An Automatic Detection Method for Subsurface Cavities Using Ground Penetrating Radar Based on a RefineDet Network Model

Abstract

This invention proposes an automatic detection method for underground cavities using a RefineDet network model based on ground-penetrating radar (GPR). The method includes: background removal of the acquired GPR echo image of an underground cavity target to obtain a GPR echo image with suppressed lateral ripples; gaining the generated GPR echo image to highlight the pixel location features of the cavity target; denoising the amplified image to suppress clutter; pre-screening the processed GPR echo image, labeling cavities in the image using labelimg based on results from manual identification and field confirmation; creating an image detection dataset from the obtained data; training a RefineDet network model using the obtained detection dataset to obtain network weight parameters; and using the trained network model to detect underground cavity targets in GPR echo images. This invention solves the problem of existing methods' difficulty in detecting underground cavity targets.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology in ground-penetrating radar echo image post-processing, and in particular relates to an automatic detection method for underground cavity targets using a RefineDet network model. Background Technology

[0002] Ground-penetrating radar (GPR) is a detection instrument used for non-destructive exploration of shallow underground environments. GPR utilizes the differences in the electromagnetic permittivity of the underground medium, reflecting these differences as parameters in radar echo data. By processing this echo data, efficient detection and a clear understanding of the distribution of the underground environment can be achieved. To visually present the echo data for manual analysis, a common method is to horizontally display multi-channel echo data; the B-Scan image, a commonly used data format in GPR analysis, is based on this approach.

[0003] Ground-penetrating radar (GPR), as a fast, high-resolution, and non-destructive geophysical method, plays a significant role in the research and engineering practice of detecting underground cavities. GPR technology is applicable to various road conditions and does not cause irreversible damage to the target. Its detection results are characterized by good real-time performance and high accuracy, meeting the requirements of wide application, non-destructive testing, and high efficiency in highway defect detection, thus perfectly suited for the task of detecting underground cavities. A GPR system can consist of single or multiple pairs of transmitting and receiving antennas. Each pair of transmitters and receivers can acquire a single B-Scan image by scanning candidate areas. Analysis and verification of the B-Scan images can reveal the distribution of the underground environment. Currently, in practical engineering, B-Scan images acquired require manual interpretation, a method that is inefficient and prone to missed or false detections. Image detection is an advanced task in image recognition, requiring not only the identification of the presence of cavities in the GPR echo image but also the identification of the target's approximate location and size. Using some of the current mainstream deep learning methods to detect and identify underground cavities also presents challenges. Obtaining relevant pattern information about underground cavities through confirmation, verification, and localization is difficult in engineering practice, and the patterns and shapes of underground cavities in B-Scan images are not fixed. Furthermore, acquiring a large number of ground-penetrating radar echo image samples of underground cavities is also a demanding engineering task. Summary of the Invention

[0004] The purpose of this invention is to solve the problem that existing methods are difficult to detect and identify underground cavity targets, and to propose an automatic detection method for underground cavity targets using ground penetrating radar based on the RefineDet network model.

[0005] This invention is achieved through the following technical solution: This invention proposes an automatic detection method for underground cavity targets using a RefineDet network model for ground-penetrating radar. The method specifically includes:

[0006] Step 1: Perform background removal on the acquired ground-penetrating radar echo image of the underground cavity target to obtain a ground-penetrating radar echo image with suppressed transverse ripples.

[0007] Step 2: Gain the ground-penetrating radar echo image generated in Step 1, suppress the background, and extract the cavity features submerged in the image;

[0008] Step 3: Denoise the image data after gain processing in Step 2 to suppress the influence of clutter;

[0009] Step 4: Pre-screen the ground-penetrating radar echo images processed in Step 3. Based on the results of manual identification and on-site confirmation, use labelimg to mark the holes in the images.

[0010] Step 5: Convert the image data obtained in Step 4 into a detection network dataset in PASCAL VOC dataset format;

[0011] Step 6: Input the training set from the dataset obtained in Step 5 into the RefineDet network and train it to obtain the weight model;

[0012] Step 7: Input the test set from the dataset obtained in Step 5 into the obtained weight model to perform target detection on the ground-penetrating radar echo image of the underground cavity target.

[0013] Furthermore, in step 1, the image background is eliminated by the transverse ripple suppression filtering method to obtain a ground-penetrating radar echo image with suppressed transverse ripples.

[0014] Furthermore, the nodal mean linear gain method is used to extract the features of the hollow target from the redundant background information. The nodal mean linear gain method can highlight the curve features of the hollow in the background and obtain the position and shape features of the hollow target more clearly.

[0015] Furthermore, the nodalized mean linear gain method is specifically as follows:

[0016] First, the image is divided into 7 parts vertically on an even basis. Then, the starting row of each part and the last row of the image are each assigned a node, resulting in a total of 8 nodes.

[0017] Then, the average value of the maximum pixel value of each row in each part is taken as the gain value of the corresponding node. The pre-gain curve is obtained by linear interpolation, and the maximum value is used to calibrate the pre-gain curve to obtain the gain curve. Each row of the image corresponds to a point on the gain curve, and the size of the point is the gain value of the row. The image is then gained according to the gain curve.

[0018] Furthermore, the noise reduction employs fast nonlocal means denoising.

[0019] Further, in step 5, the image data is made into a detection network dataset in PASCAL VOC dataset format, wherein a portion of the targetless images and holes are allocated to the network for training, and the remaining portion of the targetless images and holes are allocated to the network for testing.

[0020] Furthermore, step 6 specifically includes:

[0021] The RefineDet network is divided into 4, 5 or 6 layers; the backbone network is divided into ResNet50, Ghost_ResNet50 and their respective hybrid pooling models with different proportions, which include max pooling weights and average pooling weights increasing from 0 to 1 with a step size of 0.1; the connection blocks are divided into original connection blocks and connection blocks with added attention mechanisms.

[0022] The Ghost module works as follows: In the original process, an input image of size w×h is processed by n sets of k×k convolution kernels to generate an output with n channels and size w'×h'. In the Ghost module, a simple operation Φ is used to replace the traditional convolution operation to obtain the Ghost feature map: First, the input image is processed by m sets of k×k convolution kernels to generate an intrinsic feature map with m channels and size w'×h'. Then, the operation Φ is applied to the intrinsic feature map to generate the Ghost feature map. The combined results are the output of the module.

[0023] The operation Φ uses depthwise convolution (DWC). In DWC, one convolution kernel corresponds to one channel, and a single convolution kernel only processes the corresponding channel. Operation Φ is equivalent to a d×d convolution kernel operation.

[0024] The attention mechanism is specifically as follows:

[0025] The design is carried out from two dimensions: channel and space. The channel attention mechanism is specifically implemented as follows: the input feature image is pooled in two modes to obtain two descriptions of the background information from two perspectives. Then, the results of the two pooling processes are processed by a multilayer perceptron with shared weights. The two results are then summed and input into the channel weight judgment function, and the sigmoid function is selected.

[0026] Output Mc Represented as M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))), where F represents the input feature map and σ represents the sigmoid function;

[0027] The spatial attention mechanism works as follows: For the input feature image, two abstraction methods are applied to the channel of each feature point to obtain a multidimensional description of the feature image. After concatenating the abstract results, a convolutional kernel is used to further abstract and extract features from the concatenated result. Finally, the result is fed into the spatial weight judgment function using the sigmoid function. At this point, the module obtains the weight of each feature point in the input feature layer. Based on this spatial feature weight, multiplying this weight by the original input feature layer yields the result after spatial attention processing; the output result M... s Represented as:

[0028] M s (F)=σ(f n×n ([AvgPool(F); MaxPool(F)])), where F represents the input feature map, f n×n This represents a convolution kernel operation of size n×n.

[0029] Furthermore, the hybrid pooling of the backbone network is specifically as follows: when performing feature extraction, max pooling is used in shallow layers and average pooling is used in deep layers.

[0030] This invention proposes an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the automatic detection method for underground cavity targets using a RefineDet network model of ground penetrating radar.

[0031] This invention proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the automatic detection method for underground cavity targets using a RefineDet network model of ground penetrating radar.

[0032] The beneficial effects of this invention are as follows:

[0033] This invention performs background removal, gain enhancement, and noise reduction on existing ground-penetrating radar (GPR) echo images of underground cavities. The resulting images are then flipped and augmented to create a dataset in PASCAL VOC format. A RefineDet network model is trained using a portion of this dataset, and the model is tested using the remaining data. Based on the test results, this invention can be applied to data not previously encountered by real-world networks. This invention is for target image detection tasks, involving the annotation of targets of interest. Building upon target identification and classification, it annotates the target's location, representing an advanced task in target recognition. The output results are used to detect underground cavities in GPR echo images. Using the method of this invention to detect underground cavities in GPR echo images effectively improves the detection and recognition probability. This invention achieves an accuracy of over 87% and a false negative rate of less than 8.3% for underground cavity target detection.

[0034] In practice, when ground-penetrating radar (GPR) collects data related to underground cavities, the shape of these cavities is random and difficult to predict, and their depth, size, and location are unknown. This not only poses a challenge to the detection task but also significantly hinders data acquisition and subsequent deep learning-based classification and detection. The purpose of this invention is to train a RefineDet network model using a dataset of underground cavity target echo images, and then use the resulting model to detect underground cavity target echo images using GPR. Attached Figure Description

[0035] Figure 1 This is a flowchart of the automatic detection method for underground cavity targets using ground-penetrating radar based on the RefineDet model, according to the present invention.

[0036] Figure 2 Here are the RefineDet network model diagrams, where (a) is a 4-layer model, (b) is a 5-layer model, and (c) is a 6-layer model.

[0037] Figure 3 This is a schematic diagram of the TCB structure in the RefineDet network model, where (a) is the original structure and (b) is the structure after adding the attention mechanism.

[0038] Figure 4 These are schematic diagrams of the original convolution and Ghost convolution operations, where (a) is the original convolution and (b) is the Ghost operation.

[0039] Figure 5 This is a schematic diagram of max pooling and average pooling.

[0040] Figure 6 It is a raw ground-penetrating radar echo image of a single underground cavity target.

[0041] Figure 7 It is a ground-penetrating radar echo image of a single underground cavity target after background removal.

[0042] Figure 8 It is a ground-penetrating radar echo image of a single underground cavity target, after gaining.

[0043] Figure 9 This is a noise-reduced image of a single underground cavity target's ground-penetrating radar echo image.

[0044] Figure 10 This is a schematic diagram of the RefineDet training curve.

[0045] Figure 11 These are the training curves of the RefineDet model after adding the attention mechanism, where (a) is the training curve of the 4-layer model, (b) is the training curve of the 5-layer model, and (c) is the training curve of the 6-layer model.

[0046] Figure 12 These are the results of underground cavity detection, where (a) represents a correct detection, (b) represents an incorrect detection, and (c) represents a missed detection. Detailed Implementation

[0047] 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.

[0048] Example 1:

[0049] This invention proposes an automatic detection method for underground cavities using ground-penetrating radar based on the RefineDet network model. The automatic detection method includes the following steps:

[0050] Step 1: Perform background removal on the acquired ground-penetrating radar echo image of the underground cavity target to obtain a ground-penetrating radar echo image with suppressed transverse ripples.

[0051] Step 2: Gain the ground-penetrating radar echo image generated in Step 1, suppress the background, and extract the cavity features submerged in the image;

[0052] Step 3: Denoise the image data after gain processing in Step 2 to suppress the influence of clutter;

[0053] Step 4: Pre-screen the ground-penetrating radar echo images processed in Step 3. Based on the results of manual identification and on-site confirmation, use labelimg to mark the holes in the images.

[0054] Step 5: The image data obtained in Step 4 is used to create a detection network dataset in the classic PASCAL VOC dataset format. A certain proportion of the targetless images and dilated images are allocated to the network for training, and the remaining targetless images and dilated images are allocated to the network for testing.

[0055] Step 6: Input the training set from the dataset obtained in Step 5 into the RefineDet network and train it to obtain the weight model;

[0056] Step 7: Input the test set from the dataset obtained in Step 5 into the obtained weight model to perform target detection on the ground-penetrating radar echo image of the underground cavity target.

[0057] In step 1,

[0058] The preprocessing includes: background removal of the acquired ground-penetrating radar echo images of underground cavity targets;

[0059] By using a transverse ripple suppression filtering method to eliminate the image background, a ground-penetrating radar echo image with suppressed transverse ripples is obtained.

[0060] In step 2,

[0061] The nodalized mean linear gain method is used to extract the features of cavitary targets from redundant background information. Nodalized mean linear gain can highlight the curve features of cavities in the background, providing a clearer understanding of the location and shape of cavitary targets.

[0062] In step 2, the nodal mean linear gain method specifically includes:

[0063] First, the image is divided into 7 parts vertically, according to the depth of the ground-penetrating radar echo image. Then, the starting row of each part and the last row of the image are each assigned a node, resulting in a total of 8 nodes.

[0064] Then, the average value of the maximum pixel value of each row in each part is taken as the gain value of the corresponding node. The pre-gain curve is obtained by linear interpolation, and the maximum value is used to calibrate the curve to obtain the gain curve. Each row of the image corresponds to a point on the curve, and the size of the point is the gain value of the row. The image is then gained according to the gain curve.

[0065] In step 3,

[0066] The noise reduction process described is specifically: fast nonlocal mean denoising.

[0067] The Fast Non-Local Means (NL-means) denoising algorithm is an accelerated version of Non-Local Means. NL-means filters based on the similarity between pixels.

[0068] For an image, a search bounding box of size D×D is selected, and a neighborhood bounding box of size d×d centered at x and y is selected. The similarity between two neighborhoods is measured as follows: Here, the size of the neighborhood is m×n, and x(i,j) and y(i,j) are the pixel values ​​within two neighborhoods. This similarity is used to obtain the metric for each pixel. Where h represents the smoothing factor, which affects the distortion of the filter. The final filtered result for point x is: NLmeans(x)=∑w(x,y)·y

[0069] The fast algorithm addresses the time-consuming point-by-point calculation of the original method by constructing an integral image of the pixels, thereby accelerating filtering. This invention constructs a 5×5 search window and a 3×3 neighborhood window.

[0070] In step 4,

[0071] The processed ground-penetrating radar echo images are pre-screened. Based on the results of manual identification and on-site confirmation, the holes in the images are marked using labelimg, and the hole targets are labeled. There may be more than one hole target in an image.

[0072] In step 5,

[0073] The PASCAL VOC format dataset consists of three main parts: JPEGImages, ImageSets, and Annotations. JPEGImages stores the images used for the task assignments. The ImageSets folder contains a Main folder containing related .txt files, such as train.txt for training and test.txt for testing. These .txt files list the filenames of the image files used to perform their respective tasks. The Annotations folder stores .xml files, which store relevant information for each image, such as image address, bounding box coordinates, the category of the object within the bounding box, and whether the object is a hard sample.

[0074] In step 6,

[0075] The RefineDet network, developed based on SSD, is a single-stage detection algorithm. RefineDet consists of three parts: the Anchor Refinement Module (ARM), the Object Detection Module (ODM), and the Transfer Connection Block (TCB). The ARM performs foreground and background segmentation of the image, reducing the number of samples for the ODM, providing refined anchor boxes, and optimizing the bounding box regression starting point. The ODM uses this result for classification and object localization. The TCB transforms the feature map from the ARM into the form required by the ODM, avoiding region-by-region RoI pooling.

[0076] The RefineDet network model used in this invention consists of 4, 5, or 6 layers. The backbone network of the RefineDet model includes ResNet50, Ghost_ResNet50, and hybrid pooling models with different proportions, where the max pooling weights and average pooling weights increase from 0 to 1 with a step size of 0.1. See Table 1 for details.

[0077] Table 1 ResNet50 network model structure

[0078]

[0079] The Ghost_ResNet50 is based on the fact that feature maps of the same depth have a certain degree of similarity, and obtains the required feature maps in a less computationally intensive way.

[0080] The Ghost module works as follows: Assuming the original process involves an input image of size w×h, after being processed by n sets of k×k convolutional kernels, an output with n channels and size w'×h' is generated. The Ghost module uses a simple operation Φ to replace the traditional convolution operation, obtaining the Ghost feature map: First, the input image is processed using m sets of k×k convolutional kernels to generate an intrinsic feature map with m channels and size w'×h'. Then, the operation Φ is applied to the intrinsic feature map to generate the Ghost feature map. The combined output of the module is then obtained. See Table 2 for details.

[0081] Table 2 Results of Ghost module utilization

[0082]

[0083] For the operation Φ, this invention uses depth-wise convolution (DWC). Unlike ordinary convolution, in DWC, one convolution kernel corresponds to one channel, and a single convolution kernel only processes the corresponding channel. The operation Φ can be equivalent to a d×d convolution kernel operation.

[0084] The RefineDet model's connection blocks are divided into original connection blocks and connection blocks with added attention mechanisms.

[0085] The attention mechanism is as follows:

[0086] Attention mechanisms are designed primarily from two dimensions, with different emphases depending on the task. The Convolutional Block Attention Module (CBAM) combines channel and spatial information.

[0087] The core idea of ​​the channel attention mechanism is to focus primarily on meaningful information. Two pooling modes are applied to the input feature image, corresponding to two perspectives describing the background information. The results of the two pooling processes are then processed using a multilayer perceptron with shared weights. The two results are then summed and input into the channel weight judgment function; this invention chooses the sigmoid function.

[0088] Output M c Represented as M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) where F represents the input feature map and σ represents the sigmoid function.

[0089] The core idea of ​​the spatial attention mechanism is to focus on the location of meaningful information. For the input feature image, two abstraction methods are used on the channel of each feature point to obtain a multi-dimensional description of the feature image. After concatenating the abstract results, a convolutional kernel is used to further abstract and extract features from the concatenated result, which is then fed into a spatial weighting function; this invention uses the sigmoid function. At this point, the module obtains the weights for each feature point in the input feature layer. Based on these spatial feature weights, multiplying these weights by the original input feature layer yields the result after spatial attention processing. The output result M is... s It can be represented as: M s (F)=σ(f n×n ([AvgPool(F); MaxPool(F)])) where F represents the input feature map, f n ×n This represents a convolution kernel operation of size n×n. Using CBAM allows for more efficient allocation of network computing power.

[0090] Hybrid pooling methods specifically include:

[0091] During feature extraction, max pooling is used in shallow layers and average pooling in deep layers. Both pooling methods divide the feature map into regions. Max pooling outputs the maximum value within its effective range, while average pooling averages the effective range to obtain the final result. Comparatively, max pooling focuses on preserving edge information, while average pooling focuses on preserving background information. However, for GPR echo images, both edge and background information contain significant amounts of information. Combining the two can achieve higher model accuracy. This invention uses hybrid pooling to optimize the model. Hybrid pooling is based on probability, selecting between the two pooling methods and combining their characteristics. See Tables 3 and 4 for details.

[0092] Table 3 shows the results of identifying hybrid pooling weights in shallow pooling.

[0093]

[0094] Table 4 shows the results of identifying hybrid pooling weights in deep pooling.

[0095]

[0096] In step 6,

[0097] In the network parameter settings, the batch size is 16, the initial learning rate is 0.001, which is halved as the number of iterations increases, the optimizer is SGD, its momentum is 0.9, and the weight decay is 0.0005.

[0098] In step 7,

[0099] The specific method for target recognition and classification of ground-penetrating radar (GPR) underground cavity target echo images involves using a trained RefineDet model to extract features from GPR underground cavity target echo images that have not been previously input into the system. These features are then fed into the model to automatically detect targets within the GPR underground cavity target echo images. The detected values ​​output the cavity location, confidence level, length, and height.

[0100] Example 2:

[0101] Combination Figures 1-12 This invention proposes an automatic detection method for underground cavity targets using a RefineDet network model for ground-penetrating radar. The method specifically includes:

[0102] Step 1: Perform background removal on the acquired ground-penetrating radar echo image of the underground cavity target to obtain a ground-penetrating radar echo image with suppressed transverse ripples.

[0103] Step 2: Gain the ground-penetrating radar echo image generated in Step 1, suppress the background, and extract the cavity features submerged in the image;

[0104] Step 3: Denoise the image data after gain processing in Step 2 to suppress the influence of clutter;

[0105] Step 4: Pre-screen the ground-penetrating radar echo images processed in Step 3. Based on the results of manual identification and on-site confirmation, use labelimg to mark the voids in the images and label them with "void".

[0106] Step 5: The image data obtained in Step 4 is used to create a detection network dataset in PASCAL VOC dataset format, where 80% of the targetless and holeless images are assigned to the network for training, and 20% of the targetless and holeless images are assigned to the network for testing.

[0107] Step 6: Input the training set from the dataset obtained in Step 5 into the RefineDet network and train it to obtain the weight model;

[0108] Step 7: Input the test set from the dataset obtained in Step 5 into the obtained weight model to perform target detection on the ground-penetrating radar echo image of the underground cavity target.

[0109] Step 1, which involves preprocessing the acquired ground-penetrating radar echo image of the underground cavity target, specifically involves removing standing waves and filtering the acquired ground-penetrating radar echo image of the underground cavity target. Standing waves are removed by direct cropping, and the image is filtered using a transverse mean filtering method to obtain a ground-penetrating radar echo image with suppressed transverse ripples.

[0110] In step 2,

[0111] The nodalized mean linear gain method is used to extract the features of cavitary targets from redundant background information. Nodalized mean linear gain can highlight the curve features of cavities in the background, providing a clearer understanding of the location and shape of cavitary targets.

[0112] In step 2, the nodal mean linear gain method specifically includes:

[0113] First, the image is divided into 7 equal parts vertically. Then, the starting row of each part and the last row of the image are each assigned a node, resulting in a total of 8 nodes.

[0114] Then, the average value of the maximum pixel value of each row in each part is taken as the gain value of the corresponding node. The pre-gain curve is obtained by linear interpolation, and the maximum value is used to calibrate the curve to obtain the gain curve. Each row of the image corresponds to a point on the curve, and the size of the point is the gain value of the row. The image is then gained according to the gain curve.

[0115] In step 3,

[0116] The noise reduction process specifically includes: fast nonlocal mean denoising, wavelet denoising, singular value decomposition, discrete cosine transform, and Gaussian filtering.

[0117] The Fast Non-Local Means (NL-means) denoising algorithm is an accelerated version of Non-Local Means. NL-means filters based on the similarity between pixels.

[0118] For an image, a search bounding box of size D×D is selected, and a neighborhood bounding box of size d×d centered at x and y is selected. The similarity between two neighborhoods is measured as follows: Here, the size of the neighborhood is m×n, and x(i,j) and y(i,j) are the pixel values ​​within two neighborhoods. This similarity is used to obtain the metric for each pixel. Where h represents the smoothing factor, which affects the distortion of the filter. The final filtered result for point x is: NLmeans(x)=∑w(x,y)·y

[0119] The fast algorithm addresses the time-consuming point-by-point calculation of the original method by constructing an integral image of the pixels, thereby accelerating filtering. This invention constructs a 5×5 search window and a 3×3 neighborhood window.

[0120] In step 4,

[0121] The processed ground-penetrating radar echo images were pre-screened. Based on the results of manual identification and on-site confirmation, the voids in the images were marked using labelimg, and the void targets were labeled "void".

[0122] In step 5,

[0123] The object detection dataset consists of three main parts: JPEGImages, ImageSets, and Annotations. JPEGImages stores the images of the target objects. The ImageSets folder contains a Main folder containing related .txt files, such as train.txt for training and test.txt for testing. These .txt files store a list of image filenames used to perform their respective tasks. The Annotations folder stores .xml files, which contain relevant information for each image, such as image address, bounding box coordinates, the category of the object within the bounding box, and whether the object is a hard sample.

[0124] In step 6,

[0125] The RefineDet network model used in this invention has 4, 5, or 6 layers. The backbone network of the RefineDet model consists of ResNet50, Ghost_ResNet50, and their hybrid pooling models with different proportions, which include max pooling weights and average pooling weights increasing from 0 to 1 with a step size of 0.1.

[0126] The Ghost_ResNet50 is based on the fact that feature maps of the same depth have a certain degree of similarity, and obtains the required feature maps in a less computationally intensive way.

[0127] The Ghost module works as follows: Assuming the original process involves an input image of size w×h, after being processed by n sets of k×k convolutional kernels, an output with n channels and size w'×h' is generated. The Ghost module uses a simple operation Φ to replace the traditional convolution operation, obtaining the Ghost feature map: First, the input image is processed with m sets of k×k convolutional kernels to generate an intrinsic feature map with m channels and size w'×h'. Then, the operation Φ is applied to the intrinsic feature map to generate the Ghost feature map. The combined results are the module's output.

[0128] For the operation Φ, this invention uses depth-wise convolution (DWC). Unlike ordinary convolution, in DWC, one convolution kernel corresponds to one channel, and a single convolution kernel only processes the corresponding channel. The operation Φ can be equivalent to a d×d convolution kernel operation.

[0129] The RefineDet model's connection blocks are divided into original connection blocks and connection blocks with added attention mechanisms.

[0130] The attention mechanism is as follows:

[0131] Attention mechanisms are designed primarily from two dimensions: channels and space.

[0132] The core idea of ​​the channel attention mechanism is to focus on meaningful information. Two pooling modes are applied to the input feature image, corresponding to two perspectives describing the background information. The results of the two pooling processes are then processed using a multilayer perceptron with shared weights. Finally, the two results are summed and input into the channel weight judgment function, where the sigmoid function is chosen.

[0133] Output M c Represented as M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) where F represents the input feature map and σ represents the sigmoid function.

[0134] The core idea of ​​the spatial attention mechanism is to focus on the location of meaningful information. For the input feature image, two abstraction methods are used on each feature point's channel to obtain a multi-dimensional description of the feature image. After concatenating the abstract results, a convolutional kernel is used to further abstract and extract features from the concatenated result, which is then fed into a spatial weighting function using the sigmoid function. At this point, the module obtains the weights for each feature point in the input feature layer. Based on these spatial feature weights, multiplying these weights by the original input feature layer yields the result after spatial attention processing. The output result M is... s It can be represented as: M s (F)=σ(f n×n ([AvgPool(F); MaxPool(F)])) where F represents the input feature map, f n×n This represents a convolution kernel operation of size n×n. Using CBAM allows for more efficient allocation of network computing power. See Tables 5 and 6 for details.

[0135] Table 5. RefineDet detection results for the attention mechanism of the 4-layer structure.

[0136]

[0137] Table 6. RefineDet test results for multi-layer structures

[0138]

[0139] Hybrid pooling methods specifically include:

[0140] When performing feature extraction, max pooling is used in shallow layers and average pooling is used in deep layers. This invention uses hybrid pooling to optimize the model. Hybrid pooling is based on probability and selects between two pooling methods, combining the characteristics of both.

[0141] In step 6,

[0142] In the network parameter settings, the batch size is 16, the initial learning rate is 0.001, which is halved as the number of iterations increases, the optimizer is SGD, its momentum is 0.9, and the weight decay is 0.0005.

[0143] In step 7,

[0144] The specific method for target recognition and classification of ground-penetrating radar (GPR) underground cavity target echo images involves using a trained RefineDet model to extract features from GPR underground cavity target echo images that have not been previously input into the system. These features are then fed into the model to automatically detect targets within the GPR underground cavity target echo images. The detected values ​​output the cavity location, confidence level, length, and height.

[0145] This invention proposes an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the automatic detection method for underground cavity targets using a RefineDet network model of ground penetrating radar.

[0146] This invention proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the automatic detection method for underground cavity targets using a RefineDet network model of ground penetrating radar.

[0147] The memory in this application embodiment can be volatile memory or non-volatile memory, or it can include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0148] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0149] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0150] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0151] The above provides a detailed description of the automatic detection method for underground cavity targets using a RefineDet network model proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A ground penetrating radar underground cavity target automatic detection method of a RefineDet network model, characterized in that, The method specifically comprises: Step 1: background elimination is performed on the obtained ground penetrating radar echo image of the underground cavity target, so as to obtain a ground penetrating radar echo image in which transverse ripples are suppressed; Step 2: gain is performed on the ground penetrating radar echo image generated in step 1, background is suppressed, and a cavity feature submerged in the image is extracted; Step 3: noise reduction is performed on the image data after gain processing in step 2, and the influence of clutter is suppressed; Step 4: pre-screening is performed on the ground penetrating radar echo image after processing in step 3, a cavity in the image is labeled by referring to the results of manual identification and field confirmation, and using labelimg; Step 5: the image data obtained in step 4 is made into a detection network data set in the PASCAL VOC data set format; Step 6: the training set in the data set obtained in step 5 is input into a RefineDet network, and the training is performed, so as to obtain a weight model; Step 7: the test set in the data set obtained in step 5 is input into the obtained weight model, and target detection is performed on the ground penetrating radar echo image of the underground cavity target; The node mean linear gain method is used to extract the cavity target feature from the redundant background information, the node mean linear gain can highlight the curve feature of the cavity in the background, and the position and shape feature of the cavity target can be more clearly obtained; The node mean linear gain method specifically comprises: First, the picture is divided into 7 parts in the vertical direction, and then the starting row of each part and the final row of the image correspond to a node, that is, there are 8 nodes in total; Then, the mean value of the maximum value of each row of pixels in each part is taken as the gain size of the corresponding node, a pre-gain curve is obtained through linear interpolation, and the pre-gain curve is calibrated using the maximum value to obtain a gain curve, each row of the image corresponds to a point on the gain curve, and the size of the point is the gain size of the row, and the image is gain processed according to the gain curve.

2. The method of claim 1, wherein, In step 1, the image background is eliminated by using a transverse ripple suppression filtering method, so as to obtain a ground penetrating radar echo image in which transverse ripples are suppressed.

3. The method of claim 1, wherein, The noise reduction adopts fast non-local mean noise reduction.

4. The method of claim 1, wherein, In step 5, the image data is made into a detection network data set in the PASCAL VOC data set format, wherein part of the target-free image and the cavity image are allocated to the training of the network in a certain proportion, and the remaining part of the target-free image and the cavity image are allocated to the test of the network.

5. The method of claim 1, wherein, The step 6 specifically comprises: The RefineDet network is divided into 4, 5 or 6 layers; the backbone network is divided into ResNet50, Ghost_ResNet50 and their respective mixed pooling models with different proportions, wherein the maximum pooling weight and the average pooling weight increase from 0 to 1 with a step of 0.1, and the connection block is divided into an original connection block and a connection block with an attention mechanism; The Ghost module specifically is: assuming that in the original process, the input image with the size of acts on the convolution kernel of groups , an output with the channel number of and the size of is generated; in the Ghost module, the simple operation is used instead of the traditional convolution operation to obtain the Ghost feature map: first, the input image is processed by the convolution kernel of groups , an intrinsic feature map with the channel number of and the size of is generated, then the operation is applied to the intrinsic feature map to generate the Ghost feature map, and the output of the module is obtained after integration; The operation Adopt deep convolution DWC, One convolution kernel corresponds to one channel, and a single convolution kernel only processes the corresponding channel Equivalent to The convolution kernel operation The attention mechanism specifically comprises: The design is performed from two dimensions of channels and spaces. The processing of the channel attention mechanism is specifically as follows: for an input feature image, two modes of pooling are respectively performed to obtain two angle descriptions of background information. Then, the results of the two modes of pooling are processed by using a multi-layer perception with shared weights, and then the two results are added and input into a channel weight judgment function to select a function. Output is represented as wherein represents an input feature map, is represented as a function; The processing of the spatial attention mechanism is specifically as follows: for an input feature image, two abstraction methods are used on each channel of each feature point to obtain a multi-dimensional description of the feature image; after the abstraction results are spliced, a convolution kernel is used to abstract and further extract features from the spliced results; finally, the spatial weight judgment function is used At this point, the module obtains the weight of each feature point of the input feature layer, and based on the spatial feature weight, the weight is multiplied by the original input feature layer to obtain the result after spatial attention processing. Output result is represented as: wherein represents an input feature map, represents a convolution kernel operation of size .

6. The method of claim 5, wherein, The mixed pooling of the backbone network specifically comprises: when performing feature extraction, maximum pooling is used in the shallow layer, and average pooling is used in the deep layer. 7.An electronic device comprising a memory and a processor, the memory storing a computer program, wherein, The processor executes the computer program to realize the steps of the method in any one of claims 1-6.

8. A computer readable storage medium for storing computer instructions, characterized in that, The computer instructions, when executed by a processor, implement the steps of the method of any of claims 1-6.

Images