A three-dimensional ground penetrating radar underground cavity target automatic identification method based on a joint CNN
By using a joint CNN approach to preprocess and construct datasets from ground-penetrating radar echo images, and combining this with multi-layer network training, we achieved efficient and accurate automatic identification of 3D hollow targets, solving the problem of low efficiency in traditional methods.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2023-08-03
- Publication Date
- 2026-05-19
AI Technical Summary
Traditional methods and two-dimensional neural networks are inefficient, computationally intensive, and have poor real-time performance when identifying underground cavities. Furthermore, the detection process for three-dimensional echo data is cumbersome and has low accuracy.
A joint CNN-based approach is adopted. By preprocessing and annotating ground-penetrating radar echo images, two-dimensional and three-dimensional detection datasets are constructed. RetinaNet, Faster R-CNN, ResNet and Res2Net networks are used for training. By combining two-dimensional and three-dimensional convolutions, the automatic identification of hollow targets is achieved.
It improves the probability and speed of identifying underground cavity targets, achieving a recognition rate of 95%, and overcomes the shortcomings of two-dimensional convolutional networks in being insensitive to spatial information.
Smart Images
Figure CN117173617B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of target detection in ground-penetrating radar echo image post-processing; specifically, it relates to an automatic identification method for underground cavity targets in three-dimensional ground-penetrating radar based on joint CNN. Background Technology
[0002] Ground-penetrating radar (GPR) is a radio frequency electromagnetic technology that uses the physical premise of detecting the differences in the electromagnetic properties of different underground media. The equipment moves along the survey line, while the transmitting antenna emits electromagnetic waves that propagate underground. During propagation, these waves encounter interfaces with different electromagnetic properties, resulting in forward and backscattering. The scattered waves then produce further scattering when they encounter different target interfaces. Forward-scattered waves continue to propagate deeper, while backscattered waves propagate towards the surface and are received by the receiving antenna. By analyzing the echo information, the distribution of the underground media can be determined. GPR offers advantages such as speed, high resolution, non-destructive testing, and real-time performance, making it suitable for identifying underground cavities.
[0003] Traditional identification methods rely on professionals to distinguish underground cavities based on echo energy, waveform, and phase characteristics. The biggest problem with manual identification is its low efficiency, making it unsuitable for today's vast amounts of data. Therefore, there is an urgent need to find suitable automated methods for interpreting echo data. Ground-penetrating radar (GPR) echo data has its own characteristics. On the B-scan, it resembles traditional image data, with multiple parallel B-scans forming three-dimensional GPR echo data. For cavities of interest, they all appear as hyperbolic curves on the B-scan, as do some other interfering targets. On the C-scan, cavities exhibit unique characteristics, but directly detecting them on the C-scan leads to high computational costs and poor real-time performance. Furthermore, performing two-dimensional detection channel by channel on three-dimensional echo data inevitably results in a cumbersome detection process, neglecting information between target channels and leading to low detection accuracy. Summary of the Invention
[0004] This invention provides an automatic identification method for underground cavity targets using three-dimensional ground-penetrating radar based on a combined CNN, in order to solve the problem of low efficiency in identifying underground cavity targets using traditional methods and two-dimensional neural networks.
[0005] This invention is achieved through the following technical solution:
[0006] An automatic identification method for underground cavities using 3D ground-penetrating radar based on joint CNN, the automatic identification method comprising the following steps:
[0007] Step 1: Perform direct wave removal and gain control on the acquired ground-penetrating radar echo image of the underground cavity target to obtain the processed ground-penetrating radar echo image;
[0008] Step 2: Pre-screen the ground-penetrating radar echo images generated in Step 1, mark the pixel positions of the target holes, and create a two-dimensional detection dataset. Based on this, repeatedly crop fixed positions from the two-dimensional images of different channels to obtain a three-dimensional detection dataset.
[0009] Step 3: Randomly divide the two-dimensional detection dataset from Step 2 into a training set and a validation set. Train the training set of the two-dimensional dataset using RetinaNet and Faster R-CNN networks with different backbones to obtain a two-dimensional weight model.
[0010] Step 4: Randomly divide the 3D detection dataset from Step 2 into a training set and a validation set. Use ResNet and Res2Net classification networks to train the training set of the 3D dataset to obtain the 3D weight model.
[0011] Step 5: Use the validation set of the two-dimensional detection dataset to detect the two-dimensional weight model obtained in Step 3 to narrow down the target range. Then, based on the two-dimensional detection results, extract the three-dimensional data of the detection location from the three-dimensional dataset obtained in Step 2 and use the three-dimensional weight model obtained in Step 4 for recognition.
[0012] Furthermore, step 1 specifically involves filtering out the direct wave, one of the main sources of interference, using mean filtering, and then using envelope extrema to generate a gain compensation function to amplify the characteristic information of the underground target, thereby obtaining the processed ground-penetrating radar echo image.
[0013] Furthermore, step 2, which involves creating a two-dimensional detection dataset, specifically involves using bounding boxes to label the targets in the two-dimensional detection dataset, obtaining all the labeling information, and then creating the dataset according to the COCO dataset format.
[0014] Furthermore, the three-dimensional detection dataset obtained in step 2 is processed into images channel by channel, and then the target in the image is cropped to a fixed size. The fixed position is repeatedly cropped for each channel to obtain the complete three-dimensional data of the target.
[0015] Furthermore, step 3 specifically involves training the RetinaNet network and Faster R-CNN network with 50 and 101 layers respectively, using ResNet and Res2Net as backbone networks for the two-dimensional dataset.
[0016] Furthermore, step 4 specifically involves, for the 3D dataset, expanding the 2D convolutions in the ResNet and Res2Net networks into 3D convolutions to reduce the number of feature channels, and performing training with 10, 18, and 34 layers respectively.
[0017] Furthermore, step 5 specifically involves taking one B-Scan image from the two-dimensional detection dataset for target detection, obtaining the two-dimensional detection result, which includes the target category, position, and size in the image. Then, based on the two-dimensional detection result, three-dimensional data of the detection position is extracted from multiple B-Scan images in the three-dimensional dataset obtained in step 2. The obtained three-dimensional data is then fed into the three-dimensional classification network in separate channels. Finally, by combining the information obtained from the detection network and the classification network, the hollow targets contained in the input three-dimensional data and their positions in the three-dimensional data are identified.
[0018] The beneficial effects of this invention are:
[0019] This invention uses images labeled with manually categorized underground cavities as a training set to train a convolutional neural network model capable of automatically identifying underground cavity targets in subsequent ground-penetrating radar 3D echo images. Employing this method to detect underground cavity targets in ground-penetrating radar 3D echo images can effectively improve both the recognition probability and speed.
[0020] This invention can increase the probability of identifying underground cavity targets to 95% on a three-dimensional dataset.
[0021] In practice, ground-penetrating radar (GPR) data, due to the complexity of the underground environment, cannot fully represent the underground target using only two-dimensional B-Scan images. This makes subsequent target identification prone to misjudgment and missed detection. The purpose of this invention is to input the three-dimensional echo image of GPR into a 3D neural network. Through 3D convolutional kernels, features of the three-dimensional echo image are extracted and jointly judged with the two-dimensional echo image, overcoming the limitation of two-dimensional convolutional networks being insufficiently sensitive to spatial information. Finally, the trained model is used to identify underground cavity targets in GPR three-dimensional echo images. Attached Figure Description
[0022] Figure 1 This is a flowchart of the automatic identification method for underground cavity targets based on three-dimensional ground-penetrating radar using a combined CNN, according to the present invention.
[0023] Figure 2 This is a diagram of the two-dimensional training model structure.
[0024] Figure 3 This is a diagram of the structure of the three-dimensional training model used.
[0025] Figure 4 It is a pre-screened image of a single underground cavity target's ground-penetrating radar echo.
[0026] Figure 5 These are the loss function and accuracy curve during the training process.
[0027] Figure 6 This is a flowchart of the joint testing process.
[0028] Figure 7 This is a schematic diagram of the combined test results. Detailed Implementation
[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] An automatic identification method for underground cavities using 3D ground-penetrating radar based on joint CNN, the automatic identification method comprising the following steps:
[0031] Step 1: Perform direct wave removal and gain control on the acquired ground-penetrating radar echo image of the underground cavity target to obtain the processed ground-penetrating radar echo image;
[0032] Step 2: Pre-screen the ground-penetrating radar echo images generated in Step 1, mark the pixel positions of the target holes, and create a two-dimensional detection dataset. Based on this, repeatedly crop fixed positions from the two-dimensional images of different channels to obtain a three-dimensional detection dataset.
[0033] Step 3: Randomly divide the two-dimensional detection dataset from Step 2 into a training set and a validation set. Train the training set of the two-dimensional dataset using RetinaNet and Faster R-CNN networks with different backbones to obtain a two-dimensional weight model.
[0034] Step 4: Randomly divide the 3D detection dataset from Step 2 into a training set and a validation set. Use ResNet and Res2Net classification networks to train the training set of the 3D dataset to obtain the 3D weight model.
[0035] Step 5: Use the validation set of the two-dimensional detection dataset to detect the two-dimensional weight model obtained in Step 3 to narrow down the target range. Then, based on the two-dimensional detection results, extract the three-dimensional data of the detection location from the three-dimensional dataset obtained in Step 2 and use the three-dimensional weight model obtained in Step 4 for recognition.
[0036] Furthermore, step 1, which preprocesses the acquired ground-penetrating radar echo image of the underground cavity target, specifically involves removing surface reflection waves and filtering the acquired ground-penetrating radar echo image of the underground cavity target. Surface reflection waves are removed by direct cropping, and the filtering adopts the transverse mean filtering method to obtain a ground-penetrating radar echo image with suppressed transverse ripples.
[0037] Furthermore, in step 1, the obtained ground-penetrating radar echo image of the underground cavity target is subjected to direct wave removal and gain control to obtain the processed ground-penetrating radar echo image. Specifically, the direct wave, one of the main sources of interference, is filtered out by mean filtering, and then the envelope extremum is used to generate a gain compensation function to amplify the feature information of the underground target, thus obtaining the processed ground-penetrating radar echo image.
[0038] Furthermore, step 2 pre-screens the ground-penetrating radar echo image generated in step 1, clearly marking the pixel positions of the target holes to create a two-dimensional detection dataset. Based on this, a three-dimensional detection dataset is obtained by repeatedly cropping fixed positions from the two-dimensional images of different channels. Specifically, for the two-dimensional detection dataset, the target is marked with a rectangle, and after obtaining all the marking information, the dataset is created according to the COCO dataset format. For the three-dimensional detection dataset, the original data is processed into images channel by channel, and then the target in the image is cropped to a fixed size. The fixed positions are repeatedly cropped for each channel to obtain the complete three-dimensional data of the target.
[0039] Furthermore, in step 3, the two-dimensional detection dataset in step 2 is randomly divided into a training set and a validation set. The training set of the two-dimensional dataset is trained using RetinaNet and Faster R-CNN networks with different backbone networks to obtain the two-dimensional weight model. Specifically, for the two-dimensional dataset, the RetinaNet network and Faster R-CNN network are trained with ResNet and Res2Net as backbone networks with 50 and 101 layers respectively.
[0040] In step 4, the 3D detection dataset from step 2 is randomly divided into a training set and a validation set. The training set of the 3D dataset is trained using ResNet and Res2Net classification networks to obtain the 3D weight model. Specifically, for the 3D dataset, the 2D convolutions in the ResNet and Res2Net networks are expanded into 3D convolutions to reduce the number of feature channels. Training is performed with 10, 18, and 34 layers respectively.
[0041] As shown in steps 3 and 4, the ResNet network addresses network degradation by modifying its model structure. The basic convolutional module of the ResNet network, the residual block, superimposes the original input onto the output after convolution. During network training, when the processing of a certain convolutional layer negatively impacts model optimization, the model can use an identity mapping to preserve the original result, thus preventing network performance degradation. The Res2Net network aims to enhance the network's multi-scale expressive capabilities by constructing a residual-layer-like connection within a residual block. This not only increases the receptive field size but also diversifies the feature scale. After a 1×1 convolutional layer, the feature map is divided into s subsets, using x... i Let be the subset of features, where i ∈ {1, 2, ..., s}. Except that the number of feature map channels in each subset is 1 / s of the original number, each feature map subset x... i They all have the same spatial size as the original feature map set before partitioning. Except for x i Each feature map subset has its corresponding 3×3 convolutional layer, using K... i () represents. Define K. i The output of () is y i A subset x of the feature map i and K i-1 (x i-1 After adding them together, send them into K. i () performs convolution processing. To reduce the number of parameters and increase the number of s, the 3×3 convolution to be performed on x1 is ignored, therefore y i It can be represented as: Each 3×3 convolution kernel K i () can subtly receive all the feature map subsets that precede it, i.e., {x j Feature information for the subset x of the feature map, where j ≤ i. j After passing through a 3×3 convolution kernel, the output feature map can have a ratio of x. j A larger sensory field.
[0042] Furthermore, in step 4, the two-dimensional validation set is first input into the two-dimensional weight model obtained in step 3 for detection to narrow down the target range. Then, based on the two-dimensional detection results, the three-dimensional data of the detection location is extracted from the original three-dimensional dataset and input into the three-dimensional weight model obtained in step 3 for recognition. Specifically, one B-Scan image from the two-dimensional detection dataset is taken for target detection to obtain the two-dimensional detection results, which include the target category, position, and size in the image. Then, based on the two-dimensional detection results, the three-dimensional data of the detection location is extracted from multiple B-Scan images in the three-dimensional dataset obtained in step 2. The obtained three-dimensional data is sent to the three-dimensional classification network by channel. Finally, the information obtained by the detection network and the classification network is combined to find the hollow targets contained in the input three-dimensional data and their positions in the three-dimensional data.
Claims
1. An automatic identification method for underground cavities using three-dimensional ground-penetrating radar based on joint CNN, characterized in that, The automatic identification method includes the following steps: Step 1: Perform direct wave removal and gain control on the acquired ground-penetrating radar echo image of the underground cavity target to obtain the processed ground-penetrating radar echo image; Step 2: Pre-screen the ground-penetrating radar echo images generated in Step 1, mark the pixel positions of the target holes, and create a two-dimensional detection dataset. Based on this, repeatedly crop fixed positions from the two-dimensional images of different channels to obtain a three-dimensional detection dataset. Step 3: Randomly divide the two-dimensional detection dataset from Step 2 into a training set and a validation set. Train the training set of the two-dimensional dataset using RetinaNet and Faster R-CNN networks with different backbones to obtain a two-dimensional weight model. Step 4: Randomly divide the 3D detection dataset from Step 2 into a training set and a validation set. Use ResNet and Res2Net classification networks to train the training set of the 3D dataset to obtain the 3D weight model. Step 5: Use the validation set of the two-dimensional detection dataset to detect the two-dimensional weight model obtained in Step 3 to narrow down the target range. Then, based on the two-dimensional detection results, extract the three-dimensional data of the detection location from the three-dimensional dataset obtained in Step 2 and use the three-dimensional weight model obtained in Step 4 for recognition. Step 1 specifically involves filtering out the direct wave, one of the main sources of interference, using mean filtering, and then using envelope extrema to generate a gain compensation function to amplify the feature information of the underground target, thereby obtaining the processed ground-penetrating radar echo image. The specific steps of step 2, which involves creating a two-dimensional detection dataset, are as follows: for a two-dimensional detection dataset, the target is labeled using rectangular boxes, and after obtaining all the labeling information, the dataset is created according to the COCO dataset format. Specifically, step 3 involves training the RetinaNet network and Faster R-CNN network with 50 and 101 layers respectively, using ResNet and Res2Net as backbone networks for a two-dimensional dataset. Step 5 specifically involves taking one B-Scan image from the two-dimensional detection dataset for target detection, obtaining the two-dimensional detection result, which includes the target's category, position, and size in the image. Then, based on the two-dimensional detection result, three-dimensional data of the detection location is extracted from multiple B-Scan images in the three-dimensional dataset obtained in step 2. The obtained three-dimensional data is then fed into the three-dimensional classification network through different channels. Finally, by combining the information obtained from the detection network and the classification network, the hollow targets contained in the input three-dimensional data and their positions in the three-dimensional data are identified.
2. The automatic identification method for underground cavity targets based on joint CNN using three-dimensional ground-penetrating radar according to claim 1, characterized in that, The three-dimensional detection dataset obtained in step 2 is processed into images channel by channel. Then, the target in the image is cropped to a fixed size, and the fixed position is repeatedly cropped for each channel to obtain the complete three-dimensional data of the target.
3. The automatic identification method for underground cavity targets based on joint CNN using three-dimensional ground-penetrating radar according to claim 1, characterized in that, Specifically, step 4 involves expanding the two-dimensional convolutions in the ResNet and Res2Net networks into three-dimensional convolutions for the three-dimensional dataset, reducing the number of feature channels, and training with 10, 18, and 34 layers respectively.