Geological radar image abnormal region identification method and device, equipment and storage medium

Abstract

The invention provides a geological radar image abnormal area identification method and device, equipment and a storage medium, and relates to the technical field of geological engineering and artificial intelligence crossing. The method comprises the following steps: performing feature extraction on a target geological radar image to be identified to obtain a feature map, and performing numeralization on construction background information corresponding to the target geological radar image to obtain construction background features; carrying out feature fusion on the feature map and the construction background features to obtain fusion features; and inputting the fusion features into a trained geological radar image abnormal region recognition model to obtain an abnormal region recognition result of the target geological radar image. According to the invention, the abnormal region analysis and identification speed and accuracy of the geological radar image can be improved.

Description

Technical Field

[0001] The present invention relates to the cross - technical field of geological engineering and artificial intelligence, and particularly to a method, device, equipment and storage medium for identifying abnormal areas in ground penetrating radar images. Background Technique

[0002] Ground penetrating radar imaging technology, also known as Ground Penetrating Radar (GPR), is a non - invasive geophysical method that uses high - frequency electromagnetic waves to detect the structure and properties of underground media, and can be widely applied to fields such as geological exploration, tunnel construction, mine exploitation, and urban underground facility detection.

[0003] Existing ground penetrating radar imaging technology usually can only generate a single radar image, which requires experts to analyze these images by combining various information such as construction background, so as to identify abnormal areas such as underground pipelines and cavities in the ground penetrating radar images. This process is time - consuming and laborious, and highly dependent on the experience and professional knowledge of experts. Therefore, how to improve the analysis speed and accuracy of ground penetrating radar imaging has become an urgent technical problem to be solved. Summary of the Invention

[0004] Embodiments of the present invention provide a method, device, equipment and storage medium for identifying abnormal areas in ground penetrating radar images to solve the problem that the current analysis and identification of ground penetrating radar images are time - consuming and laborious.

[0005] In a first aspect, embodiments of the present invention provide a method for identifying abnormal areas in ground penetrating radar images, including:

[0006] Performing feature extraction on a target ground penetrating radar image to be identified to obtain a feature map, and numericalizing the construction background information corresponding to the target ground penetrating radar image to obtain construction background features;

[0007] Performing feature fusion on the feature map and the construction background features to obtain fusion features;

[0008] Inputting the fusion features into a trained model for identifying abnormal areas in ground penetrating radar images to obtain an identification result of the abnormal areas in the target ground penetrating radar image.

[0009] In a possible implementation manner, the performing feature extraction on a target ground penetrating radar image to be identified to obtain a feature map includes:

[0010] Performing initial convolution extraction on the target ground penetrating radar image to be identified based on an initial convolution layer to obtain an initial feature map;

[0011] Gradually extract the initial feature map based on multiple Bottleneck modules to obtain the intermediate feature map output by the last Bottleneck module;

[0012] Perform global average pooling operation on the intermediate feature map based on the global average pooling layer to obtain the global average pooling map;

[0013] Perform a fully connected operation on the global average pooling map based on the first fully connected layer to obtain a feature map.

[0014] In a possible implementation, the construction background information includes geological type and / or construction method;

[0015] Numericalize the construction background information corresponding to the target ground penetrating radar image to obtain construction background features, including:

[0016] Based on all categories corresponding to the geological type and / or all methods corresponding to the construction method, perform one-hot encoding on the geological type and / or construction method corresponding to the target ground penetrating radar image to obtain construction background features.

[0017] In a possible implementation, perform feature fusion on the feature map and the construction background features to obtain fusion features, including:

[0018] Map the construction background features to a low-dimensional feature space based on the second fully connected layer to obtain low-dimensional construction background features;

[0019] Perform feature fusion on the feature map and the low-dimensional construction background features to obtain fusion features.

[0020] In a possible implementation, the mapping of the construction background features to a low-dimensional feature space based on the second fully connected layer to obtain low-dimensional construction background features includes:

[0021] Based on B = φ(W b x b + b b ), obtain low-dimensional construction background features;

[0022] where B is the low-dimensional construction background feature, φ() is the activation function, W b is the weight matrix, x b is the construction background feature, and b b is the bias vector.

[0023] In a possible implementation, perform feature fusion on the feature map and the low-dimensional construction background features to obtain fusion features, including:

[0024] Concatenate the feature map and the low-dimensional construction background features in the feature dimension to obtain fused features.

[0025] In a second aspect, an embodiment of the present invention provides a device for identifying abnormal regions in ground penetrating radar images, including:

[0026] A processing module, configured to extract features from a target ground penetrating radar image to be identified to obtain a feature map, and digitize the construction background information corresponding to the target ground penetrating radar image to obtain construction background features;

[0027] A fusion module, configured to perform feature fusion on the feature map and the construction background features to obtain fused features;

[0028] An identification module, configured to input the fused features into a trained ground penetrating radar image abnormal region identification model to obtain an abnormal region identification result of the target ground penetrating radar image.

[0029] In a third aspect, an embodiment of the present invention provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the method in the first aspect or any possible implementation manner of the first aspect above is implemented.

[0030] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium that stores a computer program, and when the computer program is executed by a processor, the method in the first aspect or any possible implementation manner of the first aspect above is implemented.

[0031] In a fifth aspect, an embodiment of the present invention provides a computer program product, including a computer program, and when the computer program is executed by a processor, the method in the first aspect or any possible implementation manner of the first aspect above is implemented.

[0032] In the embodiments of the present invention, by first extracting features from a target ground penetrating radar image to be identified to obtain a feature map, and digitizing the construction background information corresponding to the ground penetrating radar image to obtain construction background features, and then performing feature fusion on the feature map and the construction background features to obtain fused features, so that by fusing the construction background information, the obtained fused features are helpful for enhancing the identification ability of the subsequent ground penetrating radar image abnormal region identification model. Then, the fused features are input into a trained ground penetrating radar image abnormal region identification model to obtain an abnormal region identification result of the target ground penetrating radar image, realizing fast and accurate identification of ground penetrating radar images, and significantly improving the efficiency of geological exploration and construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flowchart of the implementation of the method for identifying abnormal regions in ground penetrating radar images provided by the embodiments of the present invention;

[0034] Figure 2 It is a flowchart of the implementation of the method for identifying abnormal regions in ground penetrating radar images provided by another embodiment of the present invention;

[0035] Figure 3 It is a schematic structural diagram of the device for identifying abnormal regions in ground penetrating radar images provided by an embodiment of the present invention;

[0036] Figure 4 It is a schematic diagram of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0037] Next, the embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0038] See Figure 1 , which shows a flowchart of the implementation of the method for identifying abnormal regions in ground penetrating radar images provided by an embodiment of the present invention, and is described in detail as follows:

[0039] Step 101: Extract features from the target ground penetrating radar image to be recognized to obtain a feature map, and numericalize the construction background information corresponding to the target ground penetrating radar image to obtain construction background features.

[0040] Exemplarily, before extracting features from the target ground penetrating radar image to be recognized and numericalizing the construction background information corresponding to the target ground penetrating radar image, data collection and preprocessing can be performed.

[0041] Among them, the purpose of data preprocessing is to clean and transform the original data so that it is suitable for the training and inference requirements of the model. This can include preprocessing of ground penetrating radar images.

[0042] Among them, data collection includes the collection of ground penetrating radar images and the collection of construction background information. Among them:

[0043] Collection of ground penetrating radar images: Collect a large number of ground penetrating radar images under different geological conditions, and label the target areas (i.e., abnormal areas, such as underground pipelines, cavities, etc.) by professionals.

[0044] Collection of construction background information: Collect the construction background information corresponding to the ground penetrating radar image, including geological type, construction method, historical data, etc.

[0045] Among them, the preprocessing of ground penetrating radar images includes:

[0046] Normalization processing: Normalize the image pixel values to the range of [0, 1] to accelerate the convergence of model training:

[0047]

[0048] Among them, Inorm is the pixel value of the normalized image, ranging from [0, 1], I is the pixel value of the original image, I min and I max are the minimum and maximum values of the pixel values.

[0049] Data augmentation: Expand the dataset by methods such as rotation, translation, scaling, adding noise, etc., to prevent the model from overfitting.

[0050] Through data collection and preprocessing, the data quality can be improved, noise and bias can be reduced, the generalization ability of the model can be enhanced, and overfitting can be prevented.

[0051] After data collection and preprocessing, exemplarily, feature extraction is performed on the target ground penetrating radar image to be recognized to obtain a feature map, which may include:

[0052] Performing initial convolution extraction on the target ground penetrating radar image to be recognized based on the initial convolutional layer to obtain an initial feature map.

[0053] Performing step-by-step extraction on the initial feature map based on multiple Bottleneck modules to obtain an intermediate feature map output by the last Bottleneck module.

[0054] Performing global average pooling operation on the intermediate feature map based on the global average pooling layer to obtain a global average pooling map.

[0055] Performing a fully connected operation on the global average pooling map based on the first fully connected layer to obtain a feature map.

[0056] In this embodiment, high-level feature representations are extracted from the preprocessed ground penetrating radar image through feature extraction, and these features will be used for subsequent target detection and recognition.

[0057] Exemplarily, the MobileNetV3 model can be used as the feature extraction network, and its specific process includes the following steps:

[0058] (1), Initial convolution to extract low-level features:

[0059] Using a standard convolutional layer (i.e., the initial convolutional layer) to extract initial features (i.e., the initial feature map):

[0060] F 1 = Conv(I norm );

[0061] where, F 1 is the initial feature map, Conv() is the convolution operation for extracting the initial features of the image, and I norm is the normalized input image.

[0062] (2), Stacking Bottleneck modules

[0063] Adopt the bottleneck module in MobileNetV3, gradually increase the model depth and the number of channels, and the output of each module is:

[0064] F n+1 = Bottleneck(F n );

[0065] where F n+1 is the intermediate feature map output by the (n + 1)-th bottleneck module, Bottleneck() is the bottleneck module, including depthwise separable convolution and activation function, and F n is the intermediate feature map output by the n-th bottleneck module and serves as the input to the (n + 1)-th bottleneck module.

[0066] In this embodiment, through multiple stacks of the bottleneck module, gradual extraction of features and enhancement of the expression ability can be achieved.

[0067] Among them, in some bottleneck modules, a Squeeze-and-Excitation (SE) module can be embedded to enhance the channel attention of features:

[0068] F SE = SE(F n ) ⊙ F n ;

[0069] where F SE is the intermediate feature map enhanced by the SE module, SE(F n ) is the channel attention vector calculated by the SE module, ⊙ is the element-wise multiplication operation, and F n is the intermediate feature map output by the n-th bottleneck module and serves as the input to the SE module.

[0070] Among them, an efficient h-swish activation function can be used to improve the computational performance and non-linear expression ability:

[0071]

[0072] where h-swish(x) is the output of the h-swish activation function, x is the input value, and ReLU6(x + 3) is the ReLU6 function, which limits the output range to [0, 6].

[0073] (3) Global pooling and fully connected layer generate the final output:

[0074] Among them, the global average pooling formula is:

[0075] F global = GlobalAvgPool(F n );

[0076] Among them, F global is the feature vector after global average pooling, that is, the global average pooling graph. GlobalAvgPool() is the global average pooling operation, which averages the spatial dimensions (height and width) of the feature map to generate a feature vector of fixed length. F n is the intermediate feature map output by the nth bottleneck module and serves as the input to the global average pooling layer.

[0077] Among them, the formula for the fully connected layer (i.e., the first fully connected layer) to generate the final feature vector is:

[0078] F = FC(F global );

[0079] Among them, F is the output feature vector of the first fully connected layer, that is, the feature map. F ∈ R H×W×C , H and W are the height and width of the feature map, and C is the number of channels.

[0080] Exemplarily, the cross-entropy loss and the optimization algorithm Adam can be used to train the MobileNetV3 model and evaluate its performance on the validation set and the test set.

[0081] In this embodiment, by performing feature extraction on the target ground penetrating radar image to be recognized, the spatial and semantic features of the target ground penetrating radar image can be extracted, information such as the shape and texture of the target can be captured, a rich feature representation can be provided, and subsequent multi-modal fusion and target detection tasks can be supported.

[0082] Exemplarily, the construction background information may include geological types and / or construction methods.

[0083] Numericalize the construction background information corresponding to the target ground penetrating radar image to obtain construction background features, which may include:

[0084] Based on all categories corresponding to the geological type and / or all methods corresponding to the construction method, perform one-hot encoding on the geological type and / or construction method corresponding to the target ground penetrating radar image to obtain construction background features.

[0085] In this embodiment, in order to convert text or categorical construction background information into numerical features, one-hot encoding can be used for representation.

[0086] For example, when numericalizing the construction background information corresponding to the target ground penetrating radar image:

[0087] First, determine all categories, collect all possible categories in the construction background information, and establish a category set.

[0088] For example, assume there are 4 types of geological types: clay (class1), sand (class2), rock (class3), and gravel (class4).

[0089] Secondly, assign a one-hot vector to each category: each category corresponds to a vector of length N (total number of categories), where the position corresponding to the category is 1 and the rest are 0.

[0090] The formula is expressed as: v i =[v i1 ,v i1 ,…,v iN , where

[0091] Specific example: clay (class1): v 1 =[1,0,0,0], sand (class2): v 2 =[0,1,0,0], rock (class3): v 3 =[0,0,1,0], gravel (class4): v 4 =[0,0,0,1].

[0092] Exemplarily, for the construction background information corresponding to the target ground penetrating radar image, according to the category corresponding to its geological type, the corresponding one-hot vector is assigned. For example, assume the geological type in the construction background information corresponding to the target ground penetrating radar image is "sand", then the construction background feature obtained after numericalization is v 2 =[0,1,0,0].

[0093] In this embodiment, by numericalizing the construction background information, the construction background information can be understood and utilized by the model, providing additional context information.

[0094] Step 102: Perform feature fusion on the feature map and the construction background feature to obtain a fused feature.

[0095] In this embodiment, fusing the construction background information with the image features can improve the recognition accuracy of the model using multi-modal information. Among them, the construction background information provides additional context and can assist the model in making more accurate judgments.

[0096] Optionally, performing feature fusion on the feature map and the construction background feature to obtain a fused feature may include:

[0097] Mapping the construction background feature to a low-dimensional feature space based on the second fully connected layer to obtain a low-dimensional construction background feature.

[0098] Perform feature fusion on the feature map and the low-dimensional construction background feature to obtain a fused feature.

[0099] Exemplarily, mapping the construction background features to a low-dimensional feature space based on the second fully connected layer to obtain low-dimensional construction background features may include:

[0100] Based on B = φ(W b x b + b b ), the low-dimensional construction background features are obtained.

[0101] Among them, B is the low-dimensional construction background feature, B ∈ R D , φ() is an activation function, such as ReLU, h-swish, etc., W b is a weight matrix, the size can be D × N, D is the dimension of the mapped feature, x b is the construction background feature, that is, the background information represented by the one-hot vector, and b b is the bias vector.

[0102] In this embodiment, considering that the dimension of the one-hot vector may be relatively high, it can be mapped to a lower-dimensional feature space through a fully connected layer.

[0103] Exemplarily, fusing the feature map and the low-dimensional construction background features to obtain fused features may include: concatenating the feature map and the low-dimensional construction background features in the feature dimension to obtain fused features.

[0104] In this embodiment, based on B = φ(W b x b + b b ), the numerical construction background information is mapped to the feature vector B ∈ R D through the fully connected layer to realize the embedding of the construction background information, and then based on F fusion = [F; B], F and B are concatenated in the feature dimension. Or, it can also be based on F fusion = αF + (1 - α)B, F and B are weighted and fused, where α is a learnable weight parameter, between 0 and 1.

[0105] In this embodiment, by fusing multi-modal information, when the subsequent model performs recognition, it not only depends on the image features but also can utilize the construction background information, improving the recognition accuracy, and at the same time can provide a richer feature representation to support the training of the subsequent recognition model.

[0106] Exemplarily, before performing the feature extraction in step 101 and the feature fusion in step 102, the feature extraction network and the multi-modal fusion module can be trained first, and the training strategy can be:

[0107] 1) Adopt a multi-task learning strategy to train the feature extraction network and the multi-modal fusion module simultaneously to achieve end-to-end model training.

[0108] 2) Train the model using cross - entropy loss and the Adam optimizer, setting appropriate learning rates (such as 0.001) and momentum parameters.

[0109] 3) Use a learning rate decay strategy, such as reducing the learning rate when the performance on the validation set does not improve.

[0110] The training steps can be as follows:

[0111] 1) Forward propagation:

[0112] Input the ground - penetrating radar image I and the construction background information x b .

[0113] Extract the feature map F through the MobileNetV3 feature extraction network.

[0114] The construction background information x b Pass through the embedding layer to get B.

[0115] Use the multi - modal feature fusion module to generate the fused feature F fusion .

[0116] 2) Calculate the loss

[0117] Calculate the cross - entropy classification loss according to the true labels:

[0118]

[0119] where L cls is the classification loss, N is the number of samples, p i is the probability that the i - th sample is predicted as the positive class, is the true class label of the i - th sample (positive class is 1, negative class is 0).

[0120] The regression loss (smooth L 1 loss) formula:

[0121]

[0122] where L reg is the regression loss, N reg is the number of positive samples for the regression loss, is the true class label of the i - th candidate region (non - background class is 1, background class is 0), t i is the regression parameter predicted by the model for the i - th candidate region, is the true regression target parameter of the i - th candidate region.

[0123] The smooth L 1 loss definition:

[0124]

[0125] Among them, is the smooth L1 loss function, which is used to reduce the influence of outliers. x is the regression error, that is, the difference between the predicted value and the true value.

[0126] Calculate the total loss L total :

[0127] L total = L cls + λL reg ;

[0128] Among them, L total is the total loss, and λ is the balance factor, which is used to adjust the weights of the classification loss and the regression loss, and is usually set to 1.

[0129] 3) Backpropagation and parameter update

[0130] Calculate the gradient of the loss with respect to the model parameters, and use the optimizer to update the model parameters.

[0131] 4) Iterative training

[0132] Repeat the above steps until the model converges or reaches the preset number of training epochs.

[0133] After the training is completed, the model can be tested for performance evaluation. Among them, the evaluation metrics for performance evaluation can include:

[0134] Precision: Measures the proportion of samples predicted as positive by the model that are actually positive:

[0135]

[0136] Among them, TP is True Positives, that is, true positives, which represents the number of samples correctly predicted as positive by the model. FP is False Positives, that is, false positives, which represents the number of samples incorrectly predicted as positive by the model (actually negative).

[0137] Recall: Measures the proportion of samples that are actually positive and are correctly predicted as positive by the model:

[0138]

[0139] Among them, FN is False Negatives, that is, false negatives, which represents the number of samples incorrectly predicted as negative by the model (actually positive).

[0140] Mean Average Precision (mAP): The average precision of all classes:

[0141]

[0142] Among them, M is the total number of categories, and AP m is the average precision of the m-th category.

[0143] Intersection over Union (IoU): Measures the overlap between the predicted bounding box and the ground truth bounding box:

[0144]

[0145] The process of testing the model is as follows:

[0146] 1. Use an independent test set and input it into the model for prediction.

[0147] 2. Compare the prediction results with the ground truth annotations and calculate the evaluation metrics.

[0148] After passing the test, it can be used for feature extraction and feature fusion to facilitate the input into the subsequent ground penetrating radar image anomaly region recognition model for anomaly region recognition, thereby improving the speed and accuracy of anomaly region recognition in ground penetrating radar images.

[0149] Step 103: Input the fused features into the trained ground penetrating radar image anomaly region recognition model to obtain the anomaly region recognition result of the target ground penetrating radar image.

[0150] Exemplarily, the fused feature F fusion can be recognized based on a deep neural network (DNN) and a multi-layer perceptron (MLP).

[0151] Among them, the fused feature F fusion is a feature vector synthesized from the feature map F and the low-dimensional construction background feature B, and this vector can contain multi-modal data from both the image and background information.

[0152] In the multi-modal fusion module, after concatenation or weighted fusion, a fused feature vector F fusion is obtained, and its dimension is D fusion .

[0153] To recognize F fusion , first, a multi-layer perceptron (MLP) can be used. It is a neural network composed of multiple fully connected layers and is used to learn the complex relationships in the fused features.

[0154] Among them, the input layer receives the fused feature F fusion , and the feature dimension is D fusion .

[0155] Among them, the hidden layer: Feature mapping is performed through multiple hidden layers to extract high-level semantic information. Each hidden layer h kThe calculation process is as follows:

[0156] h k = σ(W k h k-1 + b k );

[0157] Among them, h k-1 is the output of the previous hidden layer, h 0 = F fusion That is, the input layer, W k is the weight matrix of the k-th layer, b k is the bias term, and σ() is the activation function. Commonly used ones include the ReLU activation function: σ(x) = max(0, x), and it can also be the h-swish activation function, etc.

[0158] Among them, the output layer is used to predict the final recognition result and is usually used for classification problems. Assuming a multi-classification task is being carried out, the calculation of the output layer is as follows:

[0159] y = softmax(W out h k + b out );

[0160] Among them, W out is the weight matrix of the output layer, b out is the bias term of the output layer, and the softmax function converts the neural network output into a probability distribution:

[0161]

[0162] Among them, x i is the score of the i-th class, and N is the number of classes.

[0163] To further illustrate the geological radar image abnormal area recognition method provided by the embodiments of the present invention, in combination with the above steps 101 to 103, the specific steps of geological radar image abnormal area recognition can be:

[0164] 1. The geological radar image is preprocessed and feature extracted to obtain the feature map F, and the construction background information is numerically processed and mapped to obtain the low-dimensional construction background feature B.

[0165] 2. Adopt a splicing or weighted fusion strategy to combine the feature map F and the low-dimensional construction background feature B into the fusion feature F fusion .

[0166] 3. Through the DNN training model, input the fusion feature F fusion , and the output is the predicted class or the target area position.

[0167] 4. To improve the model accuracy, the identified fusion feature Ffusion Perform forward propagation by importing the above model training steps.

[0168] Exemplarily, when applying the ground penetrating radar image anomaly region recognition method provided by the embodiments of the present invention, model deployment is performed:

[0169] Among them, model export: Save the trained model parameters and structure in a deployment format, such as ONNX.

[0170] Deployment environment:

[0171] Hardware: GPU servers, edge computing devices, etc.

[0172] Software: Deployment frameworks (such as TensorRT), necessary dependency libraries.

[0173] Real-time inference:

[0174] Data input: Real-time acquired ground penetrating radar images and construction background information.

[0175] Data preprocessing: Keep consistent with the preprocessing method during training.

[0176] Model inference: Use the deployed model for forward propagation to obtain the recognition result.

[0177] Result display: Visualize the recognition result and provide it for engineers' reference.

[0178] Among them, in order to visualize the recognition result and implement user interaction and feedback, it is possible to:

[0179] Graphical interface: Design an intuitive and easy-to-use user interface that supports uploading ground penetrating radar image data and construction background information.

[0180] Dynamic optimization: Users can annotate and correct the recognition result through a feedback mechanism, and the system optimizes the model accordingly.

[0181] An embodiment of the present invention provides a rapid recognition technology for ground penetrating radar images combined with construction background information. By using the MobileNetV3 model for feature extraction and integrating construction background information, the MobileNetV3 model has lightweight and efficient feature extraction capabilities, can quickly process a large amount of radar imaging data, reduce calculation time, and the multi-modal feature fusion combined with construction background information improves the recognition accuracy of the model. Moreover, the model is easily deployable and has a user-friendly interface, enabling even non-professionals to use it efficiently. Thus, an artificial intelligence model can be used to intelligently analyze ground penetrating radar images and construction background information, intelligently identify abnormal areas in ground penetrating radar images, quickly obtain conclusions, significantly improve analysis efficiency and accuracy, provide real-time and accurate geological information support for construction, and achieve efficient, accurate, and easy-to-use ground penetrating radar image recognition. It can be widely applied to fields such as geological exploration, tunnel construction, mine exploitation, earthquake monitoring, and urban underground facility detection. It can efficiently identify underground abnormal structures in geological exploration, monitor construction risks in real time during tunnel construction, locate potential geological hazards during mine exploitation, and evaluate the safety of underground pipelines and other facilities during urban facility detection, having great market potential and application prospects.

[0182] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not imply the order of execution. The order of execution of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0183] The following is an apparatus embodiment of the present invention. For details not described in detail therein, reference may be made to the corresponding method embodiments above.

[0184] Figure 3 The structural schematic diagram of the ground penetrating radar image abnormal area recognition apparatus provided by the embodiment of the present invention is shown. For the sake of convenience of description, only the parts related to the embodiment of the present invention are shown and are described in detail as follows:

[0185] As Figure 3 shown, the ground penetrating radar image abnormal area recognition apparatus includes: a processing module 31, a fusion module 32, and an identification module 33.

[0186] The processing module 31 is configured to perform feature extraction on the target ground penetrating radar image to be recognized to obtain a feature map, and numericalize the construction background information corresponding to the target ground penetrating radar image to obtain construction background features;

[0187] The fusion module 32 is configured to perform feature fusion on the feature map and the construction background features to obtain fusion features;

[0188] The identification module 33 is configured to input the fusion features into a trained ground penetrating radar image abnormal area recognition model to obtain the abnormal area recognition result of the target ground penetrating radar image.

[0189] In a possible implementation, the processing module 31 can be used to perform initial convolution extraction on the target ground penetrating radar image to be recognized based on the initial convolution layer, so as to obtain an initial feature map; perform step-by-step extraction on the initial feature map based on a plurality of Bottleneck modules to obtain an intermediate feature map output by the last Bottleneck module; perform global average pooling operation on the intermediate feature map based on the global average pooling layer to obtain a global average pooling map; perform a fully connected operation on the global average pooling map based on the first fully connected layer to obtain a feature map.

[0190] In a possible implementation, the construction background information includes geological type and / or construction method; the processing module 31 can be used to perform one-hot encoding on the geological type and / or construction method corresponding to the target ground penetrating radar image based on all categories corresponding to the geological type and / or all methods corresponding to the construction method to obtain construction background features.

[0191] In a possible implementation, the fusion module 32 can be used to map the construction background features to a low-dimensional feature space based on the second fully connected layer to obtain low-dimensional construction background features; perform feature fusion on the feature map and the low-dimensional construction background features to obtain fusion features.

[0192] In a possible implementation, the fusion module 32 can be used to obtain low-dimensional construction background features based on B = φ(W b x b +b b ).

[0193] Wherein, B is the low-dimensional construction background feature, φ() is an activation function, W b is a weight matrix, x b is the construction background feature, and b b is a bias vector.

[0194] In a possible implementation, the fusion module 32 can be used to splice the feature map and the low-dimensional construction background features in the feature dimension to obtain fusion features.

[0195] Figure 4 is a schematic diagram of an electronic device provided by an embodiment of the present invention. As Figure 4 shown, the electronic device 4 in this embodiment includes: a processor 40 and a memory 41. The memory 41 stores a computer program 42. When the processor 40 executes the computer program 42, the steps in the above-mentioned various method embodiments are implemented. Alternatively, when the processor 40 executes the computer program 42, the functions of each module / unit in the above-mentioned various device embodiments are implemented.

[0196] Exemplarily, the computer program 42 can be divided into one or more modules / units, which are stored in the memory 41 and executed by the processor 40 to implement the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program 42 in the electronic device 4.

[0197] The electronic device 4 may include, but is not limited to, the processor 40 and the memory 41. Those skilled in the art can understand that Figure 4 merely examples of the electronic device 4, which do not constitute a limitation to the electronic device 4, may include more or fewer components than shown in the figure, or combine certain components, or different components. For example, the electronic device 4 may further include input / output devices, network access devices, buses, etc.

[0198] The processor 40 can be a central processing unit (CPU), or can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0199] The memory 41 can be an internal storage unit of the electronic device 4, such as the hard disk or memory of the electronic device 4. The memory 41 can also be an external storage device of the electronic device 4, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 4. Further, the memory 41 can also include both the internal storage unit and the external storage device of the electronic device 4. The memory 41 is used to store the computer program 42 and other programs and data required by the electronic device 4. The memory 41 can also be used to temporarily store the data that has been output or will be output.

[0200] For the convenience and simplicity of description, only the above division of each functional module / unit is used as an example. In practical applications, the above functions can be allocated to different functional modules / units according to needs. The above modules / units can be implemented in the form of hardware, or in the form of software, or in the form of a combination of hardware and software.

[0201] An embodiment of the present invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the methods in the above method embodiments are implemented.

[0202] An embodiment of the present invention also provides a computer program product including a computer program. When the computer program is executed by a processor, the methods in the above method embodiments are implemented.

[0203] Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0204] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments. Without special instructions and logical conflicts, the terms and / or descriptions between different embodiments are consistent and can be mutually referred to, and the technical features in different embodiments can be combined to form new embodiments according to their internal logical relationships.

[0205] The above-described embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for identifying abnormal areas in geological radar images, characterized in that: include: Extracting features from the target geological radar image to be identified to obtain a feature map, and digitizing the construction background information corresponding to the target geological radar image to obtain construction background features; Performing feature fusion on the feature map and the construction background feature to obtain a fusion feature; The fusion features are input into the trained geological radar image abnormal area recognition model to obtain the abnormal area recognition result of the target geological radar image.

2. The method for identifying abnormal areas in geological radar images according to claim 1, characterized in that: The feature extraction of the target geological radar image to be identified to obtain a feature map includes: Based on the initial convolution layer, the initial convolution extraction is performed on the target geological radar image to be identified to obtain the initial feature map; The initial feature map is gradually extracted based on multiple Bottleneck modules to obtain an intermediate feature map output by the last Bottleneck module; Performing a global average pooling operation on the intermediate feature map based on the global average pooling layer to obtain a global average pooling map; Based on the first fully connected layer, a fully connected operation is performed on the global average pooling map to obtain a feature map.

3. The method for identifying abnormal areas in geological radar images according to claim 1, characterized in that: The construction background information includes geological type and / or construction method; The construction background information corresponding to the target geological radar image is digitized to obtain construction background features, including: Based on all categories corresponding to the geological type and / or all methods corresponding to the construction method, the geological type and / or construction method corresponding to the target geological radar image are one-hot encoded to obtain construction background features.

4. The method for identifying abnormal areas in geological radar images according to claim 1, characterized in that: The feature map and the construction background feature are subjected to feature fusion to obtain fusion features, including: Mapping the construction background features to a low-dimensional feature space based on a second fully connected layer to obtain low-dimensional construction background features; The feature map and the low-dimensional construction background feature are subjected to feature fusion to obtain a fusion feature.

5. The method for identifying abnormal areas in geological radar images according to claim 4, characterized in that: The second fully connected layer maps the construction background features to a low-dimensional feature space to obtain low-dimensional construction background features, including: Based on B = φ (W b x b +b b ), and obtain low-dimensional construction background features; Among them, B is the low-dimensional construction background feature, φ() is the activation function, and W b is the weight matrix, x b is the construction background feature, b b is the bias vector.

6. The method for identifying abnormal areas in geological radar images according to claim 4, characterized in that: The feature map and the low-dimensional construction background feature are subjected to feature fusion to obtain fusion features, including: The feature map and the low-dimensional construction background feature are spliced ​​in the feature dimension to obtain a fusion feature.

7. A device for identifying abnormal areas in geological radar images, characterized in that: include: A processing module is used to extract features from the target geological radar image to be identified to obtain a feature map, and to digitize the construction background information corresponding to the target geological radar image to obtain construction background features; A fusion module, used for fusing the feature map with the construction background feature to obtain a fusion feature; The recognition module is used to input the fusion features into the trained geological radar image abnormal area recognition model to obtain the abnormal area recognition result of the target geological radar image.

8. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

10. A computer program product, characterized in that The method comprises a computer program, which implements the method according to any one of claims 1 to 6 when being executed by a processor.

Images