Railway steel truss bridge coating corrosion detection and evaluation method and system

Abstract

The invention provides a railway steel truss bridge coating corrosion detection and evaluation method and system, and belongs to the technical field of target recognition, and the method comprises the steps: obtaining a to-be-processed railway steel truss bridge image; processing the obtained to-be-processed railway steel truss bridge image by using a pre-trained target divider model, predicting to obtain a gusset plate corrosion part in the gusset plate image in the railway steel truss bridge, and calculating a gusset plate and a pixel number corresponding to gusset plate corrosion; and processing the gusset plate image and the pixel number corresponding to gusset plate corrosion to obtain the gusset plate corrosion degree state. According to the method, gusset plate area positioning and gusset plate corrosion segmentation are realized, the gusset plate corrosion degree and the corrosion grade are obtained, and the interpretability and the practicability of a detection result are effectively improved; through joint enhancement of pixel-level features and a spatial distribution structure, it is ensured that enhanced data has high availability on the basis of maintaining original semantics, and therefore the model training bottleneck caused by insufficient samples is effectively relieved.

Description

Technical Field

[0001] This invention relates to the field of target recognition technology, specifically to a method and system for detecting and evaluating corrosion of coatings on railway steel truss bridges. Background Technology

[0002] Steel truss bridges, as crucial structural facilities in railway sections crossing rivers, are prone to corrosion due to long-term use and rainwater erosion, especially in the node plates of connecting components. This corrosion accelerates the aging and damage of the steel structure, and in severe cases, may endanger train operation safety. Furthermore, constructing or replacing corroded components is not only complex but also costly; therefore, regular inspections of the node plates of steel truss bridges are particularly important.

[0003] Although existing models have achieved encouraging results in segmentation accuracy, they still require a large amount of computational resources, which limits their application in real-time scenarios. For the detection of nodal plate corrosion in bridge remote sensing images acquired by UAVs, the following challenges still exist: (1) Images taken by UAVs often contain irrelevant areas such as the sky, vegetation, and non-target bridge components. These areas introduce a lot of noise, distracting the model from the nodal plate, reducing segmentation accuracy, and easily leading to false detection when irrelevant areas and target areas have similar visual features. (2) Nodal plates usually lack obvious features or textures, resulting in limited semantic information acquired by the model during the learning process. This situation easily leads to repetitive and inefficient gradient updates during training, thus slowing down the model's convergence speed. At the same time, the subtle differences between the nodal plate and the surrounding environment also increase the difficulty of the model to establish clear and robust feature representations. (3) The corrosion on the nodal plate varies greatly in shape, size, and severity, which poses a challenge to the model's generalization ability. In addition, due to the scarcity of corrosion samples and the high labor intensity of the data annotation process, it is difficult to obtain a sufficient number of defective samples for training. This scarcity of samples limits the effectiveness of the model in detecting new or complex corrosion.

[0004] Semantic segmentation algorithms aim to achieve pixel-level image classification, focusing on a balance between high accuracy and real-time performance. Early semantic segmentation methods were primarily based on encoder-decoder structures. In these methods, the encoder expands the receptive field through operations such as convolution and pooling, while the decoder recovers detailed information from high-level semantic features. However, downsampling often leads to the loss of spatial details, limiting the ability to capture fine-grained information. The DeepLab series introduced dilated convolutions with different dilation rates, enhancing multi-scale feature extraction and effectively mitigating the resolution degradation problem. Although this method improves segmentation accuracy, it is prone to raster artifacts and has high computational complexity, making it unsuitable for real-time applications. SwiftNet captures high-level semantic information by using low-resolution input while preserving spatial details with high-resolution input, enabling its lightweight decoder to achieve efficient and accurate segmentation. Similarly, DFANet introduces a lightweight backbone network into the Xception architecture based on depthwise separable convolutions, reducing input size and inference time while maintaining competitive segmentation performance. However, encoder-decoder based networks inherently suffer from latency due to their multi-stage encoding and decoding processes. To overcome these limitations, dual-branch networks, such as BiSeNet and DDRNet, have emerged. These architectures employ one branch for contextual feature modeling and another for detail features, efficiently integrating the information from both branches through a fusion module. PIDNet further proposes a three-branch structure, focusing on detail, context, and boundary features respectively, achieving efficient multi-scale context aggregation and feature fusion through a parallel aggregation pyramid pooling module (PAPPM) and a boundary attention-guided fusion module (Bag). Despite these advancements, balancing feature representation capabilities with branch interaction efficiency remains a challenge for multi-branch networks, especially in small datasets or complex scenarios, where the model's generalization ability is easily affected by the limited number of training samples.

[0005] Surface defect detection is crucial for ensuring bridge safety and extending their service life, as failure to detect or properly address defects in a timely manner can lead to structural failure and costly repairs. Among numerous methods, computer vision-based detection techniques have attracted widespread attention due to their advantages such as low cost, ease of implementation, and non-contact operation. These methods make defect detection for large-scale infrastructure efficient and scalable. For example, Dung et al. proposed a concrete crack detection method based on a fully convolutional neural network (FCN), demonstrating good performance in crack identification. However, this method relies on pixel-level classification, and achieving automated quantitative analysis in complex backgrounds or scenarios with overlapping defects remains a challenge. Similarly, Hoang et al. developed a concrete spalling detection method that combines a Gabor filter for texture extraction with a logistic regression model optimized based on adaptive moment estimation. While this method is effective in detecting large-area spalling, it has limitations in identifying small defects or subtle surface changes, which are often early signs of major structural problems. Khayatazad et al. focused on steel structure corrosion detection, accurately locating corrosion areas through roughness and color features. However, this method is highly sensitive to changes in lighting conditions and performs poorly in images containing impurities or overlapping structures. While the aforementioned methods have made significant progress in detecting general defects in bridges, research on gusset plates remains limited. Existing studies largely focus on stress distribution analysis, corrosion repair, and buckling behavior of gusset plates, with a lack of automated methods for gusset plate corrosion detection and quantitative analysis. This gap is particularly concerning because gusset plates play a crucial role in the structural stability of steel truss bridges. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for detecting and evaluating corrosion of coatings on railway steel truss bridges, so as to solve at least one of the technical problems existing in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for detecting and evaluating corrosion of coatings on railway steel truss bridges, comprising:

[0009] Acquire images of the railway steel truss bridge to be processed;

[0010] The acquired railway steel truss bridge image is processed using a pre-trained target segmenter model to predict the node plate structure in the railway steel truss bridge; the predicted node plate is cropped to obtain the node plate image; the node plate image is then input into the target segmenter model for further processing to predict the rusted parts in the node plate, and the number of pixels corresponding to the node plate and the rusted node plate is calculated.

[0011] The corrosion degree evaluator is used to process the node plate image and the number of pixels corresponding to the corrosion of the node plate to obtain the corrosion degree status of the node plate.

[0012] As a further limitation of the first aspect of the present invention, the target segmenter model includes: a feature extraction unit and a semantic segmentation unit; the feature extraction unit is used to extract features of railway steel truss bridge node plates and node plate corrosion; the feature extraction unit uses detail branches, context branches, and boundary branches, the detail branches are used to retain large-scale feature maps and capture fine-grained details in the image; the context branches extract local and global features through continuous downsampling operations, improve the ability to model context information, and enhance the network's understanding of complex scenes; the boundary branches extract high-frequency features, predict boundary regions, and simultaneously perform upsampling operations on the extraction results of the detail branches, context branches, and boundary branches; the semantic segmentation unit fuses the features extracted by the detail branches, context branches, and boundary branches, and uses a convolutional structure as the segmentation head to achieve pixel-level classification of the image.

[0013] As a further limitation of the first aspect of the present invention, the detail branch successively uses the residual module, the separable ECA residual module and the residual bottleneck module; the context branch and the detail branch use the same module, extracting local and global features through continuous downsampling operations, and at the same time introducing an improved parallel channel attention pyramid pooling module to improve the ability to model context information and enhance the network's ability to understand complex scenes.

[0014] As a further definition of the first aspect of the invention, the boundary branch predicts the boundary region through a residual module and a deformable residual bottleneck module. Specifically, the separable ECA residual module replaces the second layer with a depthwise separable convolution, decomposing the standard convolution into depthwise and pointwise convolutions to reduce computational complexity. It also embeds an efficient channel attention module to adaptively allocate channel weights in a lighter manner. The deformable residual bottleneck module introduces a deformable convolution module onto the classic residual bottleneck module. The 3×3 convolution in the classic residual bottleneck module is replaced by a deformable convolution, whose offset is predicted by an independent 3×3 convolutional layer and works in conjunction with the main convolutional layer to complete deformable feature sampling and processing. Simultaneously, the deformable residual bottleneck module retains the efficient bottleneck structure of the classic residual bottleneck module, achieving feature compression and expansion through two 1×1 convolutional layers.

[0015] As a further limitation of the first aspect of the present invention, the parallel channel attention pyramid pooling module consists of multi-scale pooling branches, channel attention mechanisms, and aggregation operations. It extracts multi-scale features through three branches with different downsampling ratios, and further optimizes feature processing efficiency by using grouped convolution in the fusion stage. Finally, by introducing a lightweight channel attention mechanism, the correlation between channels is enhanced after compressing channel information, thereby effectively reducing the adverse effects that grouped convolution may cause.

[0016] As a further limitation of the first aspect of the present invention, the loss function used for training the target segmenter model is as follows: In the detail branch, a loss L0 is calculated using a cross-entropy loss function on the intermediate feature layer through an additional semantic head to supervise the learning of fine-grained features, thereby improving the network's ability to capture complex details in the image; in the boundary branch, a weighted binary cross-entropy loss function is used to calculate the loss L1 to solve the class imbalance problem caused by sparse boundary pixels; the semantic context branch uses the standard cross-entropy loss L2 to supervise the semantic segmentation results. This loss focuses on global context information and guides the network to better understand complex regions by integrating multi-scale features, thereby improving the segmentation accuracy of large-scale regions; another cross-entropy loss L3 is used as a coordination loss to jointly optimize the semantic segmentation and boundary detection tasks, thereby enhancing the synergy between the context branch and the boundary branch.

[0017] Secondly, the present invention provides a corrosion detection and evaluation system for railway steel truss bridge coatings, comprising:

[0018] The acquisition module is used to acquire images of railway steel truss bridges to be processed;

[0019] The prediction module is used to process the acquired railway steel truss bridge image to be processed using a pre-trained target segmenter model to predict the node plate structure in the railway steel truss bridge; the predicted node plate is cropped to obtain the node plate image; the node plate image is then input into the target segmenter model for further processing to predict the rusted parts in the node plate, and the number of pixels corresponding to the node plate and the rusted node plate is calculated.

[0020] The evaluation module is used to process the node plate image and the number of pixels corresponding to the node plate corrosion using a corrosion degree evaluator to obtain the corrosion degree status of the node plate.

[0021] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the railway steel truss bridge coating corrosion detection and evaluation method as described in the first aspect.

[0022] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the railway steel truss bridge coating corrosion detection and evaluation method as described in the first aspect.

[0023] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the railway steel truss bridge coating corrosion detection and evaluation method as described in the first aspect.

[0024] The beneficial effects of this invention are as follows: The proposed target segmenter achieves an intersection-to-union (IoU) ratio of 76.5% in the rusted area. After introducing the data augmentation module, the IoU ratio is further improved to 79.0%, which is significantly better than existing methods for rust segmentation and data augmentation of railway steel truss bridge node plates.

[0025] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments 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.

[0027] Figure 1 The basic flowchart of the design of the railway steel truss bridge coating corrosion detection and evaluation method according to the embodiments of the present invention.

[0028] Figure 2 This is an example diagram of the node plate and corrosion of the node plate in the railway steel truss bridge according to an embodiment of the present invention.

[0029] Figure 3 This is a rendering of the node plate corrosion image data enhancement module according to an embodiment of the present invention.

[0030] Figure 4 This is a structural diagram of the separable ECA residual module (DE-RB) according to an embodiment of the present invention.

[0031] Figure 5 This is a structural diagram of the deformable residual bottleneck module (D-RBB) according to an embodiment of the present invention.

[0032] Figure 6This is a structural diagram of the parallel channel attention pyramid pooling module (PCA-PPM) according to an embodiment of the present invention.

[0033] Figure 7 This is a complete model architecture diagram as described in the embodiments of the present invention. Detailed Implementation

[0034] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0035] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0037] In the description of this specification, 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. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. 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 those different embodiments or examples.

[0038] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0039] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0040] Example 1

[0041] In this embodiment 1, a training method for a target segmenter model of corrosion of railway steel truss bridge node plates is provided. The target segmenter model trained using this method can accurately identify the corrosion area of ​​railway steel truss bridge node plates.

[0042] Specifically, in this embodiment, the target segmenter model training method includes the following steps:

[0043] Step 1: Obtain multiple images of the railway steel truss bridge, which include the node plate area and the rusted area of ​​the node plate.

[0044] Step 2: Perform data annotation. Using the image data obtained in Step 1, create a dataset for the location and corrosion segmentation of the node plate area of ​​the railway steel truss bridge. This dataset contains two sub-datasets: the node plate area annotation dataset (dataset 1) and the corrosion segmentation annotation dataset (dataset 2). Then, send the two datasets into the data augmentation module for data augmentation.

[0045] Step 3: Based on the two datasets created, feed them into the target segmenter model for training. By adopting an incremental learning strategy, use one target segmenter to train twice to gradually achieve accurate segmentation of the rust on the railway steel truss bridge.

[0046] In step 1, the image of the railway steel truss bridge can be obtained in the following way:

[0047] 1) Obtained through public datasets and academic resources:

[0048] Traffic / bridge image datasets: such as Bridges Dataset (released by MIT), Structurae (bridge engineering database), Open Infrastructure Map, etc. Universities or research institutions may have released bridge inspection datasets (such as the "Railway Bridge Defect Image Dataset").

[0049] Figures in academic papers: Searching for keywords such as "steel truss bridge detection" and "railway bridge images" on platforms such as Google Scholar and CNKI often reveals example images in the papers.

[0050] Competition datasets: such as infrastructure detection competitions on Kaggle (e.g., "bridge crack detection").

[0051] 2) Field data collection:

[0052] By cooperating with railway departments and engineering units: cooperate with railway bureaus, bridge inspection companies, or construction companies to obtain on-site inspection photos (pay attention to safety agreements and confidentiality).

[0053] Drone aerial photography: Use drones for photography in permitted areas (airspace regulations must be followed to avoid interfering with railway operations).

[0054] Open source maps and street views: Baidu / Google Street View, Mapillary and other open platforms may contain bridge images (note the terms of use).

[0055] 3) Web crawlers (caution and compliance required):

[0056] Image search engine: Use Bing or Google Image Search keywords (such as "railway truss bridge" or "steel truss bridge") and download images in batches via script (copyrighted images must be filtered and robots.txt must be followed).

[0057] Professional stock photo libraries and social media: Flickr, Pixabay (free copyrighted images); engineering-related hashtags (such as #bridgephotography) on platforms like Instagram and Weibo.

[0058] Satellite and aerial imagery services: Google Earth, Tianditu, and other services provide satellite imagery that can capture aerial views of bridges.

[0059] 4) Synthetic data generation

[0060] 3D Modeling and Rendering: Use software such as Blender and SketchUp to create steel truss bridge models and generate virtual images from multiple angles with different lighting conditions.

[0061] Game / Simulation Engine: Generates data using bridge assets from Unity and Unreal Engine (suitable for algorithm training).

[0062] In practical applications, the acquisition of the aforementioned image data should be conducted with caution regarding copyright and privacy, avoiding the use of unauthorized commercial images, especially those involving sensitive locations related to railway infrastructure. Unauthorized entry into railway areas for photography is strictly prohibited; prior communication with the relevant management authorities is required.

[0063] In step 2, data augmentation on the labeled training set can effectively improve the model's generalization ability. The key principle is that augmentation operations must be applied simultaneously to the image and its corresponding annotations (such as bounding boxes, segmentation masks, and keypoints). In this embodiment, the labeled data can be augmented in the following way:

[0064] 1) Basic image enhancement (annotations need to be updated synchronously):

[0065] These transformations typically do not change the geometric position information of the annotations; they only require processing the image itself.

[0066] Color / Brightness Transformation: Operations: Adjust brightness, contrast, saturation, and hue; add random lighting noise; apply histogram equalization; grayscale (partial channels). Annotation Impact: No annotation modification required. Only pixel values ​​change; annotation coordinates or masks remain unchanged.

[0067] Blur and Noise: Operations: Gaussian blur, motion blur, mean blur; add Gaussian noise, salt-and-pepper noise. Annotation Impact: No annotation modification required.

[0068] 2) Geometric transformation enhancement (annotations must be transformed simultaneously):

[0069] Such transformations change pixel positions, so the same geometric transformation must be applied to the annotations.

[0070] Flip: Operations: Horizontal flip, vertical flip. Annotation transformation: The coordinates of the bounding box need to be flipped accordingly. For example, when flipping horizontally, the x-coordinate becomes width - x_max and width - x_min (the order needs to be swapped).

[0071] Rotation: Operation: Random rotation (within ±30°). Note that large-angle rotations may introduce unnatural backgrounds (requiring padding). Annotation Transformation: Bounding boxes must be transformed using the same rotation matrix, which usually results in skewed boxes. To simplify processing, the minimum bounding rectangle after rotation can be calculated, but this will introduce redundant background. For instance segmentation tasks, simply rotating the mask is sufficient.

[0072] Scaling and Cropping: Operations: Random scaling (e.g., 0.8~1.2 times); random cropping of a portion of the image. Label Transformation: Scaling: Directly scale all coordinates proportionally. Cropping: Use the cropped region as the origin of the new image and adjust the label coordinates (subtracting the offset). Labels completely outside the cropped region need to be filtered out, and some labels within the boundaries need to be processed (usually retaining bounding boxes with an IoU greater than a threshold).

[0073] Translation: Operation: Randomly move the image along the X / Y axes. Label Transformation: Add or subtract offsets to coordinates and handle labels that have moved out of bounds.

[0074] 3) Advanced and hybrid enhancement strategies:

[0075] Mixup and CutMix: Operations: Mixup: Linearly merges two images I_new = α*I1 + (1-α)*I2, and the corresponding labels (such as classification probabilities, bounding box coordinates) are also merged in the same proportion. CutMix: Crops a region from image A and pastes it into the corresponding region of image B, and the labels are also mixed accordingly (such as merging bounding box sets, and weighting the classification labels according to the area ratio).

[0076] Mosaic Enhancement: Operation: Consolidates four training images into one and merges their bounding boxes accordingly. Advantages: Provides multi-scale context at once, significantly improving small object detection performance, and resulting in more stable batch normalization statistics.

[0077] Random Erasing / CutOut: Operation: Randomly select a rectangular region in the image and fill it with random or mean values. Annotation Impact: No annotation modification is required. Simulates partial occlusion of objects, enhancing model robustness.

[0078] Style Transfer / Domain Randomization: Operation: Use GANs or filters to change the style of images (e.g., simulating different weather conditions or seasons). Annotation Impact: No annotation modification is required. Primarily used to improve the model's adaptability to different environments.

[0079] In practical applications, considering the characteristics of bridge images, the following enhancements can be emphasized:

[0080] Multi-scale and multi-angle: Steel truss bridges have many components and repetitive structures. Multi-scale scaling and random cropping are used to allow the model to learn components at different distances and locations. Appropriate rotation (small angles) simulates different shooting angles.

[0081] Lighting and Weather Simulation: Railway bridges present complex environments. Random variations in brightness and contrast are added to simulate lighting conditions at different times of day. Fog, rain, and snow simulations are added (using filtering or GANs) to improve detection robustness under adverse weather conditions.

[0082] Enhanced local occlusion: Bridges may be obscured by vegetation or other structural elements. Use random rectangular occlusion, or a more advanced grid occlusion (GridMask).

[0083] Background blending: After cutting out the labeled bridge components, paste them onto different backgrounds (such as sky, mountains, and cities) to increase background diversity (requires precise instance segmentation masks).

[0084] Augmented images should remain visually plausible. For example, bridges should not be inverted (large-angle rotations can be disabled). After transformation, always check that the labels still correctly correspond to the targets. More augmentation is not necessarily better; the intensity and combination of augmentations should be adjusted based on the dataset size and model performance. Never perform data augmentation on the validation and test sets (except for necessary operations such as normalization).

[0085] Step 3, the incremental training phase of the target segmenter, specifically includes:

[0086] First, a target segmenter is trained using dataset 1 to achieve accurate localization of node plates in railway steel truss bridge images. Then, training datasets 1 and 2 are run on the same target segmenter. In the initial training phase, the model parameters are initialized using the model weights generated in "Model Training Step 1". The trained model possesses two capabilities: 1) it retains the ability of "Model Training Step 1" to locate and identify node plates in railway steel truss bridges; 2) it adds the ability to segment rust on node plates. The complete data processing flow after model architecture is as follows: Figure 7 As shown.

[0087] In this embodiment, considering that directly extracting features from the image would greatly increase the complexity of the model, preprocessing of the training set images is necessary. When the stride of the convolutional module is 2, it indicates that the image is downsampled, and no specific restrictions are placed on the design of the convolutional module in the preprocessing module. The image first undergoes two downsampling processes, reducing the image size to 1 / 4 of the initial image. Then, in order to enable the model to have continuous learning capabilities, the image is processed by the residual module (RB), ultimately reducing the image size to 1 / 8 of the initial image, as shown in Table 1.

[0088] Table 1

[0089]

[0090] In this embodiment, the target segmenter specifically includes the following functional module (network) layer:

[0091] Feature Extraction Module: To accurately extract features of railway steel truss bridge node plates and their corrosion, this embodiment employs a three-branch approach: detail branch, context branch, and boundary branch. Optimized for semantic segmentation tasks, as shown in Table 4, the detail branch utilizes a residual module (RB), a separable ECA residual module (DE-RB), and a residual bottleneck module (RBB) with a step size of 1. This preserves large-scale feature maps to capture fine-grained details in the image, crucial for accurate segmentation of the node plate corrosion region. The context branch uses the same modules as the detail branch, but with a step size of 2. This involves continuous downsampling to extract local and global features. A novel parallel channel attention pyramid pooling module (PCA-PPM) is introduced to effectively enhance the model's ability to model contextual information, thereby strengthening the network's understanding of complex scenes. The boundary branch focuses on extracting high-frequency features, using the residual module (RB) and deformable residual bottleneck module (D-RBB) to accurately predict boundary regions. Upsampling is employed to facilitate the subsequent merging of the three branches.

[0092] Among them, the separable ECA residual module (DE-RB) replaces the second layer with a depthwise separable convolution based on the residual module (RB). By decomposing the standard convolution into depthwise convolution and pointwise convolution, it reduces computational complexity. At the same time, it embeds an efficient channel attention (ECA) module to adaptively allocate channel weights in a lighter way, thereby improving the ability to represent the rust features of the node board. Figure 4 As shown.

[0093] The Deformable Residual Bottleneck Module (D-RBB) introduces a deformable convolutional module to the classic Residual Bottleneck Module (RBB). The 3×3 convolutions in the classic RBB are replaced by deformable convolutions, whose offsets are predicted by a separate 3×3 convolutional layer and work in conjunction with the main convolutional layer to perform deformable feature sampling and processing. Simultaneously, the D-RBB retains the efficient bottleneck structure of the classic RBB, achieving feature compression and expansion through two 1×1 convolutions, and combining residual connections to ensure gradient flow. This enhances the ability to model complex boundary scenarios while maintaining computational efficiency. Figure 5 As shown.

[0094] The Parallel Channel Attention Pyramid Pooling Module (PCA-PPM) consists of multi-scale pooling branches, a channel attention mechanism, and aggregation operations. The module extracts multi-scale features through three branches with different downsampling ratios and further optimizes feature processing efficiency during the fusion stage using grouped convolution. Finally, by introducing a lightweight channel attention mechanism, the correlation between channels is enhanced after compressing channel information, thereby effectively mitigating the potentially adverse effects of grouped convolution. Figure 6 As shown.

[0095] Semantic segmentation module: In this implementation, in order to achieve efficient feature fusion and high-performance semantic segmentation with low computational cost, a lightweight boundary attention-guided fusion module (Light-Bag) is used to fuse the features extracted from the three branches, and a simple convolutional structure is used as the segmentation head to achieve pixel-level classification of the image, as shown in Table 4.

[0096] In this embodiment, the loss function used for training the above model is: the loss function aims to optimize the semantic segmentation task from multiple dimensions and includes four different components.

[0097] In the detail branch, an additional semantic head is used to calculate the loss (denoted as L0) on the intermediate feature layer using the cross-entropy loss function. This loss is used to supervise the learning of fine-grained features, thereby improving the network's ability to capture complex details in images, satisfying the following equation:

[0098] ;

[0099] in, This represents the total number of valid pixels in the corresponding loss function. Represents pixels The one-hot encoded value of the semantic label in category c. This represents the predicted probability that pixel i belongs to category c.

[0100] In the boundary branch, to address the class imbalance caused by sparse boundary pixels, a weighted binary cross-entropy loss function (denoted as L1) is used to calculate the loss, satisfying the following equation:

[0101] ;

[0102] in, , These represent the predicted boundary map and the true boundary map output by the boundary head, respectively.

[0103] The semantic context branch uses standard cross-entropy loss (L2) to supervise the semantic segmentation results. This loss focuses on global contextual information and, by integrating multi-scale features, guides the network to better understand complex regions, improving the segmentation accuracy of large regions, satisfying the following equation:

[0104] ;

[0105] in, It is the confidence threshold for online hard sample mining;

[0106] Finally, another cross-entropy loss (L3) is used as a coordination loss to jointly optimize the semantic segmentation and boundary detection tasks, thereby enhancing the synergy between the context branch and the boundary branch, improving the overall segmentation accuracy and boundary localization capability, satisfying the following equation:

[0107] ;

[0108] in, It is the confidence threshold for boundary awareness. Represents the pixels output by the boundary header. Boundary confidence level;

[0109] Finally, the segmenter loss function in this embodiment is composed of the above four loss functions, and the total loss function is expressed as follows:

[0110]

[0111] In this embodiment, the weight ratio of each loss is... Set them to 0.4, 20, 1 and 1 respectively.

[0112] The target segmenter iteratively calculates the prediction results of the data read in each iteration, calculates the total loss function based on the prediction results and labels, and uses stochastic gradient descent to optimize the convolutional parameters of each module in the segmenter and a series of related updatable parameters. The iteration continues until the pre-set number of iterations is reached, ending model training. This comprehensive loss function design effectively supervises the learning process of each branch and balances their contributions to the overall network performance.

[0113] Example 2

[0114] In this embodiment 2, a method for detecting corrosion of railway steel truss bridge node plates is first provided. This method uses the target segmenter model trained in embodiment 1 to detect corrosion of railway steel truss bridge node plates and obtain the corrosion area of ​​railway steel truss bridge node plates.

[0115] Specifically, first load the model weights of the target segmenter trained in Example 1;

[0116] The images of the railway steel truss bridge acquired on-site are input into the trained target segmenter for inference, predicting the node plate structure in the railway steel truss bridge. The predicted node plates are then cropped to obtain node plate images. The node plate images are then input into the target segmenter for inference, predicting the corroded parts in the node plates.

[0117] In this embodiment, the degree of corrosion of the node plate predicted above is also evaluated, and the number of pixels corresponding to the node plate and the corrosion of the node plate is calculated. The node plate and the number of pixels corresponding to the corrosion of the node plate are input into the corrosion degree evaluator trained on the node plate, and the corrosion degree status of the node plate is output.

[0118] In this embodiment, the corrosion rate of the node plate is calculated according to the following formula:

[0119] ;

[0120] in, This represents the total number of pixels in the node board. This represents the number of pixels in the rusted area.

[0121] In this embodiment, the rust level and rust degree of the node plate are calculated according to the rust rate of the node plate. The correspondence between the rust rate of the node plate and the rust level and rust degree of the node plate set in the rust degree evaluator is shown in Table 2.

[0122] Table 2

[0123]

[0124] Example 3

[0125] In this embodiment 3, a system for detecting and evaluating the corrosion of coatings on railway steel truss bridges is first provided. This system includes: an acquisition module for acquiring images of the railway steel truss bridge to be processed; a prediction module for processing the acquired images of the railway steel truss bridge using a pre-trained target segmenter model to predict the node plate structure in the railway steel truss bridge; cropping the predicted node plates to obtain node plate images; further inputting the node plate images into the target segmenter model for processing to predict the corroded parts in the node plates and calculating the node plates and the number of pixels corresponding to the corrosion; and an evaluation module for processing the node plate images and the number of pixels corresponding to the corrosion using a corrosion degree evaluator to obtain the corrosion degree status of the node plates.

[0126] In this embodiment, the above-described system is used to implement a method for detecting and evaluating the corrosion of coatings on railway steel truss bridges. The method includes: first, acquiring an image of the railway steel truss bridge to be processed using an acquisition module; then, using a prediction module with a pre-trained target segmenter model to process the acquired image and predict the node plate structure in the bridge; cropping the predicted node plate to obtain a node plate image; further inputting the node plate image into the target segmenter model for processing to predict the corroded portion of the node plate and calculate the node plate and the number of pixels corresponding to the corrosion; finally, using a corrosion degree evaluator to process the node plate image and the number of pixels corresponding to the corrosion, obtaining the corrosion degree status of the node plate.

[0127] The prediction module loads a target segmenter model trained using the training method described in Example 1. This target segmenter model includes a feature extraction unit and a semantic segmentation unit. The feature extraction unit extracts features of railway steel truss bridge node plates and node plate corrosion. The feature extraction unit uses detail branches, context branches, and boundary branches. The detail branch preserves large-scale feature maps and captures fine-grained details in the image. The context branch extracts local and global features through continuous downsampling operations, improving the modeling ability of contextual information and enhancing the network's understanding of complex scenes. The boundary branch extracts high-frequency features, predicts boundary regions, and simultaneously performs upsampling operations on the extraction results of the detail branch, context branch, and boundary branch. The semantic segmentation unit fuses the features extracted by the detail branch, context branch, and boundary branch, and uses a convolutional structure as the segmentation head to achieve pixel-level classification of the image.

[0128] The detail branch uses the residual module, the separable ECA residual module, and the residual bottleneck module in succession. The context branch uses the same module as the detail branch, extracting local and global features through continuous downsampling operations. At the same time, an improved parallel channel attention pyramid pooling module is introduced to improve the ability to model context information and enhance the network's understanding of complex scenes.

[0129] The boundary branch predicts the boundary region through the residual module and the deformable residual bottleneck module. The separable ECA residual module, based on the residual module, replaces the second layer with a depthwise separable convolution. By decomposing the standard convolution into depthwise and pointwise convolutions, computational complexity is reduced. Simultaneously, an efficient channel attention module is embedded to adaptively allocate channel weights in a lighter manner. The deformable residual bottleneck module introduces a deformable convolution module on top of the classic residual bottleneck module. The 3×3 convolution in the classic residual bottleneck module is replaced by a deformable convolution, whose offset is predicted by an independent 3×3 convolutional layer, and works with the main convolutional layer to complete deformable feature sampling and processing. At the same time, the deformable residual bottleneck module retains the efficient bottleneck structure of the classic residual bottleneck module, achieving feature compression and expansion through two 1×1 convolutions.

[0130] The parallel channel attention pyramid pooling module consists of multi-scale pooling branches, channel attention mechanisms, and aggregation operations. It extracts multi-scale features through three branches with different downsampling ratios and further optimizes feature processing efficiency by using grouped convolutions in the fusion stage. Finally, by introducing a lightweight channel attention mechanism, the correlation between channels is enhanced after compressing channel information, thereby effectively mitigating the adverse effects that grouped convolutions may cause.

[0131] The evaluation module is loaded with a corrosion degree evaluator, which calculates the corrosion level and degree of the node plate according to the corrosion rate of the node plate. The correspondence between the corrosion rate of the node plate and the corrosion level and degree of the node plate set in the corrosion degree evaluator is shown in Table 2 of Example 2.

[0132] Example 4

[0133] This embodiment 4 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the railway steel truss bridge coating corrosion detection and evaluation method described above. The method includes:

[0134] Acquire images of the railway steel truss bridge to be processed;

[0135] The acquired railway steel truss bridge image is processed using a pre-trained target segmenter model to predict the node plate structure in the railway steel truss bridge; the predicted node plate is cropped to obtain the node plate image; the node plate image is then input into the target segmenter model for further processing to predict the rusted parts in the node plate, and the number of pixels corresponding to the node plate and the rusted node plate is calculated.

[0136] The corrosion degree evaluator is used to process the node plate image and the number of pixels corresponding to the corrosion of the node plate to obtain the corrosion degree status of the node plate.

[0137] Example 5

[0138] This embodiment 5 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the railway steel truss bridge coating corrosion detection and evaluation method as described above, the method including:

[0139] Acquire images of the railway steel truss bridge to be processed;

[0140] The acquired railway steel truss bridge image is processed using a pre-trained target segmenter model to predict the node plate structure in the railway steel truss bridge; the predicted node plate is cropped to obtain the node plate image; the node plate image is then input into the target segmenter model for further processing to predict the rusted parts in the node plate, and the number of pixels corresponding to the node plate and the rusted node plate is calculated.

[0141] The corrosion degree evaluator is used to process the node plate image and the number of pixels corresponding to the corrosion of the node plate to obtain the corrosion degree status of the node plate.

[0142] Example 6

[0143] This embodiment 6 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device performs the above-described method for detecting and evaluating the corrosion of the coating of a railway steel truss bridge.

[0144] Acquire images of the railway steel truss bridge to be processed;

[0145] The acquired railway steel truss bridge image is processed using a pre-trained target segmenter model to predict the node plate structure in the railway steel truss bridge; the predicted node plate is cropped to obtain the node plate image; the node plate image is then input into the target segmenter model for further processing to predict the rusted parts in the node plate, and the number of pixels corresponding to the node plate and the rusted node plate is calculated.

[0146] The corrosion degree evaluator is used to process the node plate image and the number of pixels corresponding to the corrosion of the node plate to obtain the corrosion degree status of the node plate.

[0147] In summary, this invention proposes a two-stage detection and evaluation method for railway steel truss bridge node plate modification, combining corrosion segmentation (using only a single target segmenter with a simple incremental training strategy) and a corrosion degree evaluator (manually designed evaluation criteria). A node plate region localization and node plate corrosion segmenter based on three-branch decoupled semantic segmentation are designed. Only one target segmenter is used, and a simple incremental learning strategy is fused for two training iterations to achieve node plate region localization and node plate corrosion segmentation. Simultaneously, a node plate corrosion degree evaluator is designed. By inputting the node plate region and the corresponding number of corroded pixels, it outputs the degree and level of node plate corrosion, effectively improving the interpretability and practicality of the detection results. A data augmentation method fusing "pixel enhancement space" and "spatial enhancement space" is proposed. Through joint enhancement of pixel-level features and spatial distribution structure, and by employing a controllable combination strategy during the enhancement process, the enhanced data maintains high usability while preserving the original semantics, thereby effectively alleviating the model training bottleneck caused by insufficient samples.

[0148] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A method for detecting and evaluating corrosion of coatings on railway steel truss bridges, characterized in that, include: Acquire images of the railway steel truss bridge to be processed; The acquired images of the railway steel truss bridge to be processed are processed using a pre-trained target segmenter model to predict the node plate structure in the railway steel truss bridge. The predicted node board is cropped to obtain the node board image; The node plate image is then input into the target segmenter model for processing to predict the corroded parts in the node plate and to calculate the number of pixels corresponding to the node plate and the corroded parts. The corrosion degree evaluator is used to process the node plate image and the number of pixels corresponding to the corrosion of the node plate to obtain the corrosion degree status of the node plate.

2. The method for detecting and evaluating corrosion of coatings on railway steel truss bridges according to claim 1, characterized in that, The target segmenter model includes a feature extraction unit and a semantic segmentation unit. The feature extraction unit is used to extract features of railway steel truss bridge node plates and node plate corrosion. The feature extraction unit uses detail branches, context branches, and boundary branches. The detail branches are used to retain large-scale feature maps and capture fine-grained details in the image. The context branches extract local and global features through continuous downsampling operations, improving the ability to model context information and enhancing the network's understanding of complex scenes. The boundary branches extract high-frequency features, predict boundary regions, and simultaneously perform upsampling operations on the extraction results of the detail branches, context branches, and boundary branches. The semantic segmentation unit fuses the features extracted by the detail branches, context branches, and boundary branches, and uses a convolutional structure as the segmentation head to achieve pixel-level classification of the image.

3. The method for detecting and evaluating corrosion of coatings on railway steel truss bridges according to claim 2, characterized in that, The detail branch successively uses the residual module, the separable ECA residual module, and the residual bottleneck module; the context branch uses the same module as the detail branch, extracting local and global features through continuous downsampling operations, and introduces an improved parallel channel attention pyramid pooling module to enhance the modeling ability of context information and improve the network's understanding of complex scenes.

4. The method for detecting and evaluating corrosion of coatings on railway steel truss bridges according to claim 3, characterized in that, The boundary branch predicts the boundary region through the residual module and the deformable residual bottleneck module. The separable ECA residual module, based on the residual module, replaces the second layer with a depthwise separable convolution. By decomposing the standard convolution into depthwise and pointwise convolutions, computational complexity is reduced. Simultaneously, an efficient channel attention module is embedded to adaptively allocate channel weights in a lighter manner. The deformable residual bottleneck module introduces a deformable convolution module on top of the classic residual bottleneck module. The 3×3 convolution in the classic residual bottleneck module is replaced by a deformable convolution, whose offset is predicted by an independent 3×3 convolutional layer, and works with the main convolutional layer to complete deformable feature sampling and processing. At the same time, the deformable residual bottleneck module retains the efficient bottleneck structure of the classic residual bottleneck module, achieving feature compression and expansion through two 1×1 convolutions.

5. The method for detecting and evaluating corrosion of coatings on railway steel truss bridges according to claim 4, characterized in that, The parallel channel attention pyramid pooling module consists of multi-scale pooling branches, channel attention mechanisms, and aggregation operations. It extracts multi-scale features through three branches with different downsampling ratios and further optimizes feature processing efficiency by using grouped convolution in the fusion stage. Finally, by introducing a lightweight channel attention mechanism, the correlation between channels is enhanced after compressing channel information, thereby effectively mitigating the adverse effects that grouped convolution may cause.

6. The method for detecting and evaluating corrosion of coatings on railway steel truss bridges according to claim 5, characterized in that, The loss functions used in training the target segmenter model are as follows: In the detail branch, a cross-entropy loss function is used to calculate the loss L0 on the intermediate feature layer through an additional semantic head to supervise the learning of fine-grained features, thereby improving the network's ability to capture complex details in the image; in the boundary branch, a weighted binary cross-entropy loss function is used to calculate the loss L1 to solve the class imbalance problem caused by sparse boundary pixels; the semantic context branch uses the standard cross-entropy loss L2 to supervise the semantic segmentation results. This loss focuses on global context information and guides the network to better understand complex regions by integrating multi-scale features, thereby improving the segmentation accuracy of large areas; another cross-entropy loss L3 is used as a coordination loss to jointly optimize the semantic segmentation and boundary detection tasks, thereby enhancing the synergy between the context branch and the boundary branch.

7. A system for detecting and evaluating corrosion of coatings on railway steel truss bridges, characterized in that, include: The acquisition module is used to acquire images of railway steel truss bridges to be processed; The prediction module is used to process the acquired images of the railway steel truss bridge to be processed using a pre-trained target segmenter model, and to predict the node plate structure in the railway steel truss bridge. The predicted node board is cropped to obtain the node board image; The node plate image is then input into the target segmenter model for processing to predict the corroded parts in the node plate and to calculate the number of pixels corresponding to the node plate and the corroded parts. The evaluation module is used to process the node plate image and the number of pixels corresponding to the node plate corrosion using a corrosion degree evaluator to obtain the corrosion degree status of the node plate.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the railway steel truss bridge coating corrosion detection and evaluation method as described in any one of claims 1-6.

9. A computer device, characterized in that, The system includes a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the railway steel truss bridge coating corrosion detection and evaluation method as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the railway steel truss bridge coating corrosion detection and evaluation method as described in any one of claims 1-6.

Images