Bridge bank slope deformation abnormal area identification method and device, equipment, medium and program product

Abstract

The invention provides a bridge bank slope deformation abnormal area identification method and device, equipment, a medium and a program product, and relates to the technical field of geological prediction. According to the invention, an improved UNet (SEUNet) neural network model based on an SE module attention mechanism is developed, and the deformation abnormal region in the GB-SAR image can be efficiently extracted. In the application, SE modules are respectively constructed for four hierarchical encoders of a UNet model, and an SEUNet model is jointly formed. And then, a GB-SAR data set is constructed and labeled, and a data enhancement technology is combined, so that the generalization ability of the model is improved. Experimental results show that the model provided by the invention has higher accuracy and stability in deformation abnormal area detection.

Description

Technical Field

[0001] The present application relates to the field of geological prediction technology, and in particular to a method, device, equipment, medium and program product for identifying abnormal deformation areas of bridge bank slopes. Background Art

[0002] During bridge construction, human engineering activities such as excavation and unloading may significantly affect the stability of the bridge bank slope. Therefore, real-time monitoring and identification of abnormal slope deformation areas have become a key means to ensure bridge safety. In recent years, total stations, GNSS and photogrammetry technologies have been widely used in slope monitoring, and ground-based synthetic aperture radar (GB-SAR) has been widely used in stability monitoring of high and steep slopes on bridge banks due to its all-weather and all-day monitoring capabilities. However, real-time and effective detection of abnormal deformation areas in GB-SAR images still faces challenges. The identification of abnormal deformation areas in traditional GB-SAR images mainly relies on manual visual interpretation. This method is limited by image rendering methods and human subjective judgment, and the real-time and automation levels are low. With the increase in the amount of SAR data during the monitoring process, manual visual interpretation becomes time-consuming and laborious, and it is difficult to achieve real-time performance, and it is difficult to link with the early warning system. Summary of the invention

[0003] The present application provides a method, device, medium and program product for identifying abnormal deformation areas of bridge bank slopes to at least partially solve the above-mentioned problems.

[0004] In a first aspect, the present application provides a method for identifying abnormal deformation areas of a bridge bank slope, the method comprising: Obtain historical GB-SAR images corresponding to the deformation data of the bridge bank slope in the target study area; Annotating the historical GB-SAR images to obtain true values ​​of abnormal deformation regions corresponding to the historical GB-SAR images, and obtaining a training data set based on a plurality of historical GB-SAR images annotated with the true values ​​of abnormal deformation regions; The preset model is trained based on the training data set to obtain a bridge bank slope deformation abnormal area recognition model; The GB-SAR image to be identified of the bridge bank slope in the target study area is input into the bridge bank slope deformation abnormal area identification model to obtain the bridge bank slope deformation abnormal area identification result.

[0005] Optionally, the historical GB-SAR images are annotated to obtain true values ​​of abnormal deformation regions corresponding to the historical GB-SAR images, and a training data set is obtained based on a plurality of historical GB-SAR images annotated with true values ​​of abnormal deformation regions, including: Obtain annotations of abnormal deformation areas in historical GB-SAR images; Dividing the historical GB-SAR image into a plurality of image slices of preset sizes; Generate a true value mask of the abnormal deformation area of ​​each image slice based on the abnormal deformation area annotation; A training data set is obtained based on image slices of multiple preset sizes and true value masks of abnormal deformation areas corresponding to each image slice.

[0006] Optionally, training a preset model based on the training data set includes: The image slices of the preset size are used as single-channel floating-point matrices of the preset size, and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size are used as single-channel binary matrices of the preset size in binary format, which are input into the preset model to train the preset model.

[0007] Optionally, before inputting the training data into the preset model, the method further includes: The image slices of the preset size and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size are randomly flipped, and the flipped image slices and the true value masks of the abnormal deformation areas corresponding to the image slices are used as training data.

[0008] Optionally, before inputting the training data into the preset model, the method further includes: The image slices of the preset size and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size are randomly rotated, and the rotated image slices and the true value masks of the abnormal deformation areas corresponding to the image slices are used as training data.

[0009] Optionally, the preset model includes: four layers of downsampling encoders, each layer of downsampling encoders, with an SE module of the same size added, the SE module is used to calculate the weights of each channel of the corresponding layer, and recalculate its output value according to the weight as the final output of the encoder.

[0010] The second aspect of the present application provides a device for identifying abnormal deformation areas of bridge bank slopes, the device comprising: The acquisition module is used to obtain the historical GB-SAR images corresponding to the deformation data of the bridge bank slope in the target study area; A labeling module is used to label the historical GB-SAR images, obtain the true value of the abnormal deformation area corresponding to the historical GB-SAR images, and obtain a training data set based on a plurality of historical GB-SAR images labeled with the true value of the abnormal deformation area; A training module, used to train a preset model based on the training data set to obtain a bridge bank slope deformation abnormal area identification model; The identification module is used to input the GB-SAR image to be identified of the bridge bank slope in the target study area into the bridge bank slope deformation abnormal area identification model to obtain the bridge bank slope deformation abnormal area identification result.

[0011] The third aspect of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when executed, the processor implements the method for identifying abnormal deformation areas of bridge bank slopes as described in the first aspect of the present invention.

[0012] A fourth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for identifying abnormal deformation areas of bridge bank slopes as described in the first aspect of the present invention.

[0013] The fifth aspect of the present application provides a computer program product, including a computer program / instruction, which is used by a processor to implement the steps of the method for identifying abnormal deformation areas of bridge bank slopes as described in the first aspect of the present invention.

[0014] This application develops an improved UNet (SEUNet) neural network model based on the SE module attention mechanism, which can efficiently extract deformation abnormal areas in GB-SAR images. In this application, SE modules are constructed for the four-level encoders of the UNet model to form a SEUNet model. Then, by constructing and annotating the GB-SAR dataset and combining data enhancement technology, the generalization ability of the model is improved. Experimental results show that the model proposed in this application has higher accuracy and stability in the detection of deformation abnormal areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the description of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 It is a flowchart of the steps of the method for identifying abnormal deformation areas of bridge bank slopes provided in this application; Figure 2 It is a schematic diagram of the UNet basic structure in the bridge bank slope deformation abnormal area identification method provided in this application; Figure 3 It is a schematic diagram of the SE module mechanism principle in the bridge bank slope deformation abnormal area identification method provided by this application; Figure 4 It is a schematic diagram of the SEUNet model structure in the bridge bank slope deformation abnormal area identification method provided in this application; Figure 5 It is a schematic diagram of calculating the intersection-over-union (IoU) ratio in the bridge bank slope deformation abnormal area identification method provided in this application; Figure 6 It is the GB-SAR surface displacement monitoring result map (GeoTIFF image) of the study area in the bridge bank slope deformation abnormal area identification method provided in this application; Figure 7 This is an example diagram of an abnormal area marking in the bridge bank slope deformation abnormal area identification method provided in this application; Figure 8 It is an example diagram of the slices of the model input and the corresponding true value of the abnormal deformation area in the bridge bank slope deformation abnormal area identification method provided by the present application; Fig. 9 It is a comparison chart of the loss curves of the SEUNet and UNet models in the bridge bank slope deformation abnormal area identification method provided in this application in the training set and the test set; Fig.10 It is the detection result of abnormal deformation areas at different periods by the SEUNet model in the bridge bank slope deformation abnormal area identification method provided in this application; Fig.11 It is a comparison chart of the SEUNet and UNet model results in the bridge bank slope deformation abnormal area identification method provided in this application. DETAILED DESCRIPTION

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] At present, scholars at home and abroad have tried to apply deep learning models such as YOLO, CNN and UNet to the identification of deformation zones in large-scale SAR satellite images, mainly for large-scale landslide screening, and rarely involve the identification of abnormal deformation zones in ground-based GB-SAR images. Although space-based SAR and ground-based SAR are similar in principle, there are significant differences in their deployment schemes, monitoring ranges, monitoring frequencies and data analysis methods, and the identification schemes of space-based SAR images cannot be directly applied. Therefore, based on the actual deployment characteristics of GB-SAR for bridge bank slopes, it is particularly important to carry out automatic identification of abnormal deformation zones in bridge bank slopes based on GB-SAR and deep learning models.

[0019] This application develops an improved UNet (SEUNet) neural network model based on the SE module attention mechanism to efficiently extract deformation anomaly areas in GB-SAR images. In this application, SE modules are constructed for the four-level encoders of the UNet model to form a SEUNet model. Then, by constructing and annotating the GB-SAR dataset and combining data enhancement technology, the generalization ability of the model is improved. Experimental results show that the model proposed in this application has higher accuracy and stability in the detection of deformation anomaly areas.

[0020] Specifically, the steps of the bridge bank slope deformation abnormal area identification method proposed in this application are as follows: Figure 1 As shown, the method comprises the following steps: S101, obtaining historical GB-SAR images corresponding to the deformation data of the bridge bank slope in the target study area.

[0021] In the present application, historical GB-SAR images can be obtained from the target study area to train the preset model, so as to identify the GB-SAR images to be identified in the study area based on the trained model.

[0022] In this application, in order to achieve the purpose of accurate identification, after the study area is changed, the trained model can be fine-tuned based on the historical GB-SAR images of the changed study area, so that the fine-tuned model can adapt to the GB-SAR image characteristics of the changed study area.

[0023] S102, annotating the historical GB-SAR images to obtain true values ​​of abnormal deformation regions corresponding to the historical GB-SAR images, and obtaining a training data set based on a plurality of historical GB-SAR images annotated with the true values ​​of abnormal deformation regions.

[0024] In this application, the abnormal deformation areas in the historical GB-SAR images can be manually annotated, and then a training data set can be constructed based on the historical GB-SAR images with annotated abnormal deformation areas.

[0025] Specifically, in the present application, step S102 may include the following sub-steps: obtaining annotations of abnormal deformation regions in historical GB-SAR images; S1021, dividing the historical GB-SAR image into a plurality of image slices of preset sizes.

[0026] S1022: Generate a true value mask of the abnormal deformation area of ​​each image slice based on the abnormal deformation area annotation.

[0027] S1023, obtaining a training data set based on a plurality of image slices of preset sizes and a true value mask of an abnormal deformation area corresponding to each image slice.

[0028] In this application, the historical GB-SAR images can be divided into multiple image slices of preset sizes to obtain richer training samples, and then the true value mask of the abnormal deformation area of ​​each image slice is generated based on the abnormal deformation area annotation. The image slice and the corresponding true value mask of the abnormal deformation area are used as a training pair, and multiple training pairs are used as training data sets.

[0029] S103, training a preset model based on the training data set to obtain a bridge bank slope deformation abnormal area identification model.

[0030] In the present application, the image slices of the preset size can be used as a single-channel floating-point matrix of the preset size, and the true value mask of the abnormal deformation area corresponding to the image slices of the preset size can be used as a binary single-channel binary matrix of the preset size, which can be input into the preset model to train the preset model.

[0031] In the present application, in order to augment the training data set, it is proposed that: the image slices of the preset size and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size are randomly flipped, and the flipped image slices and the true value masks of the abnormal deformation areas corresponding to the image slices are used as training data.

[0032] In the present application, in order to augment the training data set, it is also proposed: randomly rotate the image slices of the preset size and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size, and use the rotated image slices and the true value masks of the abnormal deformation areas corresponding to the image slices as training data.

[0033] S104, inputting the GB-SAR image to be identified of the bridge bank slope in the target study area into the bridge bank slope deformation abnormal area identification model to obtain the bridge bank slope deformation abnormal area identification result.

[0034] In the present application, the preset model includes: four layers of downsampling encoders, each layer of downsampling encoders, with an SE module of the same size added, the SE module is used to calculate the weights of each channel of the corresponding layer, and recalculate its output value according to the weight as the final output of the encoder.

[0035] Specifically, this application designs an improved deep learning model SEUNet based on the classic UNet structure for identifying abnormal deformation areas of bridge bank slope GB-SAR images. SEUNet integrates the Squeeze-and-Excitation (SE) module to enhance feature extraction capabilities, and uses the PyTorch framework and its Lightning module for training and verification.

[0036] The core of SEUNet's design is to combine the encoding-decoding architecture of UNet with the channel attention mechanism of SE module. In this application, the SEUNet model structure includes the following key parts: Encoder part: The encoder consists of four encoding layers, each of which includes two convolutional layers followed by an SE module. The convolutional layer uses a 3x3 convolution kernel, Batch Normalization and ReLU activation function. The SE module forms channel attention through adaptive average pooling, adaptive fully connected layer and Sigmoid activation function to enhance the expression ability of the feature layer. Each encoding layer is downsampled through the maximum pooling layer.

[0037] Center layer: Located between the encoder and decoder, the center layer uses a higher number of convolution channels (1024 channels) for feature aggregation.

[0038] Decoder: It consists of four decoding layers, symmetrical with the encoder. Each decoding layer is upsampled by bilinear interpolation and its result is concatenated with the corresponding encoding layer features. Each decoding layer contains two convolutional layers, using the same convolution kernel, batch normalization and ReLU activation.

[0039] Output layer: Finally, a convolutional layer is used to map the decoder output to the target number of output channels.

[0040] The deep convolutional neural network UNet is a convolutional neural network architecture commonly used in image segmentation tasks. Its characteristic is that it uses an encoder-decoder structure to effectively capture features of different scales and semantics. The entire network can be regarded as a symmetrical U-shaped structure, so it is named UNet. Figure 2 is the basic structure diagram of the UNet network, which is divided into two parts. The left side is the encoder structure, which is used to extract input feature information, and the right side is the decoder structure, which outputs the encoded features in the form of an image.

[0041] (1) Encoder part: It consists of an input layer, a downsampling module, and an encoder module. The input layer accepts the image of the model as the input of the model; the downsampling module consists of a series of convolutional layers (usually a combination of convolution, batch normalization, and activation functions) and pooling groups, which are used to gradually reduce the size and number of channels of the feature map so as to gradually extract higher-level feature information; the encoder module reuses multiple downsampling modules to gradually reduce the size and number of channels of the feature map. Each encoder module usually adopts a downsampling module and a skip connection to connect the feature map of the previous level to the next level for feature fusion in the decoder.

[0042] (2) Decoder part: It consists of an upsampling module, a decoder, and an output layer. The upsampling module consists of a series of upsampling operations (such as deconvolution and transposed convolution) and convolution operations, which are used to gradually increase the size and number of channels of the feature map so as to gradually restore the resolution and retain more detail information. The decoder module reuses multiple upsampling modules to gradually increase the size and number of channels of the feature map. Each decoder module usually uses an upsampling module, a jump connection, and a fusion operation (such as concatenation and weighted fusion) to fuse the feature map from the encoder with the feature map of the current decoder. The last layer of the output layer is a convolution layer, which is used to generate the final segmentation result. Usually, the number of channels of the output layer is equal to the number of categories of the task, and an appropriate activation function is applied to generate a probability distribution of each pixel belonging to each category.

[0043] The working principle of the UNet network is as follows: In the encoder part on the left side of the UNet, the input image is first convolved with a 3×3 convolution kernel, and then an activation function is used to obtain a feature map. The feature extraction result is max pooled (downsampled) with a 2×2 pooling kernel. Each downsampled image becomes half of its original size, and the number of convolution kernels is doubled. The subsequent process is similar, and each layer of the image is subjected to two convolutions and two activation functions to extract image features, and then downsampled, and repeated 4 times. In the decoder part on the right side of the UNet, the feature matrix is ​​deconvolved with a 2×2 convolution kernel to achieve the upsampling process. The image size of each upsampling layer is doubled, and the number of channels is reduced by half. When the encoder downsamples, the image is cropped to the same size and then spliced ​​to reduce data loss because deconvolution can only expand the image but cannot restore it. This method can increase the feature layer so that the output on the encoder can be directly connected to the decoder for further propagation. After the matrix is ​​connected, the new feature matrix is ​​convolved with 3×3. Each layer will undergo an upsampling process, and then extract features through two convolutions and two activation functions, and repeat the above operation four times. In the last step, two 3×3 convolution kernels are selected to reduce the number of 128 feature channels to 1. UNet is a typical end-to-end network structure, which means that the input is an image and the output is also an image.

[0044] In order to enhance the feature extraction efficiency of the traditional UNet model, this application integrates the SE module component in its basic structure to enhance the feature extraction capability of the model. The SE module (Squeeze-and-Excitation Block) is an attention mechanism designed to improve the feature extraction capability of convolutional neural networks. It dynamically adjusts the channel weights of the feature map through adaptive learning of the relationship between channels, thereby improving network performance while keeping the network lightweight. The schematic diagram is shown in the figure below. Figure 3As shown. The core idea of ​​the SE module mechanism is to dynamically redistribute the importance of each channel through "squeeze" and "excitation" operations, thereby enhancing the feature representation capability. The working mechanism of the SE module can be divided into three main steps: squeeze, excitation, and recalibration. The specific process and mathematical formula are as follows: (1) Squeeze operation: In this stage, each feature map is compressed into a single value through global average pooling to capture global spatial information. This operation can effectively reduce the spatial dimension and retain only the information of the channel dimension, thereby achieving global semantic understanding of the feature map. , first perform global average pooling to compress the spatial dimension to the channel dimension. The output is: (1)

[0045] in, is the feature representation of the cth channel, H and W are the height and width of the feature map, C is the number of channels, and i, j represent the i-th row and j-th column of the feature map respectively.

[0046] (2) Excitation operation: In this stage, the SE module recalibrates the squeezed features through a pair of fully connected layers. First, a fully connected layer is used to map the features to a low-dimensional space, forming a bottleneck, and nonlinear characteristics are introduced through the ReLU activation function. Subsequently, the second fully connected layer maps the features back to the original channel dimension and limits the output value between 0 and 1 through the Sigmoid function. This process generates a weight coefficient for each channel, indicating the importance of each channel.

[0047] (2)

[0048] in and is a learnable weight matrix, represents the ReLU activation function, represents the Sigmoid activation function, z is the feature representation generated after the compression operation, and r is the scaling factor, which represents the scaling ratio of the channel.

[0049] (3) Recalibration operation: Finally, these weight coefficients are multiplied by the original feature map channel by channel to generate a recalibrated feature map. This step realizes the dynamic adjustment of the original features, allowing the network to pay more attention to important feature channels and suppress unimportant channels. The formula is as follows: (3)

[0050] The above formula re-weights the features of each channel, thereby enhancing the network's focus on key features. Through this mechanism, the SE module effectively improves the network's feature expression ability and classification performance.

[0051] For the traditional UNet model, there are four layers of downsampling encoders to extract features of different scales of the image. To improve its feature extraction efficiency, for each layer of the downsampling encoder, add a SE module of the same size to calculate the weights of each channel, and recalculate its output value according to the weight as the final output of the encoder to improve its feature extraction efficiency and the generalization ability of the model. The resulting SEUNet network structure is as follows: Figure 4 shown.

[0052] In the training process of the SEUNet model, this application selects BCEWithLogitsLoss as the loss function. This loss function combines the binary cross entropy loss (Binary Cross Entropy Loss) and the Sigmoid activation function, and is suitable for binary classification tasks. The advantage of BCEWithLogitsLoss is that it takes into account both the numerical stability and computational efficiency of the predicted value when calculating the loss, and does not require the Sigmoid activation function to be explicitly added to the network output layer. For each sample, the calculation formula of BCEWithLogitsLoss is: (4)

[0053] in, is the original output of the model, i.e. the log-odds without Sigmoid activation, and is the true label 0 or 1. Represents the Sigmoid function: (5)

[0054] In the above formula, The output of the model is converted into a probability value between [0,1]. The loss function is calculated for each pixel and the loss of all pixels is averaged to obtain the loss value of the entire image.

[0055] In order to quantitatively evaluate the recognition effect of the model, this application uses three indicators, namely accuracy, recall, and intersection over union (IoU), to quantitatively analyze the accuracy of the model evaluation results. The calculation formulas for accuracy and recall are as follows: (6) (7)

[0056] In formula (6) and formula (7), TP, TN, FP and FN are true positive examples, true negative examples, false positive examples and false negative examples respectively.

[0057] The intersection over union (IoU), also known as the Jaccard index, is often used to measure the similarity between two sets to determine the prediction effect of the model. Its calculation formula and schematic diagram are as follows: Figure 5 shown.

[0058] Specific embodiments are provided in this application to further illustrate the technical solutions provided in this application: The research area in the specific embodiments of this application is located near a place in the Tibet Autonomous Region of China, where the Nujiang River Bridge is being built. In order to monitor the stability of the bridge bank slope during construction, a GB-SAR surface displacement monitoring system was deployed there to monitor the surface deformation of the Lhasa bank and collect GB-SAR image data.

[0059] In this application, Python is used to parse the cumulative deformation field source file puredefodata.bin generated by GB-SAR to generate GeoTIFF images (such as Figure 6 QGIS software was used to manually mark the abnormal deformation areas in the GeoTIFF image (as shown in Figure 7 As shown in the figure), the image is divided into 256×256 slices to generate the corresponding abnormal deformation area mask. In the generated dataset, the input of the model is a single-channel floating-point matrix of size 256×256, as shown in the figure. Figure 8 As shown in the first row, the corresponding abnormal deformation area (true value) is a binary 256×256 single-channel binary matrix (such as Figure 8 The second row shows the white area, where the abnormal deformation area is marked.

[0060] In order to improve the generalization ability and robustness of the model, increase data diversity, and prevent model overfitting, data augmentation schemes are often used to expand the amount of training set data during model training. This application adopts a data augmentation method based on random flipping and rotation. Specifically, this application implements an augmentation class called RandomFlipRotate, which performs a series of random geometric transformations on the input image and the corresponding anomaly area each time it is called.

[0061] First, the random flipping in the data augmentation process includes horizontal flipping and vertical flipping. For each input image and label, these flipping operations have a 50% probability of being applied. Horizontal flipping, that is, flipping along the width direction of the image, can help the model better adapt to the changes in the left and right structures of the image; vertical flipping is flipping along the height direction, which increases the training samples of the changes in the upper and lower structures of the image. These operations not only increase the diversity of the data, but also enable the model to learn the characteristics of symmetry and structures in different directions, thereby improving the prediction ability of unseen data. Secondly, the random rotation operation provides more transformation possibilities for images and labels. In this application, random rotations of 0 degrees, 90 degrees, 180 degrees and 270 degrees are introduced, each of which has a 50% probability of occurring. This rotation enhances the model's adaptability to inputs in different directions by adjusting the direction of the image. In particular, in practical applications, images may come from different shooting angles and directions, and random rotation simulates this uncertainty to improve the robustness of the model to real-world data.

[0062] Based on the above method, the two images with scan numbers 24300 and 66205 collected by the GB-SAR system were annotated respectively, and used as training sets and test sets for model training and testing. The information of the constructed dataset is shown in Table 1.

[0063] Table 1 Dataset information table used by the model

[0064] The proposed SEUNet model is constructed using PyTorch, and the PyTorch Lightning framework is used for training and evaluation. The loss function curves of the obtained model on the training set and test set are shown in Figure 2. Fig. 9 As shown in the figure, the Loss curve of the proposed SEUNet model on the training set and the test set converges more quickly, and has fewer mutation points, and performs better on the test set, indicating that the SEUNet model proposed in this application has better generalization ability.

[0065] Table 2 Model training parameter settings

[0066] The trained SEUNet model is used to detect abnormal deformation areas on GB-SAR images scanned at different times in the study area. The abnormal areas in the images can be extracted and highlighted with red marks, as shown in Figure 10. It can be seen from parts (a)-(d) of Figure 10 that the abnormal deformation area of ​​the bridge bank slope in the study area continues to expand over time, and the deformation is relatively concentrated. By analyzing the GB-SAR images at different times in Figure 10, it can be seen that the effectiveness of the SEUNet model in detecting abnormal deformation areas. The red area in the figure clearly identifies the deformation characteristics and dynamic changes of the bridge bank slope over time. First, in part (a) of Figure 10 on October 2, 2023, there is almost no abnormal deformation area in the figure, indicating that the structural changes of the bridge bank slope during this period are relatively slight. However, as time progresses to December 20, 2023 (part (b) of Figure 10), it can be observed that the deformation area is significantly enlarged, especially at the key position of the bridge bank, and the deformation trend gradually emerges. This change may be related to natural factors (such as rainfall, flooding) or human activities (such as construction), which affects the stability of the rock and soil. By March 9, 2024 (part (c) of Figure 10), the deformation area continued to expand and deepen, indicating that the deformation not only expanded in scope but also intensified in degree. At this time, it becomes particularly important to monitor and evaluate the structural safety of the bridge bank slope, and a comprehensive analysis combined with other geological monitoring data is required to determine the cause of the deformation and take appropriate stabilization measures. Finally, on May 22, 2024 (part (d) of Figure 10), the abnormal deformation area further expanded, showing a highly concentrated deformation phenomenon. This continuous deformation may pose a threat to the safety of the bridge and surrounding infrastructure.

[0067] In general, the application of the SEUNet model in this study successfully achieved accurate identification and dynamic monitoring of deformation areas. This technical means can provide an important reference for early warning and prevention of geological disasters, and also provide an effective example for future monitoring work under similar terrain conditions. Future research can combine multi-source remote sensing data and ground-based measured data to further improve the prediction accuracy and reliability of the model.

[0068] In order to evaluate the effectiveness of the SEUNet algorithm compared to the traditional UNet algorithm, this application trained and tested the two models on the same training set and test set. The results are shown in Table 3. As can be seen from the table, SEUNet performs better than UNet in multiple evaluation indicators.

[0069] First, in terms of performance on the training set, SEUNet's accuracy is 0.983, slightly higher than UNet's 0.980. This shows that SEUNet exhibits better generalization ability when learning the features of the training data. In terms of recall rate, SEUNet is 0.939, which is also higher than UNet's 0.924, showing that SEUNet has a certain advantage in correctly identifying positive samples. The IoU (Intersection over Union) indicator is an indicator that measures the degree of overlap between the predicted result and the true result. SEUNet achieved 0.888, while UNet achieved 0.865, further proving the prediction accuracy advantage of SEUNet on the training set.

[0070] The results on the test set also show that SEUNet performs well. The accuracy of SEUNet is 0.976, which is significantly higher than UNet's 0.866, with an improvement of 12.70%, indicating that SEUNet can still maintain a high level of prediction accuracy on new data. In terms of recall, SEUNet is 0.890, which is higher than UNet's 0.811, with an improvement of 9.74%, which means that SEUNet can better capture positive samples on the test set. For the IoU of the test set, SEUNet achieved 0.792, which is 3.26% higher than UNet's 0.767, further illustrating the robustness of SEUNet in processing unseen data. In summary, the performance of the proposed SEUNet on the test set is significantly better than that of the traditional UNet model, indicating that SEUNet has relatively good generalization ability when conducting deformation anomaly analysis on the GB-SAR displacement field of the bridge bank slope.

[0071] Overall, SEUNet's overall superior performance in training and testing may be attributed to its improvements in network structure, such as deeper network layers, more effective feature extraction mechanisms, or better parameter optimization strategies. These improvements make SEUNet more capable of capturing complex data patterns. At the same time, SEUNet exhibits high stability and generalization capabilities in both training and testing phases, which is crucial for model deployment and performance maintenance in actual application scenarios.

[0072] Table 3 Comparison of model training and prediction results

[0073] Figure 11 compares the performance of SEUNet and UNet models in the object detection task. Fig.11It can be clearly seen that SEUNet presents more regular and coherent features when identifying the target area. In contrast, UNet produces many scattered and irregular dots during the detection process, which makes the results appear messy. This difference highlights the superiority of SEUNet in maintaining the overall structure of the detection area and can more accurately capture the shape and boundaries of the target. However, there are a small number of holes in the detection results of both models. Specifically, the holes in some areas of SEUNet are slightly larger than those of UNet. Although SEUNet can better identify the overall morphology of the target, these holes may affect the integrity and accuracy of the results. This phenomenon shows that in future research and improvements, the hole problem of SEUNet can be optimized, and the occurrence of holes may be reduced by improving the network structure or training strategy, thereby further improving the detection performance and application value of the model.

[0074] Therefore, the SEUNet model proposed in this application, which is improved based on the SE module attention mechanism, can be used to identify abnormal deformation areas in SAR images of bridge bank slope foundations. By integrating the SE module attention mechanism into the encoder of the traditional UNet model, the model's ability to capture the importance of feature channels is enhanced, and the efficiency and effect of feature extraction are improved. This application constructs a GB-SAR image dataset containing bridge bank slope deformation data, and uses data annotation and data augmentation methods to enrich training samples, improve the generalization ability of the model, and provide a reliable data foundation for model training.

[0075] The experimental results show that the SEUNet model performs well in identifying abnormal deformation areas. On the test set, the model's accuracy reached 97.6%, the recall rate was 89.0%, and the intersection-over-union ratio was 79.2%, all of which were better than the traditional UNet model. The accuracy rate was improved by 12.70%, showing higher detection accuracy and stability. This shows that the SEUNet model can effectively and automatically extract abnormal deformation areas from GB-SAR images, and can be applied to real-time monitoring and early warning of bridge bank slopes, providing strong technical support for engineering safety.

[0076] Based on the same inventive concept, the present application also provides a device for identifying abnormal deformation areas of bridge bank slopes, the device comprising: The acquisition module is used to obtain the historical GB-SAR images corresponding to the deformation data of the bridge bank slope in the target study area; A labeling module is used to label the historical GB-SAR images, obtain the true value of the abnormal deformation area corresponding to the historical GB-SAR images, and obtain a training data set based on a plurality of historical GB-SAR images labeled with the true value of the abnormal deformation area; A training module, used to train a preset model based on the training data set to obtain a bridge bank slope deformation abnormal area identification model; The identification module is used to input the GB-SAR image to be identified of the bridge bank slope in the target study area into the bridge bank slope deformation abnormal area identification model to obtain the bridge bank slope deformation abnormal area identification result.

[0077] Optionally, the marking module is used to: Obtain annotations of abnormal deformation areas in historical GB-SAR images; Dividing the historical GB-SAR image into a plurality of image slices of preset sizes; Generate a true value mask of the abnormal deformation area of ​​each image slice based on the abnormal deformation area annotation; A training data set is obtained based on image slices of multiple preset sizes and true value masks of abnormal deformation areas corresponding to each image slice.

[0078] Optionally, a training module is used to: The image slices of the preset size are used as single-channel floating-point matrices of the preset size, and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size are used as single-channel binary matrices of the preset size in binary format, which are input into the preset model to train the preset model.

[0079] Optionally, the device further comprises: A flipping module is used to randomly flip the image slices of the preset size and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size, and use the flipped image slices and the true value masks of the abnormal deformation areas corresponding to the image slices as training data.

[0080] Optionally, the device further comprises: A rotation module is used to randomly rotate the image slices of the preset size and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size, and use the rotated image slices and the true value masks of the abnormal deformation areas corresponding to the image slices as training data.

[0081] Optionally, the preset model includes: four layers of downsampling encoders, each layer of downsampling encoders, with an SE module of the same size added, the SE module is used to calculate the weights of each channel of the corresponding layer, and recalculate its output value according to the weight as the final output of the encoder.

[0082] Based on the same inventive concept, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes, the steps in the method for identifying abnormal deformation areas of bridge bank slopes as described in any of the above embodiments are implemented.

[0083] Based on the same inventive concept, the present application also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method for identifying abnormal deformation areas of bridge bank slopes described in any of the above embodiments are implemented.

[0084] Based on the same inventive concept, the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the method for identifying abnormal deformation areas of bridge bank slopes described in any of the above embodiments.

[0085] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0086] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, devices, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0087] The present application is described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (apparatus), and computer program products according to the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0088] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0089] These computer program instructions can also be loaded onto a computer or other programmable terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0090] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0091] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0092] The above is a detailed introduction to a method for identifying abnormal deformation areas of bridge bank slopes provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for identifying abnormal deformation areas of bridge bank slopes, characterized in that: The method comprises: Obtain historical GB-SAR images corresponding to the deformation data of the bridge bank slope in the target study area; Annotating the historical GB-SAR images to obtain true values ​​of abnormal deformation regions corresponding to the historical GB-SAR images, and obtaining a training data set based on a plurality of historical GB-SAR images annotated with the true values ​​of abnormal deformation regions; The preset model is trained based on the training data set to obtain a bridge bank slope deformation abnormal area recognition model; The GB-SAR image to be identified of the bridge bank slope in the target study area is input into the bridge bank slope deformation abnormal area identification model to obtain the bridge bank slope deformation abnormal area identification result.

2. The method for identifying abnormal deformation areas of bridge bank slope according to claim 1, characterized in that: The historical GB-SAR images are annotated to obtain true values ​​of abnormal deformation regions corresponding to the historical GB-SAR images, and a training data set is obtained based on a plurality of historical GB-SAR images annotated with true values ​​of abnormal deformation regions, including: Obtain annotations of abnormal deformation areas in historical GB-SAR images; Dividing the historical GB-SAR image into a plurality of image slices of preset sizes; Generate a true value mask of the abnormal deformation area of ​​each image slice based on the abnormal deformation area annotation; A training data set is obtained based on image slices of multiple preset sizes and true value masks of abnormal deformation areas corresponding to each image slice.

3. The method for identifying abnormal deformation areas of bridge bank slope according to claim 2, characterized in that: Training a preset model based on the training data set includes: The image slices of the preset size are used as single-channel floating-point matrices of the preset size, and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size are used as single-channel binary matrices of the preset size in binary format, which are input into the preset model to train the preset model.

4. The method for identifying abnormal deformation areas of bridge bank slope according to claim 3 is characterized in that: Before inputting the training data into the preset model, the method further includes: The image slices of the preset size and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size are randomly flipped, and the flipped image slices and the true value masks of the abnormal deformation areas corresponding to the image slices are used as training data.

5. The method for identifying abnormal deformation area of ​​bridge bank slope according to claim 3 or 4, characterized in that: Before inputting the training data into the preset model, the method further includes: The image slices of the preset size and the true value masks of the abnormal deformation areas corresponding to the image slices of the preset size are randomly rotated, and the rotated image slices and the true value masks of the abnormal deformation areas corresponding to the image slices are used as training data.

6. The method for identifying abnormal deformation areas of bridge bank slopes according to any one of claims 1 to 4, characterized in that: The preset model includes: four layers of downsampling encoders, each layer of downsampling encoders is added with an SE module of the same size, the SE module is used to calculate the weights of each channel of the corresponding layer, and recalculate its output value according to the weight as the final output of the encoder.

7. A device for identifying abnormal deformation areas of bridge bank slopes, characterized in that: The bridge bank slope deformation abnormal area identification device comprises: The acquisition module is used to obtain the historical GB-SAR images corresponding to the deformation data of the bridge bank slope in the target study area; A labeling module is used to label the historical GB-SAR images, obtain the true value of the abnormal deformation area corresponding to the historical GB-SAR images, and obtain a training data set based on a plurality of historical GB-SAR images labeled with the true value of the abnormal deformation area; A training module, used to train a preset model based on the training data set to obtain a bridge bank slope deformation abnormal area identification model; The identification module is used to input the GB-SAR image to be identified of the bridge bank slope in the target study area into the bridge bank slope deformation abnormal area identification model to obtain the bridge bank slope deformation abnormal area identification result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for identifying abnormal deformation areas of bridge bank slopes according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for identifying abnormal deformation areas of bridge bank slopes as described in any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the method for identifying abnormal deformation areas of bridge bank slopes as described in any one of claims 1 to 6 are implemented.

Images