Recognition and research method for apparent damage of lightweight concrete bridge
By constructing a lightweight segmentation network model and using morphological methods, the problem of low accuracy in bridge damage identification was solved, achieving efficient and accurate bridge damage detection.
Patent Information
- Application Number
- CN202511036244.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-26
- Publication Date
- 2025-11-11
AI Technical Summary
Existing technologies have low accuracy in identifying bridge damage, and traditional machine learning algorithms cannot effectively handle complex damage identification tasks.
A lightweight segmentation network model is constructed, using MobileNetV2 as the backbone network. The ASPP module and ECA attention mechanism are combined for feature extraction and fusion, and morphological methods are used to extract damage feature parameters.
It significantly improves the accuracy and computational efficiency of damage segmentation, achieves high-precision quantification of bridge damage, and reduces computational costs and model complexity.
Smart Images
Figure CN120932068A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bridge damage detection technology, specifically to a research method for identifying apparent damage in lightweight concrete bridges. Background Technology
[0002] As critical transportation hubs, bridge engineering necessitates regular safety inspections and assessments. Manual visual inspection is costly and inefficient, and its quality and accuracy are easily affected by the inspector's experience and subjective judgment. Early traditional machine learning algorithms such as artificial neural networks, support vector machines, and Naive Bayes were applied to structural damage identification and detection; however, these shallow models cannot effectively handle complex damage identification tasks, resulting in low accuracy in bridge damage identification results. Summary of the Invention
[0003] To address the issue of low accuracy in existing methods for identifying bridge damage, this invention aims to provide a method for identifying apparent damage in lightweight concrete bridges. The specific technical solution adopted is as follows:
[0004] This invention provides a method for identifying apparent damage in lightweight concrete bridges, the method comprising the following steps:
[0005] A lightweight segmentation network model is constructed, which includes an encoder and a decoder;
[0006] The bridge image is input into a lightweight segmentation network model, where the encoder extracts features from the input bridge image to generate shallow and deep features.
[0007] Deep features are input to the ASPP module; the ASPP module contains five parallel branches, each of which is connected to an ECA attention mechanism to fuse multi-scale features.
[0008] The decoder concatenates the shallow features with the deep features output by the ASPP module, upsamples them to restore the original image size, and outputs the segmented damaged image.
[0009] Morphological methods are used to process the segmented damage images and extract the feature parameters of the damage.
[0010] Preferably, the encoder uses MobileNetV2 as the backbone network. The backbone network adopts an inverted residual structure and a linear bottleneck design. Multi-layer bottleneck modules are constructed by using an expansion factor t, the number of output channels c, the number of repetitions n, and the step size s to extract features.
[0011] Preferably, the five parallel branches are 1×1 convolution, dilated convolution with dilation rates of 1, 3, and 7, and global average pooling, respectively.
[0012] Preferably, after the deep features are input to the ASPP module, the module further includes:
[0013] Perform global average pooling on deep features;
[0014] Attention weights are assigned to each channel using a one-dimensional convolutional kernel; the kernel size is determined adaptively. Preferably, the morphological method is used to process the damage segmentation results and extract feature parameters of the damage, including:
[0015] Opening and closing operations are performed on the segmented damaged image to extract the crack region. The crack region is then skeletonized to convert the crack into a single-pixel wide skeleton. The maximum length and maximum width of the crack are extracted based on the skeleton.
[0016] Calculate the damaged area and coverage based on the pixel values of the damaged region;
[0017] The characteristic parameters include the maximum crack length, maximum width, damaged area, and coverage.
[0018] The present invention has at least the following beneficial effects:
[0019] This invention constructs a lightweight segmentation network model, achieving end-to-end optimization from pixel-level segmentation to damage quantification. Combined with the ASPP module, it significantly reduces the number of network model parameters and computational cost, substantially improving computational efficiency. By capturing local to global features through multi-scale dilated convolutions and combining this with dynamic attention calibration, the damage segmentation accuracy is significantly superior to traditional models. Furthermore, morphological methods are incorporated to accurately quantify the feature parameters of bridge damage. The method provided by this invention overcomes the limitations of traditional models in terms of computational efficiency, small target detection accuracy, and damage quantification capabilities, offering a solution for automated bridge inspection that balances high accuracy with engineering practicality. Attached Figure Description
[0020] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 A flowchart illustrating a method for identifying apparent damage in lightweight concrete bridges provided in an embodiment of the present invention;
[0022] Figure 2 This is a diagram showing the results of morphological operations.
[0023] Figure 3A diagram illustrating the skeletonized structure of the crack;
[0024] Figure 4 This is an original image of a crack in a specific area of the bridge.
[0025] Figure 5 for Figure 4 Corresponding label image;
[0026] Figure 6 for Figure 4 Corresponding damage segmentation effect diagram;
[0027] Figure 7 This is an original image of steel reinforcement corrosion in a specific area of the bridge.
[0028] Figure 8 for Figure 7 Corresponding label image;
[0029] Figure 9 for Figure 7 Corresponding damage segmentation effect diagram;
[0030] Figure 10 This is a stripped-out original image of a section of the bridge.
[0031] Figure 11 for Figure 10 Corresponding label image;
[0032] Figure 12 for Figure 10 The corresponding damage segmentation effect diagram. Detailed Implementation
[0033] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of a method for identifying apparent damage in lightweight concrete bridges based on the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.
[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0035] The following description, in conjunction with the accompanying drawings, details the specific scheme of the research method for identifying apparent damage in lightweight concrete bridges provided by this invention.
[0036] Example of a research method for identifying apparent damage in lightweight concrete bridges:
[0037] This embodiment proposes a method for identifying apparent damage in lightweight concrete bridges, such as... Figure 1 As shown, the method for identifying apparent damage in lightweight concrete bridges according to this embodiment includes the following steps:
[0038] Step S1: Construct a lightweight segmentation network model, which includes an encoder and a decoder.
[0039] First, a lightweight segmentation network model based on ECAS-Deeplabv3+ is constructed to learn the apparent damage feature representation of bridges. An encoder-decoder architecture is adopted, with the encoder using MobileNetV2 as the backbone network and the decoder including feature concatenation and upsampling modules.
[0040] The lightweight segmentation network model specifically includes:
[0041] Image data partitioning: In this embodiment, 1550 bridge images are obtained from the image database. These images include undamaged bridge images and damaged bridge images. The bridge images are partitioned into training set, validation set and test set according to the proportions of 80%, 20% and 10%, respectively.
[0042] Data augmentation: Randomly flip, rotate, and crop to 512×512 resolution.
[0043] Parameter settings: Model selection: SGD optimizer, initial learning rate: 0.0005, training batch size: 8, training iterations: 200, momentum parameter: 0.9, weight decay coefficient: 0.0001.
[0044] Batch size: 8, number of iterations: 200, loss function: cross-entropy loss.
[0045] The lightweight segmentation network model is trained, validated, and tested to obtain a well-trained lightweight segmentation network model. The training process of the network model is based on existing technology and will not be described in detail here.
[0046] To verify the effectiveness of each improved part of the model, comparative experiments were conducted. The experiments were divided into six groups, targeting the backbone network, ASPP module, and ECA attention mechanism. The experimental groups are shown in Table 1. "√" indicates that the algorithm was used in the experiment, and "-" indicates that the algorithm was not used. The experimental environment and parameter settings were kept consistent. Five evaluation metrics—mIoU, mPA, F1-score, Params, and FLOPs—were used to evaluate and analyze each experimental group. The results for each group are shown in Table 2.
[0047] Table 1 Comparative Experiment Groups
[0048]
[0049]
[0050] Table 2 Comparative test results
[0051]
[0052] Based on the experimental data in the table, groups A, B, and C all used Xception as the backbone network, while groups D, E, and F all used MobilnetV2. The models in groups D, E, and F significantly outperformed groups A, B, and C in identifying and segmenting apparent damage to concrete bridges, and their parameter count and computational cost were also much lower. This demonstrates that MobilnetV2, when used as the backbone network, is more efficient than Xception in identifying apparent damage to concrete bridges, providing higher recognition accuracy at a lower computational cost.
[0053] Comparing the results of groups A and B with groups D and E, it is evident that using dilated convolutions with dilation rates of 1, 3, and 7 reduces the model's recognition accuracy, but significantly decreases its complexity and computational cost. Under the Xception architecture, the ASPP module with dilation rates of 1, 3, and 7 reduces the number of parameters by 5.18M; while under the MobilnetV2 architecture, the ASPP module with dilation rates of 1, 3, and 7 reduces the number of parameters by 14.88M. Therefore, ASPP modules with smaller dilation rates can effectively reduce the number of parameters and computational cost.
[0054] Comparing the results of groups B and C with those of groups E and F, it is evident that adding the ECA attention mechanism significantly improves the accuracy of identifying apparent damage in concrete bridges. Under the Xception architecture, the ECA attention mechanism improves mIoU by 2.81%, while under the MobilnetV2 architecture, it improves mIoU by 3.07% and F1-score by 0.54%. The ECA attention mechanism optimizes feature channel relationships, enabling MobileNetV2 to extract key features more fully, thus compensating for MobileNetV2's shortcomings in local feature representation. The ECA attention mechanism effectively improves the model's recognition accuracy, especially under the MobilnetV2 architecture, where the improvement is more significant.
[0055] Group F represents the improved lightweight damage recognition model ECAS-Deeplabv3+ proposed in this embodiment. Using MobileNetV2 as the backbone network, the addition of the ECA attention mechanism significantly enhances the model's damage recognition and segmentation capabilities. Furthermore, the use of the optimized feature extraction ASPP module with a low inflation rate reduces the model's computational load and lowers computational costs. This model achieves high recognition and segmentation accuracy while maintaining a low number of parameters and computational cost.
[0056] Step S2: Input the bridge image into the lightweight segmentation network model. The encoder extracts features from the input bridge image to generate shallow and deep features.
[0057] The encoder of the lightweight segmentation network model is used for feature extraction. Specifically, the bridge image is input into the lightweight segmentation network model, and shallow features f1 (low-level details, such as texture) and deep features f2 (high-level semantics, such as damage type) are extracted by MobileNetV2.
[0058] Step S3: Deep features are input into the ASPP module; the ASPP module contains five parallel branches, each of which is connected to an ECA attention mechanism for fusing multi-scale features.
[0059] The ASPP module in this embodiment contains five parallel branches, each of which is connected to an ECA attention mechanism. These branches are 1×1 convolution, dilated convolution with dilation rates of 1, 3, and 7, and global average pooling. Each branch is connected to an ECA attention mechanism to achieve multi-scale feature fusion.
[0060] The deep feature f2 is input into the ASPP module, and local, intermediate and global features are captured by dilated convolutions with different dilation rates (1, 3, 7), combined with ECA attention to suppress channel redundancy.
[0061] In this embodiment, the core design of the ECA attention mechanism includes global average pooling and one-dimensional convolution operations. First, global average pooling is performed on the input deep features to compress the spatial dimension and capture channel information. The compressed spatial dimension is 1×1×C, where C represents the number of channels in the input feature map. Then, an adaptively selected one-dimensional convolution kernel assigns different attention weights to each channel, automatically identifying and focusing on important channel features while suppressing irrelevant channel information. The convolution kernel size is determined using an adaptive method to achieve dynamic calibration of channel features.
[0062] In step S4, the decoder concatenates the shallow features with the deep features output by the ASPP module, upsamples them to restore the original image size, and outputs the segmented damaged image.
[0063] Next, the decoder concatenates the shallow feature f1 with the features output by the ASPP module after ECA attention, generates a damage segmentation map by four times upsampling, marks the pixels of the damage area, and obtains the segmented damage image.
[0064] Step S5: Process the segmented damage image based on morphological methods to extract the feature parameters of the damage.
[0065] After obtaining the segmented damage image, the feature parameters of the damage will be extracted by combining the segmented damage image.
[0066] Specifically, the segmented damage image is processed by opening operations (erosion followed by dilation) to eliminate noise, and then closing operations (dilation followed by erosion) are performed to connect the fractured damage regions, thus skeletonizing the cracks. In other words, a morphological method is used to process the segmented damage image. The morphological operation results are shown in the figure below. Figure 2 As shown in the figure, the crack skeletonization result is as follows. Figure 3 As shown. The crack is converted into a single-pixel skeleton, and the maximum crack length (number of skeleton pixels) and maximum width (vertical skeleton boundary spacing) are calculated. For non-crack damage, the damage area and coverage are calculated using the following formula:
[0067]
[0068] S=∑f 2 (x,y)dS (3);
[0069]
[0070] In the formula, h represents the image height, w represents the image width, S represents the damaged area, CR represents the coverage, L represents the maximum crack length, and W... max Let f(x,y) represent the maximum width, f(x,y) represent the displacement of pixels in the calibration detection image, and d represent the displacement of pixels in the calibration detection image. L (x,y) represents the distance from the skeleton point (x,y) to the left (or vertically) to the crack boundary, d R (x,y) represents the distance from the skeleton point (x,y) to the right (or vertically) to the crack boundary, h represents the length of the image, and w represents the width of the image.
[0071] Thus, the method provided in this embodiment has been used to accurately identify bridge damage.
[0072] like Figures 4-12 As shown, where, Figure 4 , Figure 5 , Figure 6 For a set of images, Figure 7 , Figure 8 , Figure 9 For a set of images, Figure 10 , Figure 11 , Figure 12 Each image is a set of images. The first image in each set is the original image that was acquired, the second image is the corresponding label image, and the third image is the corresponding damage segmentation result.
[0073] To verify the effectiveness of the quantization method presented in this paper, the quantization results of 150 damage images in the test set were analyzed, including 50 images each of cracks, spalling, and rebar corrosion. The results are as follows: Figures 5-8As shown. If the predicted value is equal to the true value, the point representing the predicted size will fall on the baseline y = x. If the marker point falls above the baseline, it indicates that the predicted value is greater than the true value; if the marker point falls below the baseline, it indicates that the predicted value is less than the true value. From Figures 5-8 As can be seen, the predicted points for the maximum crack length, maximum width, spalling damage area, and steel reinforcement corrosion area all fall near the baseline, indicating that the method provided in this embodiment can accurately quantify the characteristic parameters of various types of bridge damage.
[0074] The results obtained using the method provided in this embodiment show that the length and width of cracks, the area of spalling damage, and the coverage of rebar corrosion are all highly consistent with the actual results, further proving the accuracy of the quantification method. In all the damage prediction data, the mean absolute error (MAE) for the maximum crack length prediction was 10.22 pixels, and the mean relative error (MRE) was 2.25%; the MAE for the maximum width prediction was 2.82 pixels, and the MRE was 20.78%; the MAE for the spalling damage area prediction was 767.4 pixels, and the MRE was 5.47%; the MAE for the coverage was 0.34%, and the MRE was 4.7%; the MAE for the rebar corrosion prediction area was 217.88 pixels, and the MRE was 2.65%; the MAE for the coverage was 9.73%, and the MRE was 2.65%.
[0075] This embodiment constructs a lightweight segmentation network model, achieving end-to-end optimization from pixel-level segmentation to damage quantification. Combined with the ASPP module, the number of network model parameters and computational cost are significantly reduced, resulting in a substantial improvement in computational efficiency. Multi-scale dilated convolution captures local to global features, and dynamic attention calibration further enhances damage segmentation accuracy compared to traditional models. Furthermore, morphological methods are incorporated to accurately quantify the feature parameters of bridge damage. The method provided in this embodiment overcomes the limitations of traditional models in terms of computational efficiency, small target detection accuracy, and damage quantification capabilities, offering a solution for automated bridge inspection that balances high accuracy with engineering practicality.
[0076] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for identifying apparent damage in lightweight concrete bridges, characterized in that, The method includes the following steps: A lightweight segmentation network model is constructed, which includes an encoder and a decoder; The bridge image is input into a lightweight segmentation network model, where the encoder extracts features from the input bridge image to generate shallow and deep features. Deep features are input to the ASPP module; the ASPP module contains five parallel branches, each of which is connected to an ECA attention mechanism to fuse multi-scale features. The decoder concatenates the shallow features with the deep features output by the ASPP module, upsamples them to restore the original image size, and outputs the segmented damaged image. Morphological methods are used to process the segmented damage images and extract the feature parameters of the damage.
2. The method for identifying apparent damage in lightweight concrete bridges according to claim 1, characterized in that, The encoder uses MobileNetV2 as the backbone network. The backbone network adopts an inverted residual structure and a linear bottleneck design. Multi-layer bottleneck modules are constructed by using the expansion factor t, the number of output channels c, the number of repetitions n, and the step size s to extract features.
3. The method for identifying apparent damage in lightweight concrete bridges according to claim 1, characterized in that, The five parallel branches are 1×1 convolution, dilated convolution with dilation rates of 1, 3, and 7, and global average pooling.
4. The method for identifying apparent damage in lightweight concrete bridges according to claim 1, characterized in that, After the deep features are input into the ASPP module, the following are also included: Perform global average pooling on deep features; Attention weights are assigned to each channel using a one-dimensional convolutional kernel; the kernel size is determined using an adaptive method.
5. The method for identifying apparent damage in lightweight concrete bridges according to claim 1, characterized in that, The morphological method is used to process the damage segmentation results and extract the feature parameters of the damage, including: Opening and closing operations are performed on the segmented damaged image to extract the crack region. The crack region is then skeletonized to convert the crack into a single-pixel wide skeleton. The maximum length and maximum width of the crack are extracted based on the skeleton. Calculate the damaged area and coverage based on the pixel values of the damaged region; The characteristic parameters include the maximum crack length, maximum width, damaged area, and coverage.