A lightweight multiscale crack detection method for structural monitoring
Patent Information
- Application Number
- CN202610851440.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-12
- Publication Date
- 2026-08-18
AI Technical Summary
首先，高精度的裂缝检测通常依赖于复杂的深度神经网络模型，虽然能够实现较高的检测准确率，但其模型参数量大、计算复杂度高，难以部署在无人机、巡检机器人等资源受限的边缘设备上，导致无法实现实时检测，严重制约了其在野外或移动巡检场景中的应用
1、本申请的方法采用轻量级卷积神经网络作为特征提取主干，引入深度可分离卷积与线性瓶颈结构，在保障特征表达能力的同时大幅降低了模型参数量与计算开销，使得模型可在无人机、巡检机器人等资源受限的边缘设备上实现高效推理，有效克服了传统高精度模型因计算冗余而难以部署于移动巡检平台的技术瓶颈，实现了检测效率与精度的协同优化。
Smart Images
Figure CN122597375A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of crack detection, and more specifically to a lightweight multi-scale crack detection method for structural monitoring. Background Technology
[0002] In the field of modern infrastructure construction and operation and maintenance, the safety of engineering structures such as bridges, tunnels, dams, and road surfaces is directly related to public safety and socio-economic development. With increasing service life, these structures are highly susceptible to surface damage such as cracks under the influence of environmental erosion, load fatigue, and geological changes. Cracks are not only an early manifestation of structural performance degradation but also a potential cause of structural failure. Therefore, timely and accurate detection and assessment of surface cracks are crucial for structural health monitoring. Traditional crack detection mainly relies on manual inspections, which suffers from low efficiency, strong subjectivity, limited coverage, and difficulty in quantification, failing to meet the high-frequency, large-scale, and high-precision monitoring needs of modern large-scale infrastructure. In recent years, automated image detection technology based on computer vision has gradually become a research hotspot, especially the breakthroughs in deep learning methods for object detection and semantic segmentation tasks, providing a new technical path for intelligent crack identification.
[0003] In the existing technology, a number of crack detection methods based on deep learning have been proposed, which can be mainly divided into two categories: one is based on two-stage object detection frameworks (such as Faster R-CNN), and the other is based on single-stage detectors (such as YOLO series) or semantic segmentation networks (such as U-Net, DeepLab).
[0004] These methods still face numerous technical bottlenecks in practical engineering applications. First, high-precision crack detection typically relies on complex deep neural network models. While achieving high detection accuracy, these models have a large number of parameters and high computational complexity, making them difficult to deploy on resource-constrained edge devices such as drones and inspection robots. This hinders real-time detection and severely limits their application in field or mobile inspection scenarios. Second, cracks exhibit significant multi-scale characteristics, ranging from micron-level initial cracks that are difficult to detect with the naked eye to structural cracks several centimeters wide. Traditional convolutional networks are prone to losing detailed information about small cracks during downsampling, resulting in a high rate of missed detections for tiny cracks. Furthermore, real-world detection environments contain numerous interfering factors. These background noises are visually highly similar to real cracks, making existing models prone to misjudgment in complex environments and lacking robustness. Moreover, most models do not fully consider multi-scale feature fusion mechanisms. The simple feature pyramid structure design makes it difficult to effectively integrate deep semantic information with shallow detailed features, thus limiting detection performance. Summary of the Invention
[0005] The purpose of this invention is to provide a lightweight, multi-scale crack detection method for structural monitoring, in order to overcome the aforementioned deficiencies in the prior art.
[0006] A lightweight multi-scale crack detection method for structural monitoring includes the following steps: S1. Obtain the original image data of the surface of the structure to be monitored through the image acquisition device, and preprocess the original image data to obtain a standardized input image; S2. Input the standardized input image into the preset lightweight multi-scale crack detection network model, and use the lightweight feature extraction backbone network in the model to extract features from the image to obtain multi-level feature maps. S3. The multi-scale feature fusion module in the model performs cross-level feature interaction and fusion on multi-level feature maps to generate a fused feature map containing rich contextual information. S4. Use the attention optimization module in the model to assign weights to the fused feature map, strengthen the key features of the crack and suppress background noise; S5. Input the optimized feature map into the detection head and output the crack segmentation mask or bounding box to complete crack detection and localization.
[0007] Preferably, the preprocessing of the original image data in step S1 specifically includes the following steps: S11. Obtain the original image of the surface of the structure to be monitored. , where (x,y) represents pixel coordinates; S12. Perform illumination equalization processing on the original image to eliminate the influence of uneven ambient lighting, and obtain an illumination-corrected image. ; S13. Linearly map the pixel values of the illumination-corrected image to the [0,1] interval to obtain a normalized image. The mapping formula is as follows: ; Represents the minimum value of an image pixel. Represents the maximum value of the image pixels; S14. Adjust the normalized image to a preset network input size H×W×3H to obtain a standardized input image. This is used in subsequent feature extraction steps.
[0008] Preferably, feature extraction in step S2 utilizes a lightweight feature extraction backbone network, specifically including the following steps: S21. Standardize the input image The input is fed into a lightweight convolutional neural network, where preliminary feature extraction is performed through stacked depthwise separable convolutional layers to obtain shallow feature maps. ; S22. A nonlinear transformation is performed on the shallow feature map using a linear bottleneck structure and an inverted residual structure to reduce the feature map dimension while expanding the number of channels, generating a deep semantic feature map: ; s represents the total downsampling factor of the network, and C represents the number of channels in the feature map; S23. Select the outputs of intermediate layers of different depths in the lightweight convolutional neural network to form a multi-level feature map set containing edge details and high-level semantic information. This is used for subsequent multi-scale feature fusion.
[0009] Preferably, in step S3, the multi-scale feature fusion module integrates the axial decomposition pooling (ADP) module, which compresses redundant information through axial dimension pooling to improve the efficiency of multi-scale feature fusion.
[0010] Preferably, step S3 involves cross-level feature interaction and fusion via a multi-scale feature fusion module, specifically including the following steps: S31. For multi-level feature map sets High-level semantic feature maps Perform upsampling to obtain a feature map with the same spatial resolution as the low-level feature map. The upsampling operation uses bilinear interpolation, as shown in the following formula: ; Upsampled feature map With the corresponding low-level feature map Perform channel-level concatenation to obtain the concatenated feature map. ; S32. Concatenate feature maps using a 1×1 convolutional layer. Perform channel dimensionality reduction and feature recalibration to generate a fused feature map. ; S33. Process the fused feature map using parallel dilated convolution branches, with the dilation rates set to [values to be specified in the original text]. This method captures crack context information under different receptive fields and performs a weighted summation of the outputs of each branch to obtain a final fused feature map containing rich multi-scale information. .
[0011] Preferably, the attention optimization module in step S4 is a height-width decoupled attention (HWDA) module, which achieves precise enhancement of key crack features by independently weighting the height and width dimensions of the feature map.
[0012] Preferably, step S4 involves weight allocation of the fused feature map using the attention optimization module, specifically including the following steps: S41. The feature map after multi-scale fusion The inputs are respectively sent to the channel attention submodule and the spatial attention submodule; S42. In the channel attention submodule, channel descriptors are obtained by performing global average pooling and max pooling operations on the feature map. and After passing through a shared multilayer perceptron network, the elements are added one by one, and channel attention weights are generated using a sigmoid activation function. The calculation formula is as follows: ; in, The sigmoid function is represented by MLP, and the multilayer perceptron is represented by MLP. S43. In the spatial attention submodule, the channel-weighted feature maps are subjected to max pooling and average pooling along the channel dimension, respectively. The two feature maps are then concatenated along the channel dimension, and spatial attention weights are generated through a 7×7 convolutional layer. The calculation formula is as follows: ; in, This represents the feature map after channel weighting. , For global pooling along the channel dimension; This represents a convolution operation with a kernel size of 7×7. Represents the Sigmoid function; This indicates that feature maps A and B are concatenated along the channel dimension.
[0013] S44. Adjust channel attention weights Spatial attention weights The process is applied sequentially to the fused feature map, enhancing the feature response of key crack regions, suppressing background noise interference, and outputting the optimized feature map Fatt.
[0014] Preferably, in step S5, the detection head can output two results: a pixel-level crack segmentation mask, used for accurate crack morphology depiction; or a target-level crack bounding box, used for rapid crack location.
[0015] Preferably, the lightweight multi-scale crack detection network is deployed on edge devices such as drones and inspection robots, and is suitable for automated monitoring of cracks in bridges, tunnels, dams, and road surfaces.
[0016] The beneficial effects achieved by this invention are as follows: 1. The method of this application adopts a lightweight convolutional neural network as the backbone of feature extraction and introduces depthwise separable convolution and linear bottleneck structure. While ensuring the feature expression capability, it significantly reduces the number of model parameters and computational overhead, enabling the model to achieve efficient inference on resource-constrained edge devices such as drones and inspection robots. It effectively overcomes the technical bottleneck of traditional high-precision models being difficult to deploy on mobile inspection platforms due to computational redundancy, and achieves synergistic optimization of detection efficiency and accuracy.
[0017] 2. This application designs a multi-scale feature fusion module, which expands the receptive field and fuses deep semantics and shallow detail features by combining upsampling and cross-level feature concatenation with a parallel dilated convolution structure. This significantly enhances the perception of cracks at different scales and improves the ability to identify early structural damage. Regarding complex background interference, this invention introduces a collaborative optimization mechanism of channel attention and spatial attention. By applying dual weighting to the feature map based on channel and spatial dimensions, it automatically focuses on key crack regions, effectively suppressing non-crack interference factors such as oil stains, shadows, and surface textures, significantly reducing the false alarm rate and improving the reliability and stability of the detection results. The entire detection process is end-to-end automated, requiring no manual intervention from raw image input to crack segmentation output. Combined with illumination correction and normalization in image preprocessing, it enhances the model's adaptability to different environmental conditions, making it suitable for all-weather, multi-scenario infrastructure inspection tasks.
[0018] 3. Furthermore, thanks to the synergistic effect of multi-level feature fusion and attention mechanisms, the model exhibits excellent detection performance and generalization ability across various structural types, including bridges, tunnels, and road surfaces. It maintains high segmentation accuracy in cross-scene testing, demonstrating broad application prospects. Through the organic integration of the above technical solutions, this invention effectively solves key problems in existing technologies such as poor real-time performance, insufficient sensitivity to small crack detection, sensitivity to background interference, and insufficient feature fusion. It provides an efficient, accurate, and practical intelligent detection method for structural health monitoring, possessing significant technological advancement and engineering application value. Attached Figure Description
[0019] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Figure 1 This is a block diagram of the overall structure of the present invention.
[0020] Figure 2 This is a schematic diagram of the High Width Decoupled Attention (HWDA) module used in this invention.
[0021] Figure 3 This is a schematic diagram of the Axial Decomposition Pooling (ADP) module used in this invention.
[0022] Figure 4This is a schematic diagram of the structure of the multi-feed entry control enhancement (MGE) module used in this invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments 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.
[0024] like Figure 1-4 As shown, the present invention provides A lightweight multi-scale crack detection method for structural monitoring includes the following steps: S1. Obtain the original image data of the surface of the structure to be monitored through the image acquisition device, and preprocess the original image data to obtain a standardized input image; Step S1 involves preprocessing the original image data, specifically including the following steps: S11. Obtain the original image of the surface of the structure to be monitored. , where (x,y) represents pixel coordinates; S12. Perform illumination equalization processing on the original image to eliminate the influence of uneven ambient lighting, and obtain an illumination-corrected image. ; S13. Linearly map the pixel values of the illumination-corrected image to the [0,1] interval to obtain a normalized image. The mapping formula is as follows: ; Represents the minimum value of an image pixel. Represents the maximum value of the image pixels; S14. Adjust the normalized image to a preset network input size H×W×3H to obtain a standardized input image. This is used in subsequent feature extraction steps.
[0025] S2. Input the standardized input image into the preset lightweight multi-scale crack detection network model, and use the lightweight feature extraction backbone network in the model to extract features from the image to obtain multi-level feature maps. The S2 process utilizes a lightweight feature extraction backbone network for feature extraction, specifically including the following steps: S21. Standardize the input image The input is fed into a lightweight convolutional neural network, where preliminary feature extraction is performed through stacked depthwise separable convolutional layers to obtain shallow feature maps. ; S22. A nonlinear transformation is performed on the shallow feature map using a linear bottleneck structure and an inverted residual structure to reduce the feature map dimension while expanding the number of channels, generating a deep semantic feature map: ; s represents the total downsampling factor of the network, and C represents the number of channels in the feature map; S23. Select the outputs of intermediate layers of different depths in the lightweight convolutional neural network to form a multi-level feature map set containing edge details and high-level semantic information. This is used for subsequent multi-scale feature fusion.
[0026] S3. The multi-scale feature fusion module in the model performs cross-level feature interaction and fusion on multi-level feature maps to generate a fused feature map containing rich contextual information. In step S3, the multi-scale feature fusion module integrates the axial decomposition pooling (ADP) module, which compresses redundant information through axial dimension pooling to improve the efficiency of multi-scale feature fusion.
[0027] 5. A lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: step S3 involves cross-level feature interaction and fusion through a multi-scale feature fusion module, specifically including the following steps: S31. For multi-level feature map sets High-level semantic feature maps Perform upsampling to obtain a feature map with the same spatial resolution as the low-level feature map. The upsampling operation uses bilinear interpolation, as shown in the following formula: ; Upsampled feature map With the corresponding low-level feature map Perform channel-level concatenation to obtain the concatenated feature map. ; S32. Concatenate feature maps using a 1×1 convolutional layer. Perform channel dimensionality reduction and feature recalibration to generate a fused feature map. ; S33. Process the fused feature map using parallel dilated convolution branches, with the dilation rates set to [values to be specified in the original text]. This method captures crack context information under different receptive fields and performs a weighted summation of the outputs of each branch to obtain a final fused feature map containing rich multi-scale information. .
[0028] S4. The attention optimization module in the model is used to assign weights to the fused feature map, which strengthens the key features of the crack and suppresses background noise. The attention optimization module is the height-width decoupled attention (HWDA) module. By independently weighting the height and width dimensions of the feature map, the key features of the crack are accurately strengthened.
[0029] Step S4 involves using an attention optimization module to assign weights to the fused feature map, specifically including the following steps: S41. The feature map after multi-scale fusion The inputs are respectively sent to the channel attention submodule and the spatial attention submodule; S42. In the channel attention submodule, channel descriptors are obtained by performing global average pooling and max pooling operations on the feature map. and After passing through a shared multilayer perceptron network, the elements are added one by one, and channel attention weights are generated using a sigmoid activation function. The calculation formula is as follows: ; in, The sigmoid function is represented by MLP, and the multilayer perceptron is represented by MLP. S43. In the spatial attention submodule, the channel-weighted feature maps are subjected to max pooling and average pooling along the channel dimension, respectively. The two feature maps are then concatenated along the channel dimension, and spatial attention weights are generated through a 7×7 convolutional layer. The calculation formula is as follows: ; in, This represents the feature map after channel weighting. , For global pooling along the channel dimension; This represents a convolution operation with a kernel size of 7×7. Represents the Sigmoid function; This indicates that feature maps A and B are concatenated along the channel dimension.
[0030] S44. Adjust channel attention weights Spatial attention weights The process is applied sequentially to the fused feature map, enhancing the feature response of key crack regions, suppressing background noise interference, and outputting the optimized feature map Fatt.
[0031] S5. Input the optimized feature map into the detection head. The detection head can output two results: pixel-level crack segmentation mask, used for accurate crack morphology characterization; or target-level crack bounding box, used for rapid crack location.
[0032] The lightweight multi-scale crack detection network is deployed on edge devices such as drones and inspection robots, and is suitable for automated crack monitoring scenarios in bridges, tunnels, dams, and road surfaces.
[0033] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0034] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.
[0035] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0036] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.
Claims
1. A lightweight multi-scale crack detection method for structural monitoring, characterized in that: Includes the following steps: S1. Obtain the original image data of the surface of the structure to be monitored through the image acquisition device, and preprocess the original image data to obtain a standardized input image; S2. Input the standardized input image into the preset lightweight multi-scale crack detection network model, and use the lightweight feature extraction backbone network in the model to extract features from the image to obtain multi-level feature maps. S3. The multi-scale feature fusion module in the model performs cross-level feature interaction and fusion on multi-level feature maps to generate a fused feature map containing rich contextual information. S4. Use the attention optimization module in the model to assign weights to the fused feature map, strengthen the key features of the crack and suppress background noise; S5. Input the optimized feature map into the detection head and output the crack segmentation mask or bounding box to complete crack detection and localization.
2. The lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: Step S1 involves preprocessing the original image data, specifically including the following steps: S11. Obtain the original image of the surface of the structure to be monitored. , where (x,y) represents pixel coordinates; S12. Perform illumination equalization processing on the original image to eliminate the influence of uneven ambient lighting, and obtain an illumination-corrected image. ; S13. Linearly map the pixel values of the illumination-corrected image to the [0,1] interval to obtain a normalized image. The mapping formula is as follows: ； Represents the minimum value of an image pixel. Represents the maximum value of the image pixels; S14. Adjust the normalized image to a preset network input size H×W×3H to obtain a standardized input image. This is used in subsequent feature extraction steps.
3. The lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: The S2 process utilizes a lightweight feature extraction backbone network for feature extraction, specifically including the following steps: S21. Standardize the input image The input is fed into a lightweight convolutional neural network, where preliminary feature extraction is performed through stacked depthwise separable convolutional layers to obtain shallow feature maps. ; S22. A nonlinear transformation is performed on the shallow feature map using a linear bottleneck structure and an inverted residual structure to reduce the feature map dimension while expanding the number of channels, generating a deep semantic feature map: ； s represents the total downsampling factor of the network, and C represents the number of channels in the feature map; S23. Select the outputs of intermediate layers of different depths in the lightweight convolutional neural network to form a multi-level feature map set containing edge details and high-level semantic information. This is used for subsequent multi-scale feature fusion.
4. A lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: In step S3, the multi-scale feature fusion module integrates the axial decomposition pooling (ADP) module, which compresses redundant information through axial dimension pooling, thereby improving the efficiency of multi-scale feature fusion.
5. A lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: Step S3 involves cross-level feature interaction and fusion through a multi-scale feature fusion module, specifically including the following steps: S31. For multi-level feature map sets High-level semantic feature maps Perform upsampling to obtain a feature map with the same spatial resolution as the low-level feature map. The upsampling operation uses bilinear interpolation, as shown in the following formula: ； Upsampled feature map With the corresponding low-level feature map Perform channel-level concatenation to obtain the concatenated feature map. ; S32. Concatenate feature maps using a 1×1 convolutional layer. Perform channel dimensionality reduction and feature recalibration to generate a fused feature map. ; S33. Process the fused feature map using parallel dilated convolution branches, with the dilation rates set to [values to be specified in the original text]. This method captures crack context information under different receptive fields and performs a weighted summation of the outputs of each branch to obtain a final fused feature map containing rich multi-scale information. .
6. A lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: In step S4, the attention optimization module is a height-width decoupled attention (HWDA) module, which achieves precise enhancement of key crack features by independently weighting the height and width dimensions of the feature map.
7. A lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: Step S4 involves using an attention optimization module to assign weights to the fused feature map, specifically including the following steps: S41. The feature map after multi-scale fusion The inputs are respectively sent to the channel attention submodule and the spatial attention submodule; S42. In the channel attention submodule, channel descriptors are obtained by performing global average pooling and max pooling operations on the feature map. and After passing through a shared multilayer perceptron network, the elements are added one by one, and channel attention weights are generated using a sigmoid activation function. The calculation formula is as follows: ； in, The sigmoid function is represented by MLP, and the multilayer perceptron is represented by MLP. S43. In the spatial attention submodule, the channel-weighted feature maps are subjected to max pooling and average pooling along the channel dimension, respectively. The two feature maps are then concatenated along the channel dimension, and spatial attention weights are generated through a 7×7 convolutional layer. The calculation formula is as follows: ； in, This represents the feature map after channel weighting. , For global pooling along the channel dimension; This indicates a convolution operation with a kernel size of 7×7. Represents the Sigmoid function; This indicates that feature maps A and B are concatenated along the channel dimension. S44. Adjust channel attention weights Spatial attention weights The process is applied sequentially to the fused feature map, enhancing the feature response of key crack regions, suppressing background noise interference, and outputting the optimized feature map Fatt.
8. A lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: In step S5, the detection head can output two results: a pixel-level crack segmentation mask, used for accurate crack morphology depiction; or a target-level crack bounding box, used for rapid crack location.
9. A lightweight multi-scale crack detection method for structural monitoring according to claim 1, characterized in that: The lightweight multi-scale crack detection network is deployed on edge devices such as drones and inspection robots, and is suitable for automated crack monitoring scenarios in bridges, tunnels, dams, and road surfaces.