Geological radar image automatic identification method, equipment and medium

Abstract

The invention discloses a geological radar image automatic identification method and device and a medium, and relates to the technical field of image processing, and the geological radar image identification process comprises data acquisition, preprocessing, 3D point cloud fusion, semantic segmentation, neural network processing and real-time monitoring. The method comprises the following steps of: firstly, collecting geological radar image data and preprocessing the geological radar image data to generate sample data of which noise in an image is removed, then fusing the sample data into a 3D point cloud, and carrying out semantic segmentation to obtain multi-modal data; thirdly, inputting the data into a neural network model, and extracting a single-stage target detection feature map; then, the feature map is sent to a candidate region generation network, and a depth feature map of the detection target is obtained; and finally, identifying the position and category of the target based on the depth feature map, carrying out real-time monitoring, and outputting an identification result. According to the process, the recognition precision and efficiency are improved, manual intervention is reduced, and data multi-modality and algorithm robustness are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method, device and medium for automatic recognition of geological radar images. Background Art

[0002] Geological radar image recognition is an important technology in the field of geological exploration. It can help us understand underground structures, identify geological anomalies and potential disasters, etc. The main reasons for conducting geological radar image recognition include: Efficient non-destructive testing: Geological radar has the characteristics of high efficiency, non-destructiveness, strong penetration and high resolution, making it the mainstream method for tunnel lining disease detection. Improve safety and quality: Geological surveys and non-destructive testing of construction quality by geological radar can ensure construction safety and quality. Decision support: Accurate recognition of geological radar images is helpful for promoting advanced geological forecasting of tunnels, and can automatically identify unfavorable geological bodies in geophysical images. Problems with existing technologies in geological radar image recognition include: Traditional geological radar image recognition methods have low recognition accuracy, difficulty in identifying complex targets, cumbersome recognition process, and inability to achieve end-to-end recognition. The correct interpretation of ground penetrating radar is very complicated and requires analysis based on three-dimensional information. Converting to two-dimensional image analysis will lose a lot of information and require a large amount of prior information, which is difficult to obtain through machine learning. Geological radar images have multiple interpretations. Different technicians may get different results when interpreting the same image. The accuracy of identifying diseases mainly depends on the experience of technicians. The geological radar detection process will be affected by ground coupling effects, interference from surface blocks and strips, interference from buildings, and interference from high-voltage cables in the air. It is necessary to optimize the data acquisition process and perform image post-processing to suppress noise signals and amplify effective signals. Existing algorithms may have high requirements for signal-to-noise ratio, poor robustness, and cannot adapt to negative frequency signals, making them difficult to be applied in engineering. Geological radar image recognition technology still faces challenges in improving recognition accuracy, reducing dependence on manual interpretation, and enhancing algorithm robustness and adaptability. Summary of the invention

[0003] The technical problem to be solved by the present invention is that it is difficult to identify complex targets, the identification process is cumbersome, and the identification accuracy is low. The purpose is to provide a method, device and medium for automatic identification of geological radar images. The noise in the image can be effectively removed by wavelet transform, the image quality can be improved, and the important features of the image can be retained while removing impurities. By preprocessing and fusing the geological radar image data into 3D point cloud data, and then performing semantic segmentation, the geological structure and anomaly can be more accurately identified and located, and the accuracy and efficiency of identification can be improved. By inputting multimodal data into the neural network model, the geological situation can be understood and analyzed more comprehensively, and the utilization efficiency of data and the accuracy of analysis can be improved. The real-time performance of target detection is improved: through the single-stage target detection feature map and the candidate area generation network, the deep feature map of the detection target can be quickly identified, and the real-time monitoring of the target position and category can be realized. Through the continuous optimization of the deep learning model, the robustness and adaptability of the algorithm under different geological conditions are improved, so that it can work stably in a wider range of application scenarios. Through deep learning and multimodal data fusion, the accuracy of target detection is improved. Real-time monitoring of the detection target is realized, which is suitable for scenes requiring rapid response. Manual intervention is reduced and the ability of automatic processing is improved. The problem of how to effectively integrate and utilize data from different modalities has been solved, and the generalization ability of the model has been enhanced. The automated, accurate, and real-time analysis and monitoring of geological radar image data has been achieved, solving the limitations of traditional methods in terms of processing speed, accuracy, and automation.

[0004] The present invention is achieved through the following technical solutions: The first aspect of the present invention provides a method for automatic recognition of geological radar images, comprising the following specific steps: Acquire the collected image data, and use wavelet transform to decompose the processed image into multiple frequency levels, including high-frequency details and low-frequency approximate information at different levels; wherein the high-frequency details include the edges and other sharp details of the image, and the low-frequency approximate information includes the basic shape and structure information of the image; The image data decomposed into multiple frequency levels are respectively subjected to impurity removal to obtain processing results of multiple frequency levels; The processing results of multiple frequency levels are resynthesized through inverse wavelet transform to obtain sample data; The sample data is fused to obtain 3D point cloud data, and the 3D point cloud data is semantically segmented to obtain multimodal data; Input the multimodal data into the backbone network of the neural network model and extract the single-stage target detection feature map; Input the single-stage target detection feature map into the candidate region generation network to obtain the deep feature map of the detected target; Based on the extracted deep feature map, a region proposal network is constructed to identify the location of the detection target, and a joint sparse representation model is constructed to identify the category of the detection target; Based on the position and category of the identified detection target, the detection target is monitored in real time and the recognition result is output.

[0005] Furthermore, the resynthesizing of the processing results of multiple frequency levels by inverse wavelet transform specifically includes: Obtain wavelet coefficients of each frequency level obtained through wavelet transform, including approximate coefficients and detail coefficients; Determining a wavelet basis for inverse wavelet transform, wherein the wavelet basis for inverse wavelet transform matches the wavelet basis used when performing wavelet transform; Starting from the detail coefficients of the highest layer, inverse wavelet transform is performed layer by layer. At each layer, the detail coefficients and the approximate coefficients from the previous layer are combined using the wavelet basis to reconstruct the signal of the layer; After completing the inverse transform at all levels, the resulting approximation coefficients will represent the reconstructed signal.

[0006] Furthermore, the obtaining of multimodal data specifically includes: Get the 2D RGB image in the sample data and the three-dimensional coordinates of the binary mask predicted pixel points corresponding to the detection target; Predict the 3D coordinates of the pixel points based on the 2D RGB image and the binary mask corresponding to the detection target to obtain the predicted 3D point cloud data; Based on the predicted 3D point cloud data, a new 2D projection is generated and the loss is compared with the actual data. If the calculated loss is within the set threshold, the predicted 3D point cloud data is semantically segmented based on the fully convolutional neural network to obtain multimodal data.

[0007] Furthermore, the specific process of obtaining the single-stage target detection feature map includes: The multimodal data is subjected to convolution processing to obtain a first output; The convolution data is globally pooled through the average pooling layer to obtain the mean of each channel. The mean is compressed using the first custom network layer, inverted and expanded to make its size consistent with the output of the convolution block, to obtain the second output, which is a single-stage target detection feature map.

[0008] Furthermore, the process of obtaining the depth feature map of the detected target includes: Add the first output and the second output, and pass the added data through the relu function to obtain the third output; The relu function is used to clear the weak features of the noise, and the clearing process includes: Perform global pooling on the added data, use the second custom network layer to reconstruct the pooling result into single-channel data for two-dimensional convolution, then process the output through the fully connected layer to obtain the mapping relationship between different channels, and finally output it through the softplus function to obtain the fourth output; The third output is multiplied by the fourth output to obtain a depth feature map of the detected target.

[0009] Furthermore, the construction of the region proposal network to identify the location of the detection target specifically includes: Building a region proposal network based on the extracted deep feature map, and using the region proposal network to generate candidate regions that may contain objects; Classify and regress the candidate region to determine whether the region contains the detection target; If the candidate region contains the detection target, the specific category of the detection target is further determined; Based on the regression subnetwork, the bounding box of the candidate region is adjusted to obtain the location of the detected target.

[0010] Furthermore, the construction of the joint sparse representation model to identify the category of the detection target specifically includes: The selected multi-level deep features are characterized and analyzed using joint sparse representation to determine the target category of the input sample: Get the target category, assuming there is K associated categories to be represented, construct K associated dictionaries, and performing a joint sparse representation on the categories to be represented according to the dictionaries to obtain a joint sparse matrix; A norm correlation constraint is imposed on the joint sparse matrix, and each category solves the error in its own local dictionary to obtain the sum of the reconstruction errors of the K components; The target category is determined based on the minimum error of the sum of reconstruction errors.

[0011] Furthermore, the real-time monitoring of the detection target based on the identified position of the detection target and the category of the detection target specifically includes: The denoising diffusion probability model is used to model the distribution of sample data, and a disturbance term is introduced in the sampling process to generate training samples in the low probability density area; Use training samples to train the CNN network for anomaly detection, select specific layer features for reconstruction, and train to obtain a reconstruction model; In the reconstruction model, anomaly detection and positioning are achieved by comparing the original image and the reconstructed image. The specific detection process includes: calculating the differences between radar images acquired at different time points or under different conditions to identify the deformation area of ​​the detection target in the radar images at different times; by comparing the corresponding pixels of each image in the image set, the differences caused by external factors are extracted, and based on the image with the difference eliminated, the difference between two adjacent images is calculated to identify the deformation area.

[0012] A second aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a method for automatic recognition of geological radar images when executing the program.

[0013] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for automatic recognition of geological radar images.

[0014] Compared with the prior art, the present invention has the following advantages and beneficial effects: Wavelet transform can effectively remove noise from images and improve image quality. It can also retain important features of images while removing impurities. By preprocessing and fusing geological radar image data into 3D point cloud data and then performing semantic segmentation, geological structures and anomalies can be more accurately identified and located, and the accuracy and efficiency of identification can be improved. By inputting multimodal data into the neural network model, the geological situation can be understood and analyzed more comprehensively, and the utilization efficiency of data and the accuracy of analysis can be improved. Improve the real-time performance of target detection: Through the single-stage target detection feature map and candidate region generation network, the deep feature map of the detection target can be quickly identified, and the real-time monitoring of the target position and category can be realized. Through the continuous optimization of the deep learning model, the robustness and adaptability of the algorithm under different geological conditions are improved, so that it can work stably in a wider range of application scenarios. Through deep learning and multimodal data fusion, the accuracy of target detection is improved. Real-time monitoring of detection targets is realized, which is suitable for scenarios that require rapid response. Reduce manual intervention and improve the ability of automatic processing. Solve the problem of how to effectively fuse and utilize different modal data and enhance the generalization ability of the model. It realizes the automated, accurate and real-time analysis and monitoring of geological radar image data, and solves the limitations of traditional methods in processing speed, accuracy and degree of automation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings: Figure 1 4 is a flow chart of the automatic identification method in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0017] As a possible implementation, Figure 1 As shown, this embodiment provides a method for automatic recognition of geological radar images, including the following specific steps: obtaining geological radar image data, preprocessing the geological radar image data to obtain sample data; fusing the sample data to obtain 3D point cloud data, performing semantic segmentation on the 3D point cloud data to obtain multimodal data; inputting the multimodal data into the backbone network of the neural network model to extract a single-stage target detection feature map; inputting the single-stage target detection feature map into the candidate region generation network to obtain a depth feature map of the detection target; based on the extracted depth feature map, identifying the location and category of the detection target; based on the identified location of the detection target and the category of the detection target, real-time monitoring of the detection target is performed, and the recognition result is output. By preprocessing and fusing the geological radar image data into 3D point cloud data and then performing semantic segmentation, the geological structure and anomaly can be more accurately identified and located, and the accuracy and efficiency of recognition can be improved. By inputting the multimodal data into the neural network model, the geological situation can be more comprehensively understood and analyzed, and the utilization efficiency of the data and the accuracy of the analysis can be improved. Improve the real-time performance of target detection: Through the single-stage target detection feature map and candidate region generation network, the deep feature map of the detection target can be quickly identified, and the real-time monitoring of the target location and category can be realized. Through the continuous optimization of the deep learning model, the robustness and adaptability of the algorithm under different geological conditions are improved, so that it can work stably in a wider range of application scenarios. Through deep learning and multimodal data fusion, the accuracy of target detection is improved. Real-time monitoring of the detection target is realized, which is suitable for scenarios that require rapid response. Reduce manual intervention and improve the ability of automatic processing. Solve the problem of how to effectively integrate and utilize different modal data, and enhance the generalization ability of the model. Realize the automatic, accurate and real-time analysis and monitoring of geological radar image data, and solve the limitations of traditional methods in processing speed, accuracy and automation.

[0018] In some possible implementations, the geological radar image data is preprocessed, specifically including: first obtaining the image data to be processed, and using the wavelet transform technology to decompose the image into multiple frequency levels, including high-frequency details and low-frequency approximate information. High-frequency details capture the edges and sharp details of the image, while low-frequency approximate information contains the basic shape and structure of the image. The processed data of multiple frequency levels are resynthesized through inverse wavelet transform to obtain the final sample data. In this implementation, the noise in the image can be effectively removed by wavelet transform, and the image quality can be improved. While removing impurities, important features of the image, such as edge and structural information, are retained. The wavelet transform provides multi-scale image analysis, which helps to understand the image content at different levels and reduce the amount of data by removing unimportant information. Impurities in the image are effectively removed, while key visual information is retained, and finally the processed sample data is obtained through inverse transform.

[0019] In some possible implementations, in order to effectively obtain and process the wavelet coefficients of each frequency level obtained by wavelet transform and provide a basis for subsequent signal analysis and processing, it is necessary to re-synthesize the processing results of multiple frequency levels through inverse wavelet transform, and the specific process includes: obtaining the wavelet coefficients of each frequency level obtained by wavelet transform, including approximate coefficients and detail coefficients; determining the wavelet basis for inverse wavelet transform, wherein the wavelet basis for inverse wavelet transform matches the wavelet basis used when performing wavelet transform; starting from the detail coefficients of the highest level, inverse wavelet transform is performed layer by layer, and at each level, the detail coefficients and the approximate coefficients from the previous level are combined using the wavelet basis (the approximate coefficients represent the low-frequency or smooth part of the signal, while the detail coefficients represent the high-frequency or detail part of the signal) to reconstruct the signal of that level; after completing the inverse transform of all levels, the obtained approximate coefficients will represent the reconstructed signal.

[0020] In some possible implementations, the approximation coefficient (cA) represents the low-frequency components of the signal, usually reflecting the overall trend of the signal. The detail coefficient (cD) represents the high-frequency components of the signal, usually capturing the detail information of the signal. In multi-layer wavelet decomposition, the detail coefficient can be further divided into different levels.

[0021] In some possible implementations, multimodal data is obtained, specifically including: obtaining a 2D RGB image in the sample data and a binary mask corresponding to the detection target to predict the three-dimensional coordinates of the pixel points; using the 2D RGB image and the three-dimensional coordinates of the predicted pixel points to obtain predicted 3D point cloud data through an algorithm (such as stereo matching or multi-view geometry in deep learning); expanding the 2D image information to the 3D space, increasing the dimension and depth information of the data, and providing more information for subsequent analysis.

[0022] According to the predicted 3D point cloud data, a new 2D projection is generated and compared with the actual data for loss. If the calculated loss is within the set threshold, the predicted 3D point cloud data is semantically segmented based on the fully convolutional neural network to obtain multimodal data. The accuracy of the predicted 3D point cloud data is ensured by loss comparison, which improves the reliability of subsequent processing. Loss comparison usually uses a certain loss function (such as mean square error) to measure the difference between the predicted projection and the actual data. The calculated loss is compared with the set threshold. If the loss is within an acceptable range, that is, the loss is less than or equal to the threshold, the next step of processing is continued. Semantic segmentation of the predicted 3D point cloud data is performed based on a fully convolutional neural network (FCN) to identify and classify different objects and surfaces in the point cloud, thereby obtaining multimodal data. The multimodal data obtained by semantic segmentation can be used for more complex analysis and decision-making, such as object recognition, scene understanding, etc. Using a fully convolutional neural network for semantic segmentation can automate the processing flow and improve the accuracy of segmentation.

[0023] In some possible implementations, a single-stage target detection feature map is obtained, and the specific process includes: The multimodal data is input into the convolution layer for feature extraction. The convolution layer can capture the spatial features of the input data and generate the first output, which is the preliminary feature map; the convolution-processed data is globally pooled through the average pooling layer. Global average pooling can extract global features from each channel, that is, the spatial dimensions (height and width) of each channel are reduced to 1, and only the channel dimension is retained to obtain the mean feature of each channel; the mean after global average pooling is further processed using the first custom network layer. The specific operations of this custom network layer include compression, negation, and expansion operations: Compression: The global mean feature is compressed through one or more fully connected layers (or similar compression mechanisms) to reduce the feature dimension.

[0024] Negation: Negate the compressed features, which may be a means of feature enhancement to increase the discriminative ability of the model.

[0025] Extension: Expand the negated features back to the same size as the output of the convolutional block so that they can be fused or added to the output of the convolutional layer; The expanded second output is fused with the first output (the output of the convolutional layer) to obtain the final single-stage target detection feature map. This feature map contains both local spatial features and global feature information, providing rich feature representation for subsequent target detection tasks.

[0026] By combining local features with global features, the model's ability to identify targets is enhanced. The inversion operation helps to highlight certain features and improve the robustness of the model. Through the compression and expansion operations of the custom network layer, the dimension of the feature can be flexibly controlled to adapt to different network structures and detection tasks. Through the combination of the convolution layer and the global average pooling layer, as well as the processing of the custom network layer, a single-stage target detection feature map containing rich information is generated.

[0027] In some possible implementations, a depth feature map of a detection target is obtained, and the specific process includes: The first output (the output of the convolutional layer) and the second output (the global mean feature map processed by the custom network layer) are added together to obtain the third output; the local features and the global features are added together, and then multiplied by the mapping relationship between channels to achieve deep fusion of features; The added data is processed by the ReLU activation function. The function of the ReLU function is to retain the positive eigenvalues ​​and set the negative eigenvalues ​​to zero, which can increase nonlinearity and reduce the impact of noise. The combination of the ReLU activation function and the Softplus function can enhance important features and suppress noise. Global pooling: Perform global pooling on the added data to obtain the global features of each channel.

[0028] Reconstruction and 2D convolution: The second custom network layer is used to reconstruct the result of global pooling into single-channel data and perform 2D convolution, which helps to extract local features between channels.

[0029] Fully connected layer processing: The output of the two-dimensional convolution is processed through a fully connected layer to obtain the mapping relationship between different channels, which helps to further integrate feature information.

[0030] Softplus function output: Finally, the output of the fully connected layer is processed by the Softplus function, which is a smooth ReLU function that can output all eigenvalues ​​but assign smaller weights to smaller eigenvalues, thereby clearing the weak noise features and obtaining the fourth output. The nonlinear characteristics of ReLU and Softplus functions help the model capture more complex data distributions; Multiply the third output (the feature map after ReLU activation) and the fourth output (the feature map processed by the Softplus function) to obtain the final deep feature map of the detected target.

[0031] This process generates a detection target feature map containing depth information through the addition, activation, clearing and multiplication of feature maps, providing rich feature representation for subsequent target detection tasks.

[0032] In some possible implementations, constructing a region proposal network to identify the location of a detection target specifically includes: Building RPN based on deep feature maps: Using the deep feature maps extracted from the deep learning model, a region proposal network is built. The role of RPN is to quickly and efficiently generate candidate regions that may contain objects, which are called region proposals.

[0033] Generate candidate regions: RPN generates multiple candidate regions (anchors) for each pixel by sliding a small network on the feature map. These candidate regions are rectangular boxes of different scales and proportions that are used to cover areas in the image that may contain objects.

[0034] Classification and bounding box regression: Classify the candidate region to determine whether it contains an object, and perform bounding box regression to determine the exact location of the candidate region. This step involves two branches, one for classification (determining the positive and negative samples of the object) and the other for regression (adjusting the bounding box of the candidate region).

[0035] Determine the category of the detection target: If the candidate region is classified as containing the detection target, the specific category of the detection target is further determined through the classification network.

[0036] Bounding box adjustment: Based on the regression subnetwork, the bounding box of the candidate region is adjusted to obtain the precise location of the detected object. This process usually involves fine-tuning the coordinates of the candidate region so that the predicted bounding box is more consistent with the location of the real object.

[0037] The region proposal network can effectively identify the location of the detection target from the image and provide a basis for subsequent target classification and precise positioning, which not only improves the efficiency of target detection, but also enhances the accuracy of detection.

[0038] In some possible implementations, constructing a joint sparse representation model to identify the category of the detection target specifically includes: The selected multi-level deep features are characterized and analyzed using joint sparse representation to determine the target category of the input sample: Obtain the target category. Assuming that there are K associated categories to be represented, K associated dictionaries are constructed for these categories. Each dictionary contains training samples of the corresponding category, which are used to characterize the features of the input sample. According to these dictionaries, the categories to be represented are jointly sparsely represented to obtain a joint sparse matrix. Based on the joint sparse matrix, the input sample is projected on each dictionary to find the optimal sparse coefficient. Imposing norm correlation constraints on the joint sparse matrix helps to enhance the sparsity of the model, reduce data redundancy and noise interference, and each category performs error solution in its own local dictionary, that is, to find the best sparse representation in its own dictionary space to minimize the reconstruction error. Based on the above sparse representation, the reconstruction error of each category is calculated, that is, the difference between the input sample and the linear combination of the samples in the dictionary. The sum of the reconstruction errors of the K components is used as the basis for classification, that is, the sum of the reconstruction errors of different categories is compared, and the target category is determined based on the minimum error of the sum of the reconstruction errors. The category with the smallest reconstruction error is selected as the predicted category of the input sample, because samples of the same category should have a smaller reconstruction error in the corresponding dictionary. The joint sparse representation model can effectively utilize the multi-level feature information of samples and improve the accuracy of detecting target categories.

[0039] In some possible implementations, real-time monitoring of the detection target based on the identified location of the detection target and the category of the detection target specifically includes: The distribution of sample data is modeled using a denoising diffusion probability model (DDPM). This model corrupts the training data by continuously adding Gaussian noise, and then learns to recover the data by reversing the noise process. Introducing disturbance terms in the sampling process can generate training samples in low probability density areas, enhance the model's ability to identify anomalies, and use the training samples to train a convolutional neural network (CNN) for anomaly detection. The CNN network can learn the feature representation of the image to distinguish between normal and abnormal situations. Features are extracted and reconstructed at a specific network layer, and a reconstructed model is trained to help identify and locate abnormal areas. In the reconstruction model, anomaly detection and location are achieved by comparing the original image and the reconstructed image. The specific detection process includes calculating the difference between radar images acquired at different time points or under different conditions to identify the deformation area of ​​the detection target in the radar image at different times. By comparing the corresponding pixels of each image in the image set, the differences caused by external factors are extracted. Based on the image with the difference removed, the difference between two adjacent images is calculated to obtain the deformation area. Based on the above process, the detection target can be monitored in real time and the recognition result can be output. This method can effectively identify and locate abnormal changes in the detection target, can effectively monitor the detection target in real time, promptly discover and respond to abnormal situations, and improve the accuracy and efficiency of monitoring.

[0040] As a possible implementation, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, a method for automatic recognition of geological radar images is implemented.

[0041] As a possible implementation, this embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, a method for automatic recognition of geological radar images is implemented.

[0042] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for automatic recognition of geological radar images, characterized in that: The specific steps include: Acquire the collected image data, and use wavelet transform to decompose the processed image into multiple frequency levels, including high-frequency details and low-frequency approximate information at different levels; wherein the high-frequency details include the edges and other sharp details of the image, and the low-frequency approximate information includes the basic shape and structure information of the image; The image data decomposed into multiple frequency levels are respectively subjected to impurity removal to obtain processing results of multiple frequency levels; The processing results of multiple frequency levels are resynthesized through inverse wavelet transform to obtain sample data; The sample data is fused to obtain 3D point cloud data, and the 3D point cloud data is semantically segmented to obtain multimodal data; Input the multimodal data into the backbone network of the neural network model and extract the single-stage target detection feature map; Input the single-stage target detection feature map into the candidate region generation network to obtain the deep feature map of the detected target; Based on the extracted deep feature map, a region proposal network is constructed to identify the location of the detection target, and a joint sparse representation model is constructed to identify the category of the detection target; Based on the position and category of the identified detection target, the detection target is monitored in real time and the recognition result is output.

2. The method for automatic recognition of geological radar images according to claim 1, characterized in that: The resynthesizing of the processing results of multiple frequency levels by inverse wavelet transform specifically includes: Obtain wavelet coefficients of each frequency level obtained through wavelet transform, including approximate coefficients and detail coefficients; Determining a wavelet basis for inverse wavelet transform, wherein the wavelet basis for inverse wavelet transform matches the wavelet basis used when performing wavelet transform; Starting from the detail coefficients of the highest layer, inverse wavelet transform is performed layer by layer. At each layer, the detail coefficients and the approximate coefficients from the previous layer are combined using the wavelet basis to reconstruct the signal of the layer; After completing the inverse transform at all levels, the resulting approximation coefficients will represent the reconstructed signal.

3. The method for automatic recognition of geological radar images according to claim 1, characterized in that: The obtaining of multimodal data specifically includes: Get the 2D RGB image in the sample data and the three-dimensional coordinates of the binary mask predicted pixel points corresponding to the detection target; Predict the 3D coordinates of the pixel points based on the 2D RGB image and the binary mask corresponding to the detection target to obtain the predicted 3D point cloud data; Based on the predicted 3D point cloud data, a new 2D projection is generated and the loss is compared with the actual data. If the calculated loss is within the set threshold, the predicted 3D point cloud data is semantically segmented based on the fully convolutional neural network to obtain multimodal data.

4. The method for automatic recognition of geological radar images according to claim 1, characterized in that: The specific process of obtaining a single-stage target detection feature map includes: The multimodal data is subjected to convolution processing to obtain a first output; The convolution data is globally pooled through the average pooling layer to obtain the mean of each channel. The mean is compressed using the first custom network layer, inverted and expanded to make its size consistent with the output of the convolution block, to obtain the second output, which is a single-stage target detection feature map.

5. The method for automatic recognition of geological radar images according to claim 4, characterized in that: The process of obtaining the depth feature map of the detected target includes: Add the first output and the second output, and pass the added data through the relu function to obtain the third output; The relu function is used to clear the weak features of the noise, and the clearing process includes: Perform global pooling on the added data, use the second custom network layer to reconstruct the pooling result into single-channel data for two-dimensional convolution, then process the output through the fully connected layer to obtain the mapping relationship between different channels, and finally output it through the softplus function to obtain the fourth output; The third output is multiplied by the fourth output to obtain a depth feature map of the detected target.

6. The method for automatic recognition of geological radar images according to claim 1, characterized in that: The step of constructing a region proposal network to identify the location of the detection target specifically includes: Building a region proposal network based on the extracted deep feature map, and using the region proposal network to generate candidate regions that may contain objects; Classify and regress the candidate region to determine whether the region contains the detection target; If the candidate region contains the detection target, the specific category of the detection target is further determined; Based on the regression subnetwork, the bounding box of the candidate region is adjusted to obtain the location of the detected target.

7. The method for automatic recognition of geological radar images according to claim 1, characterized in that: The step of constructing a joint sparse representation model to identify the category of the detection target specifically includes: The selected multi-level deep features are characterized and analyzed using joint sparse representation to determine the target category of the input sample: Get the target category, assuming there is K associated categories to be represented, construct K associated dictionaries, and performing a joint sparse representation on the categories to be represented according to the dictionaries to obtain a joint sparse matrix; A norm correlation constraint is imposed on the joint sparse matrix, and each category solves the error in its own local dictionary to obtain the sum of the reconstruction errors of the K components; The target category is determined based on the minimum error of the sum of reconstruction errors.

8. The method for automatic recognition of geological radar images according to claim 1, characterized in that: The real-time monitoring of the detection target based on the identified position of the detection target and the category of the detection target specifically includes: The denoising diffusion probability model is used to model the distribution of sample data, and a disturbance term is introduced in the sampling process to generate training samples in the low probability density area; Use training samples to train the CNN network for anomaly detection, select specific layer features for reconstruction, and train to obtain a reconstruction model; In the reconstruction model, anomaly detection and positioning are achieved by comparing the original image and the reconstructed image. The specific detection process includes: calculating the differences between radar images acquired at different time points or under different conditions to identify the deformation area of ​​the detection target in the radar images at different times; by comparing the corresponding pixels of each image in the image set, the differences caused by external factors are extracted, and based on the image with the difference eliminated, the difference between two adjacent images is calculated to identify the deformation area.

9. 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 program, the method for automatic recognition of geological radar images as described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for automatic recognition of geological radar images as described in any one of claims 1 to 8 is implemented.

Images