fusion network provided in the embodiments of this application is described below. The airport runway crack system based on a multi-scale global feature fusion network described below can be referred to in correspondence with the airport runway crack method based on a multi-scale global feature fusion network described above.

[0140] Figure 8This is a schematic diagram of the structure of an airport runway crack system based on a multi-scale global feature fusion network provided in an embodiment of this application. Figure 8 As shown, the system includes: an image acquisition module 801, a feature extraction module 802, and a crack monitoring module 803.

[0141] Image acquisition module 801 is used to acquire images of the airport runway surface.

[0142] The feature extraction module 802 is used to input the airport runway surface image into a multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image. The multi-scale global feature fusion network includes an encoder network and a decoder network. The encoder network extracts features from the airport runway surface image using a Deep Overlapping Block Embedding (DOPE) module and a Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map. The decoder network decodes and optimizes the feature representation extracted from the encoder network and performs global semantic fusion using an Information Decoding and Fusion IDAF module to reduce the semantic information difference between the encoder network and the decoder network, thereby obtaining a crack region segmentation result map corresponding to the airport runway surface image.

[0143] The crack monitoring module 803 is used to calculate the quantitative index of the crack based on the crack area segmentation result map, and obtain the monitoring results of the airport runway crack.

[0144] Optionally, the feature extraction module 802 is specifically used to preprocess the airport runway surface image to obtain a corresponding preprocessed image to be detected; input the preprocessed image to be detected into the encoder network to obtain a multi-granularity cascaded fusion feature map; input the multi-granularity cascaded fusion feature map into the decoder network to obtain a segmentation probability map of the airport runway crack; and perform binarization processing on the segmentation probability map of the airport runway crack to obtain a crack region segmentation result map corresponding to the airport runway surface image.

[0145] Optionally, the feature extraction module 802 is specifically used to perform multi-scale feature embedding on the preprocessed image to be detected through the Deep Overlapping Block Embedding (DOPE) module to obtain an initial feature map corresponding to the preprocessed image to be detected; and to perform cascaded feature aggregation on the initial feature map through the Progressive Patch Aggregation (PPA) module to obtain a multi-granularity cascaded fused feature map.

[0146] Optionally, the feature extraction module 802 is specifically used to perform gridded segmentation of the preprocessed image to be detected using a sliding window, set the overlap ratio of the sliding window, and obtain multiple overlapping blocks corresponding to the preprocessed image to be detected; and input the multiple overlapping blocks into the Deep Overlap Block Embedding (DOPE) module, which includes L convolutional layers, where L is an integer greater than 1; in the first convolutional layer of the L convolutional layers, the original pixel information of each overlapping block is converted into a high-dimensional feature vector through linear mapping, and nonlinearly normalized through an activation function to obtain preliminary embedding features; in each convolutional layer of the L convolutional layers other than the first convolutional layer, deep feature extraction and nonlinear activation are performed on the decoupling information of the multiple overlapping blocks in the current convolutional layer based on the embedding features output by the adjacent previous convolutional layer to obtain multi-level enhanced embedding features; and spatial dimensionality reduction is performed on the multi-level enhanced embedding features output by the last convolutional layer of the L convolutional layers to generate an initial feature map representing the spatial overlap relationship.

[0147] Optionally, the feature extraction module 802 is specifically used to perform multi-granularity feature extraction on the initial feature map of the first feature extraction layer of the encoder network through the progressive patch aggregation (PPA) module to obtain a first cascaded fusion feature map; in each feature extraction layer other than the first feature extraction layer in the N-layer feature extraction layers, based on the cascaded fusion feature map output by the adjacent previous feature extraction layer, the progressive patch aggregation (PPA) module of the current layer performs cross-scale feature aggregation on the cascaded fusion feature map to achieve progressive expansion of the receptive field and generate a multi-granularity cascaded fusion feature map of the corresponding feature extraction layer.

[0148] Optionally, the feature extraction module 802 is specifically used to perform feature mapping on the input initial feature map for multiple parallel convolutional branches in the progressive patch aggregation (PPA) module by using grouped convolutional kernels with different receptive field sizes to obtain granular features corresponding to each parallel convolutional branch; to concatenate the granular features output by each parallel convolutional branch according to the channel dimension to obtain the initial fused features; and to perform batch normalization and nonlinear activation on the initial fused features to obtain the first cascaded fused feature map.

[0149] Optionally, the feature extraction module 802 is specifically used to perform spatial reconstruction on the multi-granularity cascaded fusion feature map or the fusion feature corresponding to the previous decoding layer for each decoding layer in the decoder network using the feature reconstruction unit of the information decoding and fusion IDAF module, to obtain the reconstructed fusion feature map; to dimensionally align the reconstructed fusion feature map, the output features of the feature extraction layer of the encoder network, and the output features of the intermediate nodes, to obtain the aligned multi-scale feature flow; the feature extraction layer of the encoder network is at the same level as the current decoding layer, and the intermediate node is a convolutional node in the nested skip connection path except for the starting node of the encoder network; to input the multi-scale feature flow to the semantic fusion unit of the information decoding and fusion IDAF module, and to perform adaptive weight allocation and background noise suppression using channel attention mechanism and spatial attention mechanism, to obtain the fusion features corresponding to each decoding layer; to progressively converge the fusion features of each decoding layer to the final decoding layer through nested skip connections, generating a deep fusion feature, which is used to characterize the pixel-level edge precision and high-level semantic logic of the airport runway surface image; to perform convolutional mapping and nonlinear mapping processing on the deep fusion image to obtain the segmentation probability map of the airport runway crack.

[0150] It should be noted that this airport runway crack system based on a multi-scale global feature fusion network adopts a modular design and can be deployed on both ends of the runway inspection vehicle's edge computing equipment and the airport operation and maintenance center server to achieve distributed detection and centralized management. Specifically, the image acquisition module 801 and the crack monitoring module 803 are deployed on the edge computing equipment (such as an embedded GPU) of the runway inspection vehicle to achieve real-time image acquisition and inference, and output crack segmentation results on-site for inspection personnel to view. The feature extraction module 802 is deployed on the airport operation and maintenance center server to achieve centralized management of the dataset, iterative training and updating of the model, and receive detection data uploaded from the edge device for data aggregation, analysis and visualization, generating airport runway crack health monitoring reports, and realizing normalized and intelligent runway health management.

[0151] Figure 9 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 9 As shown, the electronic device may include a processor 910, a communications interface 920, a memory 930, and a communication bus 940. The processor 910, communications interface 920, and memory 930 communicate with each other via the communication bus 940. The processor 910 can call logical instructions from the memory 930 to execute an airport runway crack method based on a multi-scale global feature fusion network.

[0152] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0153] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the airport runway crack method based on a multi-scale global feature fusion network provided by the above methods.

[0154] In another aspect, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the airport runway crack method based on a multi-scale global feature fusion network provided by the above methods.

[0155] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0156] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for airport runway crack monitoring based on a multi-scale global feature fusion network, characterized in that, include: Acquire images of the airport runway surface; The airport runway surface image is input into a multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image. The multi-scale global feature fusion network includes an encoder network and a decoder network. The encoder network extracts features from the airport runway surface image using a Deep Overlapping Block Embedding (DOPE) module and a Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map. The decoder network decodes and optimizes the feature representation extracted from the encoder network and performs global semantic fusion using an Information Decoding and Fusion IDAF module to reduce the semantic information difference between the encoder network and the decoder network, thus obtaining the crack region segmentation result map corresponding to the airport runway surface image. The quantitative index of the crack is calculated based on the crack area segmentation result map to obtain the monitoring results of the airport runway crack.

2. The airport runway crack monitoring method based on the multi-scale global feature fusion network according to claim 1, characterized in that, The step of inputting the airport runway surface image into a multi-scale global feature fusion network to obtain the crack region segmentation result map corresponding to the airport runway surface image specifically includes: The airport runway surface image is preprocessed to obtain the corresponding preprocessed image to be detected; The preprocessed image to be detected is input into the encoder network to obtain a multi-granularity cascaded fusion feature map. The multi-granularity cascaded fusion feature map is input into the decoder network to obtain the segmentation probability map of the airport runway crack; The segmentation probability map of the airport runway cracks is binarized to obtain the crack region segmentation result map corresponding to the airport runway surface image.

3. The airport runway crack monitoring method based on the multi-scale global feature fusion network according to claim 2, characterized in that, The step of inputting the preprocessed image to be detected into the encoder network to obtain a multi-granularity cascaded fusion feature map specifically includes: The Deep Overlapping Block Embedding (DOPE) module is used to embed multi-scale features into the preprocessed image to be detected, thereby obtaining the initial feature map corresponding to the preprocessed image to be detected. The initial feature map is cascaded and aggregated using the progressive patch aggregation (PPA) module to obtain a multi-granularity cascaded fused feature map.

4. The airport runway crack monitoring method based on the multi-scale global feature fusion network according to claim 3, characterized in that, The step of performing multi-scale feature embedding on the preprocessed image to be detected using the Deep Overlapping Block Embedding (DOPE) module to obtain the initial feature map corresponding to the preprocessed image to be detected specifically includes: The image to be detected and preprocessed is divided into grids using a sliding window. The overlap ratio of the sliding window is set to obtain multiple overlapping blocks corresponding to the image to be detected and preprocessed. The multiple overlapping blocks are then input into the Deep Overlap Block Embedding (DOPE) module, which includes L convolutional layers, where L is an integer greater than 1. In the first convolutional layer of the L-layer convolutional layer, the original pixel information of each overlapping block is converted into a high-dimensional feature vector through linear mapping, and then non-linearly normalized by the activation function to obtain the preliminary embedded features. In each convolutional layer other than the first convolutional layer in the L-layer convolutional layer, deep feature extraction and nonlinear activation are performed on the decoupling information of the multiple overlapping blocks in the current convolutional layer based on the embedding features output by the adjacent previous convolutional layer to obtain multi-level enhanced embedding features. The multi-level enhanced embedding features output by the last convolutional layer in the L-layer convolutional layer are spatially reduced to generate an initial feature map representing the spatial overlap relationship.

5. The airport runway crack monitoring method based on the multi-scale global feature fusion network according to claim 3, characterized in that, The step of performing cascaded feature aggregation on the initial feature map through the progressive patch aggregation (PPA) module to obtain a multi-granularity cascaded fused feature map specifically includes: The encoder network includes N feature extraction layers, where N is an integer greater than 1; For the first feature extraction layer of the encoder network, the progressive patch aggregation (PPA) module is used to perform multi-granular feature extraction on the initial feature map to obtain the first cascaded fused feature map. In each of the N feature extraction layers other than the first feature extraction layer, based on the cascaded fusion feature map output by the adjacent previous feature extraction layer, the progressive patch aggregation (PPA) module of the current layer performs cross-scale feature aggregation on the cascaded fusion feature map to generate a multi-granularity cascaded fusion feature map corresponding to the feature extraction layer.

6. The airport runway crack monitoring method based on a multi-scale global feature fusion network according to claim 5, characterized in that, The step of extracting multi-granularity features from the initial feature map using the progressive patch aggregation PPA module to obtain the first cascaded fused feature map specifically includes: For the multiple parallel convolutional branches in the progressive patch aggregation PPA module, feature mapping is performed on the input initial feature map by grouped convolutional kernels with different receptive field sizes to obtain the granular features corresponding to each parallel convolutional branch. The granular features output by each of the parallel convolution branches are concatenated according to the channel dimension to obtain the initial fused features; The initial fusion features are batch normalized and nonlinearly activated to obtain the first cascaded fusion feature map.

7. The airport runway crack monitoring method based on a multi-scale global feature fusion network according to claim 2, characterized in that, The information decoding and fusion IDAF module includes a feature reconstruction unit and a semantic fusion unit. The step of inputting the multi-granularity cascaded fused feature map into the decoder network to obtain the segmentation probability map of the airport runway crack specifically includes: For each decoding layer in the decoder network, the feature reconstruction unit of the information decoding and fusion IDAF module is used to spatially reconstruct the multi-granularity cascaded fusion feature map or the fusion feature corresponding to the previous decoding layer to obtain the reconstructed fusion feature map; the reconstructed fusion feature map, the output features of the feature extraction layer of the encoder network, and the output features of the intermediate nodes are dimensionally aligned to obtain the aligned multi-scale feature flow; the feature extraction layer of the encoder network is at the same level as the current decoding layer, and the intermediate nodes are convolutional nodes in the nested skip connection path other than the starting node of the encoder network; The multi-scale feature stream is input into the semantic fusion unit of the information decoding and fusion IDAF module. Adaptive weight allocation and background noise suppression are performed using channel attention mechanism and spatial attention mechanism to obtain the fused features corresponding to each decoding layer. By progressively converging the fusion features of each decoding layer to the final decoding layer through nested skip connections, a deep fusion feature is generated. The deep fusion feature is used to characterize the pixel-level edge precision and high-level semantic logic of the airport runway surface image. The deep fused image is processed by convolutional mapping and nonlinear mapping to obtain the segmentation probability map of the airport runway crack.

8. A multi-scale global feature fusion network training method, characterized in that, The method is used to train a multi-scale global feature fusion network as described in any one of claims 1-7, and the method includes: Acquire sample images of the airport runway surface and perform pixel-level annotation on the crack areas in the airport runway surface sample images to obtain annotated preprocessed training samples. The preprocessed training samples are input into the initialized multi-scale global feature fusion network for forward propagation to generate a sample segmentation probability map. A loss function is constructed based on the semantic deviation between the sample segmentation probability map and the corresponding pixel-level annotation. The parameters of the multi-scale global feature fusion network are adjusted using the backpropagation algorithm until the loss function converges, thus obtaining the trained multi-scale global feature fusion network. 9.The method of claim 8, wherein, The steps of acquiring airport runway surface sample images and pixel-level annotating crack areas in the airport runway surface sample images to obtain annotated preprocessed training samples specifically include: During airport runway inspections, surface images covering different lighting conditions, weather environments, and runway areas are acquired using image acquisition equipment to obtain raw image datasets. The crack regions in the original image dataset are classified and labeled using annotation tools to generate pixel-level crack semantic segmentation label files that correspond one-to-one with those in the original image dataset; the crack categories include at least transverse cracks, longitudinal cracks, and mesh cracks. The original image dataset and the corresponding pixel-level crack semantic segmentation label file are divided into training set, validation set and test set according to a preset ratio; The training set is preprocessed, and data augmentation strategies are used to expand the number of samples to obtain the labeled preprocessed training samples.

10. An airport runway crack monitoring system based on a multi-scale global feature fusion network, characterized in that, include: Image acquisition module, used to acquire images of the airport runway surface; A feature extraction module is used to input the airport runway surface image into a multi-scale global feature fusion network to obtain a crack region segmentation result map corresponding to the airport runway surface image. The multi-scale global feature fusion network includes an encoder network and a decoder network. The encoder network extracts features from the airport runway surface image using a Deep Overlapping Block Embedding (DOPE) module and a Progressive Patch Aggregation (PPA) module to generate a multi-granularity cascaded fused feature map. The decoder network decodes and optimizes the feature representation extracted from the encoder network and performs global semantic fusion using an Information Decoding and Fusion IDAF module to reduce the semantic information difference between the encoder network and the decoder network, thereby obtaining the crack region segmentation result map corresponding to the airport runway surface image. The crack monitoring module is used to calculate the quantitative index of the crack based on the crack area segmentation result map, and obtain the monitoring result of the airport runway crack.

Images