End-to-End 3D Ground Penetrating Radar Target Recognition Method and System Based on Deep Learning

Abstract

The present invention provides an end-to-end three-dimensional ground penetrating radar target recognition method and system based on deep learning, including: Step S1, acquiring and preprocessing the original echo information to obtain one-dimensional A-scan signals and C-scan horizontal slice maps, and forming a sample set; Step S2, respectively classifying and labeling the one-dimensional A-scan signal set and the C-scan image set; Step S3, dividing the sample set into a training set, a validation set and a test set; Step S4, respectively training a non-image domain model based on the A-scan signal and an image domain model based on the C-scan horizontal slice map; Step S5, reading the echo information collected by the three-dimensional ground penetrating radar, performing preprocessing, and classifying the non-image domain model; Step S6, generating a corresponding C-scan horizontal slice map cluster according to the detected target range, and performing target detection on the image domain model. The present invention has a small amount of calculation and a high model classification accuracy.

Description

Technical Field

[0001] The present invention relates to a method for three-dimensional ground penetrating radar target recognition, and more particularly to an end-to-end three-dimensional ground penetrating radar target recognition method based on deep learning, and further relates to a system adopting the end-to-end three-dimensional ground penetrating radar target recognition method based on deep learning. Background Art

[0002] In recent years, ground penetrating radar, as an efficient and non-destructive means for detecting underground diseases, has been known to more and more people. This solution emits high-frequency pulse signals into the ground, reflects the medium distribution differences in the underground space, and can further identify potential underground hazards. With its advantages of high resolution, low cost, and convenient operation, ground penetrating radar is increasingly widely used in the detection and maintenance of urban roads. Among them, compared with traditional two-dimensional radar, three-dimensional ground penetrating radar can synchronously collect data in the vertical and horizontal directions by emitting a single electromagnetic wave, and the obtained underground information is more abundant.

[0003] However, due to the complexity of radar images, high professional requirements are imposed on interpreters, and there are problems such as long time and easy misjudgment in manual interpretation. Therefore, it has become increasingly urgent to study an automatic recognition method for radar data. Among them, the data of three-dimensional ground penetrating radar includes: (a) one-dimensional A-scan signals received by emitting a single electromagnetic wave downward; (b) two-dimensional B-scan vertical slice images converted from a series of A-scan signals received by the radar moving along the survey line direction; (c) two-dimensional C-scan horizontal profile images obtained by synchronously collecting data between antennas in different channels along the horizontal direction. Currently, most of the research on automatic recognition of ground penetrating radar data relies on mature computer image recognition technology. The principle is to first convert the echo information of underground targets into image information, and then use the B-scan vertical slice images to train an image recognition model. However, since the echo information needs to be converted into an intuitive gray-scale stack image according to the amplitude value, the time required to process two-dimensional image data is longer and the computational amount is larger. Therefore, the original A-scan signal data with smaller computational amount and richer information has once again attracted attention.

[0004] The methods for signal recognition using the original A-scan signals are divided into two categories: traditional machine learning and deep learning. One is based on traditional machine learning methods. First, features in the time domain and frequency domain are extracted from single-channel A-scan signals, and then classifiers such as BP neural networks are used for recognition to achieve the recognition of underground targets. The other is to use a deep network model to extract depth information from adjacent multi-channel A-scan signals and perform classification, so as to achieve the purpose of target recognition. However, the following problems still exist in these two types of solutions in the prior art:

[0005] First, the models trained by traditional machine learning methods have room for improvement in terms of robustness to noise. Especially in urban roads, there are various interference factors such as clutter, which can easily cause misjudgments.

[0006] Second, using only one-dimensional A-scan signals can only detect underground anomalies, but cannot classify the detected targets. Since the A-scan signals of cavities and pipelines are quite similar, it is impossible to further distinguish the two only through A-scan signals. Therefore, the classification accuracy of the models trained in this way is not high and cannot meet the requirements of actual engineering projects.

[0007] Third, using only single A-scan signals or B-scan images cannot fully utilize the advantages of three-dimensional ground penetrating radar. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide an end-to-end three-dimensional ground penetrating radar target recognition method based on deep learning, which can not only achieve the recognition of ground penetrating radar targets with complex and rich information through a small amount of calculation, but also well overcome the problem that the actual engineering project requirements cannot be met due to the low classification accuracy of the model. On this basis, a system adopting the end-to-end three-dimensional ground penetrating radar target recognition method based on deep learning is further adopted.

[0009] For this reason, the present invention provides an end-to-end three-dimensional ground penetrating radar target recognition method based on deep learning, including the following steps:

[0010] Step S1, obtain the original echo information containing underground hidden targets on the survey line, preprocess the original echo information, and obtain multiple one-dimensional A-scan signals and multiple groups of C-scan horizontal slice images in the channel to form a corresponding sample set;

[0011] Step S2, classify and label the one-dimensional A-scan signal set and the C-scan image set respectively;

[0012] Step S3, divide each sample set into a training set, a validation set, and a test set according to a preset ratio;

[0013] Step S4, train a non-image domain model based on A-scan signals and an image domain model based on C-scan horizontal slice images respectively;

[0014] Step S5, read the echo information collected by the three-dimensional ground penetrating radar, preprocess it, and input it into the trained deep learning network model based on A-scan signals for non-image domain model classification;

[0015] Step S6: Generate a corresponding cluster of C-scan horizontal slice images based on the detected target range, input them into the trained deep learning network model based on C-scan horizontal slice images for target detection in the image domain model, exclude the misidentified results in the non-image domain model in step S5, and output the final classification result.

[0016] A further improvement of the present invention is that step S1 includes the following sub-steps:

[0017] Step S101: Obtain the original echo information B formed by the A-scan signals collected at each scanning position on the survey line. k , B k = Matrix(i,j), where i = 1,2…,m; j = 1,2…,n; n represents the number of scanning positions, m represents the number of sampling points; i represents the sampling points of each A-scan data in the depth direction, j represents the number of A-scan signals collected in the survey line direction, and k represents the channel of the current data collection.

[0018] Step S102: Preprocess the original echo matrix B containing underground targets. target The preprocessing includes direct wave removal processing, global background elimination processing, reverse energy attenuation gain processing, and normalization processing to form a preprocessed echo signal matrix.

[0019] Step S103: Extract the horizontal profile scan image containing underground targets on the survey line, intercept a preset number of horizontal slice images at a preset interval in the depth direction to form a C-scan horizontal slice combined image, which is used as a sample for training the image domain model.

[0020] A further improvement of the present invention is that step S102 includes the following sub-steps:

[0021] Step S1021: When the first amplitude value in the single-channel data is lower than the preset threshold, record this position as the termination point of the direct wave, and sequentially intercept 256 points after this termination point position as the data after removing the direct wave.

[0022] Step S1022: Eliminate the global background of the new matrix by subtracting the global average value from each data.

[0023] Step S1023: Perform reverse energy attenuation gain on the data in the matrix, multiply the new matrix by the normalized gain factor g(t) to achieve reverse energy attenuation gain of the data, where A(t) represents the fitting attenuation model, and max(A(t)) represents the maximum return value of the fitting attenuation curve.

[0024] Step S1024: Through the formula Perform data normalization on the data within the matrix, where 0 ≤ a′ ij ≤ 1, a ij represents the signal data of the i-th radar sampling point in the j-th A-scan, and a′ ij represents the normalized sampling point value, a max represents the radar signal data with the maximum amplitude, and a min represents the radar signal data with the minimum amplitude.

[0025] A further improvement of the present invention is that the step S2 includes the following sub-steps:

[0026] Step S201, label the single-channel A-scan signal, where the label for intact is 0, the label for underground hidden hazards is 1, and the label for manholes is 2;

[0027] Step S202, label the C-scan horizontal slice composite map, and the labeling categories include pipelines and cavities.

[0028] A further improvement of the present invention is that the step S4 includes the following sub-steps:

[0029] Step S401, train a non-image domain model using the A-scan signal set, and the non-image domain model is a deep learning model composed of two layers of one-dimensional convolutional neural networks and two layers of GRU neural networks;

[0030] Step S402, train an image domain model using the C-scan image set, the image domain model adopts the Yolov5 model, and during the training process, use the test set to further optimize the image domain model, and update and save the optimal image domain model in real time.

[0031] A further improvement of the present invention is that in the step S401, the two layers of one-dimensional convolutional neural networks include the first convolutional layer and the second convolutional layer with the same settings. The network input size of the first convolutional layer is 256×1, the number of convolutional filters is set to 128, the activation function adopts the ReLU function, and then it is connected to a max pooling layer with a pool size of 2, and the dropout layer is set to 0.2; after the first convolutional layer and the second convolutional layer, connect the first GRU layer with 256 units and the second GRU layer with 32 units. After each GRU layer, set the dropout layer to 0.2. After converting the features into a one-dimensional feature vector through the Flatten layer, connect it to a Dense layer with 128 units and adopt the ReLU activation function; the last layer is a Dense layer with 3 units, which is used to represent three classification labels respectively, and adopts the Softmax activation function.

[0032] A further improvement of the present invention lies in that, in the step S401, the parameter settings during the training of the non-image domain model are as follows: the sample number batch_size for model training is set to 50, and the number of epochs for model training is set to 70; during the training process, a categorical cross-entropy loss function and an Adam optimizer are adopted. After each epoch of training, the validation set is used for validation, and after validation, the test set is used to test the performance of the non-image domain model, and the optimal non-image domain model is updated and saved in real time.

[0033] A further improvement of the present invention lies in that the step S5 includes the following sub-steps:

[0034] Step S501: Use a three-dimensional ground penetrating radar device to conduct full-coverage detection of the area to obtain real radar data;

[0035] Step S502: Preprocess the collected echo information. The preprocessing includes direct wave removal processing, global background elimination processing, reverse energy attenuation gain processing, and normalization processing, and all A-scan signals in 16 channels are obtained;

[0036] Step S503: Input the processed one-dimensional A-scan signals into a deep learning network based on signals for classification of the non-image domain model;

[0037] Step S504: For the continuous A-scan signals with a classification result of 1 in 16 channels, determine them as potential areas of underground hidden dangers.

[0038] A further improvement of the present invention lies in that the step S6 includes the following sub-steps:

[0039] Step S601: For the potential areas determined as underground hidden dangers, extract horizontal slice maps at preset intervals in the depth direction to form a C-scan horizontal slice composite map;

[0040] Step S602: Input the C-scan horizontal slice composite map into the trained image domain model for object detection of the image domain model to exclude misrecognition results in the non-image domain model;

[0041] Step S603: Combine the results of the image domain model and the non-image domain model to output the final object recognition result. The target objects included in the output category in the target recognition result are intact, pipelines, manholes, and hidden dangers.

[0042] The present invention also provides an end-to-end three-dimensional ground penetrating radar object recognition system based on deep learning, which adopts the above-mentioned end-to-end three-dimensional ground penetrating radar object recognition method based on deep learning, and includes:

[0043] The sample set forming module is used to obtain the original echo information of the underground hidden danger target on the survey line, pre-process the echo information, and obtain multiple one-dimensional A-scan signals and multiple groups of C-scan horizontal slices in the channel to form a corresponding sample set;

[0044] A classification and annotation module is used to classify and annotate the one-dimensional A-scan signal set and the C-scan image set respectively;

[0045] A sample set division module is used to divide each sample set into a training set, a validation set and a test set according to a preset ratio;

[0046] A model training module, used to train a non-image domain model based on an A-scan signal and an image domain model based on a C-scan horizontal slice image;

[0047] The preprocessing and training module is used to read the echo information collected by the 3D ground penetrating radar, preprocess it, and input it into the trained deep learning network model based on A-scan signals to classify the non-image domain model;

[0048] The misidentification elimination module generates a corresponding C-scan horizontal slice image cluster according to the detected target range, and inputs it into the trained deep learning network model based on the C-scan horizontal slice image to perform target detection in the image domain model, eliminate the misidentification results in the non-image domain model of the preprocessing and training module, and output the final classification result.

[0049] Compared with the prior art, the beneficial effects of the present invention are: it can save the step of converting the original signal into a grayscale accumulation map, can directly classify the pre-processed signal, improve the speed of data processing and interpretation, and has low calculation cost and simple operation; on this basis, it also combines A-scan signal data and C-scan image data, so as to better distinguish the difference between pipelines and cavities, improve the accuracy of model classification, and give full play to the advantages of three-dimensional ground penetrating radar. The present invention can realize ground penetrating radar target recognition with complex and rich information through a very small amount of calculation, and can also well overcome the problem that the model classification accuracy is not high and cannot meet the needs of actual engineering projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the working process of an embodiment of the present invention;

[0051] Figure 2 The data of different dimensions collected by the three-dimensional ground penetrating radar in one embodiment of the present invention;

[0052] Figure 3Schematic diagram of obtaining one-dimensional A-scan samples from preprocessed original echo data in an embodiment of the present invention;

[0053] Figure 4 Schematic diagram of obtaining multiple groups of C-scan horizontal slice images from a radar horizontal sectional view in an embodiment of the present invention;

[0054] Figure 5 Schematic diagram of the structure of a non-image domain model trained based on one-dimensional A-scan signals in an embodiment of the present invention;

[0055] Figure 6 Accuracy / loss curve diagram obtained by training a non-image domain model based on CNN+GRU in an embodiment of the present invention;

[0056] Figure 7 Schematic diagram of the structure of an image domain model trained based on C-scan horizontal slice images in an embodiment of the present invention;

[0057] Figure 8 Target detection results obtained by training an image domain model based on Yolov5 in an embodiment of the present invention;

[0058] Figure 9 Schematic diagram of generating corresponding C-scan pictures according to the classification results of A-scan signals in an embodiment of the present invention. Detailed implementation manners

[0059] The following further describes in detail the preferred embodiments of the present invention with reference to the accompanying drawings.

[0060] In order to solve the deficiencies of the prior art, the present invention provides an end-to-end three-dimensional ground penetrating radar target recognition method and system based on deep learning. In its technical solution, first, a three-dimensional ground penetrating radar is used for data collection to obtain multiple A-scan signals and multiple groups of C-scan horizontal slice images; then, a trained A-scan signal recognition model, also known as a non-image domain model, is used to narrow down a large detection range. Since one-dimensional A-scan signals can only reflect the medium differences in the underground space, and the A-scan signals of cavities and pipelines are quite similar, it is impossible to further identify pipelines and cavities only using A-scan signals. Therefore, in subsequent steps, a trained C-scan image recognition model, also known as an image domain model, is further used to further discriminate the corresponding horizontal slice images, thereby effectively improving the accuracy of target recognition.

[0061] Regarding this, as Figure 1As shown in the figure, this embodiment provides an end-to-end three-dimensional ground penetrating radar target recognition method based on deep learning, including the following steps:

[0062] Step S1: Obtain the original echo information containing underground hidden danger targets on the survey line, preprocess the original echo information, and obtain multiple one-dimensional A-scan signals and multiple groups of C-scan horizontal slice maps in the channel to form a corresponding sample set;

[0063] Step S2: Classify and label the one-dimensional A-scan signal set and the C-scan image set respectively;

[0064] Step S3: Divide each sample set into a training set, a validation set, and a test set according to a preset ratio; the preset ratio refers to the ratio set according to actual needs, and the default setting is 6:2:2;

[0065] Step S4: Train a non-image domain model based on the A-scan signal and an image domain model based on the C-scan horizontal slice map respectively;

[0066] Step S5: Read the echo information collected by the three-dimensional ground penetrating radar, preprocess it, and input it into the trained deep learning network model based on the A-scan signal for classification in the non-image domain model;

[0067] Step S6: Generate a corresponding C-scan horizontal slice map cluster according to the detected target range, and input it into the trained deep learning network model based on the C-scan horizontal slice map for target detection in the image domain model, exclude the misrecognized results in the non-image domain model in step S5, and output the final classification result.

[0068] The step S1 in this embodiment includes the following sub-steps:

[0069] Step S101: Obtain the original echo information B formed by the A-scan signals collected at each scanning position on the survey line k , where the data acquisition method is to use a three-dimensional ground penetrating radar device to detect underground targets along the survey line and obtain three-dimensional radar data composed of scattered echoes received by the transmitted electromagnetic waves.

[0070] As Figure 2 shown, the three-dimensional radar data includes: (a) a single-channel one-dimensional A-scan signal transmitted downward; (b) two-dimensional B-scan data formed by splicing a series of A-scan signals received by the radar moving along the survey line direction; (c) C-scan data in the vertical and horizontal directions obtained by synchronously collecting data between different channels.

[0071] The original echo information B in this embodiment kis a two-dimensional matrix of size n*m composed of n scanning positions and m sampling points, B k = Matrix(i,j), where i = 1,2…,m; j = 1,2…,n; n represents the number of scanning positions, m represents the number of sampling points; i represents the sampling points of each A-scan data in the depth direction, j represents the number of A-scan signals collected in the survey line direction, k represents the channel of the current data collection, 0 < k ≤ 16;

[0072] Step S102, preprocess the original echo matrix B containing underground targets target The preprocessing includes direct wave removal processing, global background elimination processing, reverse energy attenuation gain processing, and normalization processing to form a preprocessed echo signal matrix;

[0073] Step S103, extract the horizontal profile scan map containing underground targets on the survey line, intercept a preset number of horizontal slice maps at preset intervals in the depth direction to form a C-scan horizontal slice combination map, which is used as a sample for training the image domain model. The preset interval is the interval in the depth direction set in advance, defaulting to 10 cm; the preset number is the number of horizontal slice maps set in advance, defaulting to 24; in practical applications, both the preset interval and the preset number can be custom-set and adjusted according to actual needs. As Figure 4 shown, in this embodiment, horizontal slice maps at different depths are stitched together into a C-scan horizontal slice combination map as a sample.

[0074] More specifically, the step S102 includes steps S1021 to S1024. The step S1021 is used to remove the direct wave from each channel of data in the current channel. When the first amplitude value in a single-channel data is lower than a preset threshold, the preset threshold refers to the amplitude threshold set in advance and can be custom-set and adjusted according to the actual situation. Record this position as the termination point of the direct wave, and sequentially intercept 256 points after this termination point position as the data after removing the direct wave; expressed by the formula: B′ target = Matrix(i′,j), where i′ = start:end; j = 1,2…,n; threshold is the preset threshold, a ij represents the signal data of the i-th radar sampling point in the j-th A-scan; loc(a ij ) represents the termination point of the direct wave.

[0075] Step S1022 in this embodiment is used to perform global background elimination on the new matrix. The signal background noise is suppressed by subtracting the global average value from each data, so as to effectively improve the signal-to-noise ratio and achieve global background elimination on the new matrix; it is expressed by the formula as:

[0076] Step S1023 in this embodiment is used to perform reverse energy attenuation gain on the data in the matrix. Multiply the new matrix by the normalized gain factor g(t) to achieve reverse energy attenuation gain on the data, which is expressed by the formula as: B′ new = B new * g(t); where, A(t) represents the fitting attenuation model, and max(A(t)) represents the maximum return value of the fitting attenuation curve.

[0077] Step S1024 in this embodiment performs data normalization processing on the data in the matrix through the formula where 0 ≤ a′ ij ≤ 1, a ij represents the signal data of the i-th radar sampling point in the j-th A-scan, a′ ij represents the normalized sampling point value, a max represents the radar signal data with the largest amplitude, and a min represents the radar signal data with the smallest amplitude.

[0078] As Figure 3 shown, in this embodiment, for the preprocessed echo signal matrix, along the direction of the trace data, n one-dimensional signal matrices with a size of 1*m are intercepted as the A-scan signal samples for training the non-image domain model. Among them, the one-dimensional A-scan data of the n-th trace is expressed as A n = [a n1 , a n2 , …, a nm T . Since the three-dimensional ground-penetrating radar has 16 channels, multiple A-scan training samples are obtained for the echo data in each channel in the above manner.

[0079] Step S2 in this embodiment includes the following sub-steps:

[0080] Step S201, label the single-channel A-scan signal, where the label for intact is 0, the label for underground hidden dangers is 1, and the label for manholes is 2;

[0081] Step S202, label the C-scan horizontal slice composite map, and the labeling categories include pipeline (pipe) and cavity (hole).

[0082] ​Step S3 in this embodiment includes the following sub-steps:

[0083] Step S301: Preferably, divide the A-scan signal set into a training set, a validation set, and a test set according to the ratio of 6:2:2;

[0084] Step S302: Preferably, divide the C-scan image set into a training set, a validation set, and a test set according to the ratio of 6:2:2.

[0085] Step S4 in this embodiment includes the following sub-steps:

[0086] Step S401: Use the A-scan signal set to train a non-image domain model, where the non-image domain model is a deep learning model composed of two layers of one-dimensional convolutional neural networks and two layers of GRU neural networks;

[0087] Step S402: Use the C-scan image set to train an image domain model. The image domain model uses the Yolov5 model, and during the training process, use the test set to further optimize the image domain model, and save the optimal image domain model in real-time after updating.

[0088] As Figure 7 shown, the Yolov5 model consists of three parts: Backbone, Neck, and Head. First, the model uses a Cross Stage Partial Network (CSPNet) with 53 convolutional layers, which reduces the repetition of gradient information during backpropagation, reduces the computational amount, and improves the learning ability of the network model. Second, the model also adds a Path Aggregation Network (PANet) and a Spatial Pyramid Pooling (SPP) to further improve the performance of the model and reduce the risk of overfitting in the network. As Figure 8 shown, the object detection result obtained by training the image domain model based on Yolov5 Figure 8 In the figure, the square marking box indicates that the output label is "pipe", and the inference result is the prediction probability in the range [0, 1], where the confidence threshold is set to 0.7.

[0089] In step S401 of this embodiment, the two layers of one-dimensional convolutional neural networks include the first convolutional layer and the second convolutional layer with the same settings. As Figure 5 shown, the network input size of the first convolutional layer is 256×1, corresponding to 256 sampling points in the one-dimensional A-scan signal; the number of convolutional filters is set to 128, the activation function uses the ReLU function, and then it is connected to a max pooling layer with a pool size of 2, and a dropout layer is set to 0.2 to avoid the phenomenon of overfitting in the network.

[0090] After the first convolutional layer and the second convolutional layer, connect the first GRU layer with 256 units and the second GRU layer with 32 units. After each GRU layer, set the dropout layer to 0.2. Subsequently, convert the features into a one-dimensional feature vector through the Flatten layer. Then, connect to a Dense layer with 128 units and use the ReLU activation function; the last layer is a Dense layer with 3 units, which are used to represent three classification labels respectively, namely the label of intact, the label of underground hidden danger, and the label of manhole, and use the Softmax activation function.

[0091] Furthermore, in step S401 of this embodiment, the parameter settings in the process of training the non-image domain model are as follows: the sample number batch_size of model training is set to 50, which represents the number of one-dimensional A-scan signal samples input each time during training; the number of iterations epoch of model training is set to 70; during the training process, use the categorical cross-entropy loss function and the Adam optimizer. After each training for one iteration epoch, use the validation set for validation, and after the validation, use the test set to test the performance of the non-image domain model, and update and save the optimal non-image domain model in real time. As Figure 6 shown, it is the accuracy / loss curve diagram obtained by the non-image domain model.

[0092] Step S5 of this embodiment includes the following sub-steps:

[0093] Step S501, use a three-dimensional ground penetrating radar device to conduct full-coverage detection of the area and obtain real radar data;

[0094] Step S502, preprocess the collected echo information. The preprocessing includes direct wave removal processing, global background elimination processing, reverse energy attenuation gain processing, and normalization processing, and obtain all A-scan signals within 16 channels;

[0095] Step S503, input the processed one-dimensional A-scan signals into the deep learning network based on signals for classification of the non-image domain model; among them, the output of intact is 0, the output of underground hidden danger is 1, and the output of manhole is 2;

[0096] Step S504, for the A-scan signals with consecutive classification results of 1 in 16 channels, determine them as potential areas of underground hidden dangers. However, since the one-dimensional A-scan signals of cavities and pipelines are relatively similar, the non-image domain model based on one-dimensional A-scan signals cannot distinguish these two types well. Therefore, on this basis, this embodiment further uses C-scan horizontal slices for further confirmation, that is, through the combination of step S6 to improve the accuracy of recognition.

[0097] Step S6 in this embodiment includes the following sub-steps:

[0098] Step S601: For the potential areas determined to be underground hidden dangers, extract horizontal slice images at preset intervals in the depth direction to form a C-scan horizontal slice composite image. As Figure 9 shown, for the one-dimensional A-scan signals determined to be potential hidden dangers, generate corresponding C-scan horizontal slice images;

[0099] Step S602: Input the C-scan horizontal slice composite image into the trained image domain model for object detection in the image domain model, and exclude the misrecognition results in the non-image domain model;

[0100] Step S603: Combine the results of the image domain model and the non-image domain model, and output the final object recognition result. The object types included in the output categories in the object recognition result are intact, pipeline, manhole, and hidden danger, a total of four categories.

[0101] This embodiment also provides an end-to-end 3D ground penetrating radar object recognition system based on deep learning, which adopts the above-mentioned end-to-end 3D ground penetrating radar object recognition method based on deep learning, and includes:

[0102] A sample set formation module, which is used to obtain the original echo information containing underground hidden danger objects on the survey line, preprocess the echo information, and obtain multiple one-dimensional A-scan signals and multiple groups of C-scan horizontal slice images in the channel to form a corresponding sample set;

[0103] A classification and annotation module, which is used to classify and annotate the one-dimensional A-scan signal set and the C-scan image set respectively;

[0104] A sample set division module, which is used to divide each sample set into a training set, a validation set, and a test set according to a preset ratio;

[0105] A model training module, which is used to train a non-image domain model based on A-scan signals and an image domain model based on C-scan horizontal slice images respectively;

[0106] A preprocessing and training module, which is used to read the echo information collected by the 3D ground penetrating radar, preprocess it, input it into the trained deep learning network model based on A-scan signals, and perform classification of the non-image domain model;

[0107] The misrecognition exclusion module generates a corresponding cluster of C-scan horizontal slice images according to the detected target range, inputs them into a trained deep learning network model based on C-scan horizontal slice images for target detection in the image domain model, excludes the misrecognition results in the non-image domain model of the preprocessing and training module, and outputs the final classification result.

[0108] In summary, this embodiment can omit the step of converting the original signal into a gray-scale stacking image, directly classify the preprocessed signal, improve the speed of data processing and interpretation, and has a small calculation cost and simple operation. On this basis, it also combines A-scan signal data and C-scan image data, which can better distinguish the difference between pipelines and cavities, improve the accuracy of model classification, and give full play to the advantages of 3D ground penetrating radar. The present invention can achieve the target recognition of ground penetrating radar with complex and rich information through a small amount of calculation, and can also well overcome the problem that the actual engineering project requirements cannot be met due to the low classification accuracy of the model.

[0109] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should all be regarded as belonging to the protection scope of the present invention.

Claims

1. An end-to-end three-dimensional ground penetrating radar target recognition method based on deep learning, characterized in that, It includes the following steps: Step S1: Obtain the original echo information containing underground hidden danger targets on the survey line, preprocess the original echo information, and obtain multiple one-dimensional A-scan signals and multiple groups of C-scan horizontal slice maps in the channel to form a corresponding sample set; Step S2: Classify and label the one-dimensional A-scan signal set and the C-scan image set respectively; Step S3: Divide each sample set into a training set, a validation set, and a test set according to a preset ratio; Step S4: Train a non-image domain model based on A-scan signals and an image domain model based on C-scan horizontal slice maps respectively; Step S5: Read the echo information collected by the three-dimensional ground penetrating radar, preprocess it, and input it into the trained deep learning network model based on A-scan signals for classification of the non-image domain model; Step S6: Generate a corresponding cluster of C-scan horizontal slice maps according to the detected target range, and input it into the trained deep learning network model based on C-scan horizontal slice maps for target detection in the image domain, exclude the misidentified results in the non-image domain model in Step S5, and output the final classification result; The Step S4 includes the following sub-steps: Step S401: Train a non-image domain model using the A-scan signal set. The non-image domain model is a deep learning model composed of two layers of one-dimensional convolutional neural networks and two layers of GRU neural networks; Step S402: Train an image domain model using the C-scan image set. The image domain model adopts the Yolov5 model, and further optimizes the image domain model using the test set during the training process, and saves the optimal image domain model in real-time with updates; In the Step S401, the two layers of one-dimensional convolutional neural networks include the first convolutional layer and the second convolutional layer with the same settings. The network input size of the first convolutional layer is 256×1, the number of convolutional filters is set to 128, the activation function uses the ReLU function, and then it is connected to a max pooling layer with a pool size of 2, and the dropout layer is set to 0.2; after the first convolutional layer and the second convolutional layer, a first GRU layer with 256 units and a second GRU layer with 32 units are connected. After each GRU layer, the dropout layer is set to 0.

2. After converting the features into one-dimensional feature vectors through the Flatten layer, it is connected to a Dense layer with 128 units and uses the ReLU activation function; the last layer is a Dense layer with 3 units, which is used to represent three classification labels respectively and uses the Softmax activation function; The Step S5 includes the following sub-steps: Step S501: Use the three-dimensional ground penetrating radar equipment to conduct full-coverage detection of the area to obtain real radar data; Step S502: Preprocess the collected echo information. The preprocessing includes direct wave excision processing, global background elimination processing, reverse energy attenuation gain processing, and normalization processing, and obtain all A-scan signals in 16 channels; Step S503: Input the processed one-dimensional A-scan signal into a deep learning network based on signals for classification of a non-image domain model; Step S504: For the continuous A-scan signals with a classification result of 1 in 16 channels, determine the potential areas of underground hazards; The said step S6 includes the following sub-steps: Step S601: For the potential areas determined as underground hazards, extract horizontal slice maps at preset intervals in the depth direction to form a C-scan horizontal slice combined map; Step S602: Input the C-scan horizontal slice combined map into the trained image domain model for object detection in the image domain model to exclude mis-recognition results in the non-image domain model; Step S603: Combine the results of the image domain model and the non-image domain model to output the final object recognition result. The target objects included in the output category in the target recognition result are intact, pipelines, manholes, and hazards.

2. The end-to-end 3D ground penetrating radar target recognition method based on deep learning according to claim 1, wherein, The said step S1 includes the following sub-steps: Step S101, obtain the original echo information formed by the A-scan signals collected at each scanning position on the survey line , , where ; ; n represents the number of scanning positions, m represents the number of sampling points; i represents the sampling points of each A-scan data in the depth direction, j represents the number of A-scan signals collected in the survey line direction, k represents the channel for the current data acquisition; Step S102, perform preprocessing on the original echo matrix containing underground targets to form a preprocessed echo signal matrix, where the preprocessing includes direct wave excision processing, global background elimination processing, reverse energy attenuation gain processing, and normalization processing; Step S103: Extract the horizontal profile scan map containing underground targets on the survey line, intercept a preset number of horizontal slice maps at a preset interval in the depth direction to form a C-scan horizontal slice combined map, and use this as a sample for training the image domain model.

3. The end-to-end three-dimensional ground penetrating radar target recognition method based on deep learning according to claim 2, wherein The said step S102 includes the following sub-steps: Step S1021: When the first amplitude value in the single-channel data is lower than the preset threshold, record this position as the termination point of the direct wave, and sequentially intercept 256 points after this termination point position as the data after removing the direct wave; Step S1022: Subtract the global average value from each data to achieve global background elimination for the new matrix; Step S1023, perform reverse energy attenuation gain on the data in the matrix, and multiply the new matrix by the normalized gain factor , to achieve reverse energy attenuation gain of the data, where , represents the fitting attenuation model, represents the maximum return value of the fitting attenuation curve; Step S1024, by formula The data in the matrix is ​​normalized, wherein: , Indicates j A-scan i Signal data of radar sampling points, represents the normalized sampling point value, Represents the radar signal data with the largest amplitude, Indicates radar signal data with the smallest amplitude.

4. The end-to-end 3D ground penetrating radar target recognition method based on deep learning according to any one of claims 1 to 3, characterized in that, The said step S2 includes the following sub-steps: Step S201: Label the single-channel A-scan signal. Among them, the label for intact is 0, the label for underground hazard is 1, and the label for manhole is 2; Step S202: Label the C-scan horizontal slice combined map, and the labeling categories include pipelines and cavities.

5. The end-to-end 3D ground penetrating radar target recognition method based on deep learning according to any one of claims 1 to 3, characterized in that, In the said step S401, the parameter settings in the process of training the non-image domain model are as follows: the sample quantity batch_size for model training is set to 50, and the number of iterations epoch for model training is set to 70; during the training process, a categorical cross-entropy loss function and an Adam optimizer are adopted. After each training of an iteration epoch, the validation set is used for validation, and after validation, the test set is used to test the performance of the non-image domain model, and the optimal non-image domain model is updated and saved in real time.

6. An end-to-end 3D ground penetrating radar target recognition system based on deep learning, characterized in that, The end-to-end 3D ground penetrating radar object recognition method based on deep learning as described in any one of claims 1 to 5 is adopted, and it includes: A sample set forming module, used to obtain the original echo information containing underground hazard targets on the survey line, preprocess the echo information, and obtain multiple one-dimensional A-scan signals and multiple groups of C-scan horizontal slice maps in the channel to form a corresponding sample set; A classification and labeling module, used to classify and label the one-dimensional A-scan signal set and the C-scan image set respectively; A sample set division module, configured to divide each sample set into a training set, a validation set, and a test set according to a preset ratio; A model training module, configured to train a non-image domain model based on A-scan signals and an image domain model based on C-scan horizontal slice images respectively; A preprocessing and training module, configured to read echo information collected by a three-dimensional ground penetrating radar, preprocess the echo information, input the preprocessed echo information into a trained deep learning network model based on A-scan signals, and perform classification of the non-image domain model; A misrecognition elimination module, configured to generate a corresponding cluster of C-scan horizontal slice images according to a detected target range, input the cluster of C-scan horizontal slice images into a trained deep learning network model based on C-scan horizontal slice images, perform target detection of the image domain model, eliminate misrecognition results in the non-image domain model of the preprocessing and training module, and output a final classification result.

Images