CN114882236B - Automatic identification method for underground cavity target of ground penetrating radar based on SinGAN algorithm - Google Patents

Abstract

The present invention provides a method for automatic recognition of underground cavity targets by ground penetrating radar based on SinGAN algorithm. Step 1: Use SinGAN algorithm to perform image augmentation processing on the acquired ground penetrating radar echo image of the underground cavity target to obtain the processed ground penetrating radar echo image with similar distribution; Step 2: Annotate the ground penetrating radar echo image generated in step 1 to clarify the pixel position of the relevant target; Step 3: Randomly assign the annotated data in step 2 to a training set and a validation set; Step 4: Use the training set and validation set of step 3 to train the deep learning target recognition algorithm to obtain a weight model; Step 5: Input the weight model obtained in step 4 into the existing deep learning algorithm model to perform target recognition detection on the ground penetrating radar echo image of the underground cavity target. The present invention solves the problem that the existing method is difficult to detect and identify underground cavity targets.

Description

An automatic recognition method for underground cavity targets of ground penetrating radar based on SinGAN algorithm

Technical Field

The invention belongs to the field of target detection by post-processing of ground penetrating radar echograms, and specifically relates to a method for automatic recognition of underground cavity targets by ground penetrating radar based on a SinGAN algorithm.

Background Art

Ground penetrating radar is a non-destructive detection instrument used to detect shallow underground environments. Ground penetrating radar uses the differences in the electromagnetic dielectric constants of underground media. These differences in parameters are reflected in the radar echo data. By processing the echo data, the distribution of the underground environment can be quickly detected and intuitively understood. In order to intuitively present the echo data for manual analysis, it is a common method to list multi-channel echo data horizontally, which is how the B-Scan image commonly used in ground penetrating radar analysis is obtained.

As an important geophysical method of rapid, high-resolution and non-destructive detection, ground penetrating radar has important significance and value in the research and engineering practice of underground collapse cavity detection. Ground penetrating radar technology will not cause structural damage to the road surface and is applicable to various road conditions. Its detection results are real-time and highly accurate, meeting the requirements of highway disease detection for high efficiency, non-destructiveness, accuracy and wide application range, and is suitable for the detection of underground cavities in roads. The ground penetrating radar system can be composed of one or more pairs of transmitting and receiving antennas. Each pair of transmitters and receivers can collect a single B-Scan image by scanning the area of interest. The distribution of the underground environment can be obtained by analyzing and verifying the B-Scan image. At present, the B-Scan images collected in actual projects need to be manually interpreted and interpreted. This method is inefficient and often leads to missed detection or false detection. There are also problems with using some of the current mainstream deep learning methods to detect and identify underground cavity targets. It is difficult to obtain underground cavities that have been confirmed, verified, located and have relevant pattern information. In addition, underground cavities do not have fixed patterns and shapes in B-Scan images. Obtaining a large number of underground cavity samples is also a difficult engineering task.

Summary of the invention

The present invention provides a method for automatically identifying underground cavity targets using a ground penetrating radar based on a SinGAN algorithm, so as to solve the problem that it is difficult to detect and identify underground cavity targets using existing methods.

The present invention is achieved through the following technical solutions:

A method for automatically identifying underground cavity targets using a ground penetrating radar based on a SinGAN algorithm, the method comprising the following steps:

Step 1: Use the SinGAN algorithm to perform image augmentation processing on the acquired ground penetrating radar echo image of the underground cavity target to obtain a processed ground penetrating radar echo image with similar distribution;

Step 2: Annotate the ground penetrating radar echo image generated in step 1 to identify the relevant target pixel positions;

Step 3: Randomly assign the labeled data in step 2 to the training set and the validation set;

Step 4: Use the training set and validation set in step 3 to train the deep learning target recognition algorithm to obtain a weight model;

Step 5: Input the weight model obtained in step 4 into the existing deep learning algorithm model to perform target recognition and detection on the ground penetrating radar echo image of the underground cavity target.

Furthermore, the step 1 uses the SinGAN algorithm to perform image augmentation processing on the acquired ground penetrating radar echo image of the underground cavity target. Specifically, the SinGAN algorithm includes n GANs of different scales, namely _Gn ~ _G0 , with the scales increasing successively. The input of the _Gnth generator is a noise image, and the output result of the training is close to the original image _xn after downsampling. The input of each subsequent GAN is the output result of Gn ₊₁ plus a noise image. Except for _Gn , each remaining GAN learns the detailed information that supplements the output result of the previous GAN.

From the above, we can see that the SinGAN algorithm consists of a pyramid structure {G ₀ ,...,G _N }, and is trained on an image pyramid of x:{x0,...,xN}, where x _n is a downsampled version of x, with a multiplier of r ⁿ . For some r＞1, each generator G _n is responsible for generating a real image sample wrt the color patch distribution in the corresponding image x _n ;

As can be seen above, all generators and discriminators have the same receptive field, so the size of the structure captured during the generation process is decreasing; _Gn maps spatial Gaussian white noise _zN to image samples Right now

In addition to the spatial noise _zn , each generator _Gn also accepts an upsampled version of the coarser-scale image, i.e.

G _nExecute operation

where _ψn is a convolutional block with 5 Conv(3x3)-BatchNorm-LeakyReLus; the training loss for the nth GAN includes an adversarial formula, i.e.

Furthermore, the step 2 of labeling and clarifying the relevant target pixel positions is specifically to obtain the processed ground penetrating radar echo image with similar distribution, specifically by using the labelimg tool to mark the relevant pixel positions of the generated ground penetrating radar echo image, clarify whether each sub-pixel belongs to the hole label and perform global labeling.

Furthermore, the step 4 trains the deep learning target recognition algorithm to obtain the weight model. Specifically, the deep learning target recognition algorithm YOLOv5 is trained with training parameters of 16 batches and 1000 epochs, and finally a trained weight model is obtained.

Furthermore, the step 5 performs target recognition detection on the ground penetrating radar echo image of the underground cavity target, specifically, according to the trained deep learning algorithm model, the ground penetrating radar echo image of the underground cavity target that has not been input into the system is input into the framework, and the ground penetrating radar echo image of the underground cavity target is automatically subjected to target recognition detection, and finally a processed image is output, which includes the suspected underground cavity target selected by the selection box and its confidence.

The beneficial effects of the present invention are:

The present invention randomly generates a ground penetrating radar echo image using a pyramid structure generation network for the extracted features, and then performs deep learning algorithm training and recognition based on the output results. The output results are used to realize underground cavity target recognition in the ground penetrating radar echo image. The method of the present invention is used to detect underground cavity targets in the ground penetrating radar echo image, which can effectively improve the recognition probability.

The present invention can increase the probability of identifying underground cavity targets to more than 90%.

In practice, when ground penetrating radar collects data related to underground cavities, the shape of the underground cavities is random and difficult to predict, and their depth, size, and position are unknown, which creates a great obstacle to its own data collection and subsequent classification and detection based on deep learning. The purpose of the present invention is to use the SinGAN image augmentation algorithm to extract underground cavity target features in the ground penetrating radar echo map, and to randomly generate the ground penetrating radar echo map for the extracted features through a pyramid structure generation network, and then perform deep learning algorithm training and recognition based on the output results, and use the output results to realize the recognition of underground cavity targets in the ground penetrating radar echo image.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for automatically identifying underground cavity targets using a ground penetrating radar based on a SinGAN algorithm according to the present invention.

Figure 2 is a diagram of the generator and discriminator structure of the SinGAN algorithm.

FIG3 is a single acquired ground penetrating radar echo image of an underground cavity target.

FIG4 is a generated GPR echogram having a distribution similar to that of a known underground cavity GPR image.

Figure 5 is a target recognition effect diagram of the ground penetrating radar echo image of the underground cavity target.

DETAILED DESCRIPTION

The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Example 1

A method for automatically identifying underground cavity targets using a ground penetrating radar based on a SinGAN algorithm, the method comprising the following steps:

Step 1: Use the SinGAN algorithm to perform image augmentation processing on the acquired ground penetrating radar echo image of the underground cavity target to obtain a processed ground penetrating radar echo image with similar distribution;

Step 2: Annotate the ground penetrating radar echo image generated in step 1 to identify the relevant target pixel positions;

Step 3: Randomly assign the labeled data in step 2 to the training set and the validation set to increase the robustness of the entire system;

Step 4: Use the training set and validation set in step 3 to train the deep learning target recognition algorithm to obtain a weight model;

Step 5: Input the weight model obtained in step 4 into the existing deep learning algorithm model to perform target recognition and detection on the ground penetrating radar echo image of the underground cavity target.

Furthermore, the step 1 uses the SinGAN algorithm to perform image augmentation processing on the acquired ground penetrating radar echo image of the underground cavity target. Specifically, the SinGAN algorithm includes n GANs of different scales, namely _Gn ~ _G0 , with the scales increasing successively. The input of the _Gnth generator is a noise image, and the output result of the training is close to the original image _xn after downsampling. The input of each subsequent GAN is the result of the output of Gn ₊₁ plus a noise image. Except for _Gn , the remaining GANs learn to supplement the detailed information of the output result of the previous GAN. This is achieved through adversarial training. In the adversarial training, _Gn learns to deceive the relevant discriminator _Dn , which attempts to distinguish the generated patch blocks from the patch blocks in _xn .

From the above, we can see that the SinGAN algorithm consists of a pyramid structure {G ₀ ,...,G _N }, and is trained on an image pyramid of x:{x0,...,xN}, where x _n is a downsampled version of x, with a multiplier of r ⁿ . For some r＞1, each generator G _n is responsible for generating a real image sample wrt the color patch distribution in the corresponding image x _n ;

From the above, we can see that the generation of image samples starts from the coarsest level, and then passes through all generators in sequence until the finest level, and injects noise at each level; all generators and discriminators have the same receptive field, so the size of the structure captured in the generation process is decreasing; at the coarse scale, this generation is pure generation, _Gn maps spatial Gaussian white noise _zN to image samples Right now

The effective receptive field of this layer (which is better than a layer consisting of a generator and a discriminator, so this layer is worth the generator in the previous paragraph) is usually half the height of the image, so _Gn generates the overall layout of the image and the global structure of the object. Each generator Gn (n < N) at a smaller scale adds details that were not generated at the previous scale. Therefore, in addition to the spatial noise _zn , each generator _Gn also accepts an upsampled version of the coarser scale image, i.e.

All generators have a similar architecture, as shown in Figure 2. Specifically, noise _zn is added to the image is fed into a sequence of convolutional layers. This ensures that the GAN does not ignore the noise. The convolutional layer is used to generate the missing details. That is, G _n performs the operation

where _ψn is a convolutional block with 5 Conv(3x3)-BatchNorm-LeakyReLus; it starts with 32 kernels per block at the coarsest scale and increases by a factor of 2 every 4 scales. Because the generator is fully convolutional, it can generate images of arbitrary size and aspect ratio at test time (by changing the size of the noise map);

Our multi-scale architecture is trained sequentially from the coarsest scale to the finest scale, and once each GAN is trained, it is fixed; the training loss for the nth GAN consists of an adversarial formulation, i.e.

By training a new unconditional generative model, SinGAN, on a single natural image, we use a dedicated multi-scale adversarial training scheme to learn image patch statistics across multiple scales; this can then be used to generate realistic new image samples that preserve the original patch distribution while creating new object configurations and structures.

The SinGAN algorithm is used to perform image augmentation processing on the acquired single ground penetrating radar echo image of the underground cavity target (as shown in FIG3 ), and the processed ground penetrating radar echo image with similar distribution is shown in FIG4 .

Furthermore, the step 2 of labeling and clarifying the relevant target pixel positions is specifically to obtain the processed ground penetrating radar echo image with similar distribution, specifically by using the labelimg tool to mark the relevant pixel positions of the generated ground penetrating radar echo image, clarify whether each sub-pixel belongs to the hole label and perform global labeling.

Furthermore, the step 4 trains the deep learning target recognition algorithm to obtain the weight model. Specifically, the deep learning target recognition algorithm YOLOv5 is trained with training parameters of 16 batches and 1000 epochs, and finally a trained weight model is obtained.

Furthermore, the step 5 performs target recognition detection on the ground penetrating radar echo image of the underground cavity target, specifically, according to the trained deep learning algorithm model, the ground penetrating radar echo image of the underground cavity target that has not been input into the system is input into the framework, and the ground penetrating radar echo image of the underground cavity target is automatically subjected to target recognition detection, and finally a processed image is output, which includes the suspected underground cavity target selected by the selection box and its confidence.

Claims (4)

1. A method for automatic recognition of underground cavity targets by ground penetrating radar based on SinGAN algorithm, characterized in that the automatic recognition method comprises the following steps: Step 1: Use the SinGAN algorithm to perform image augmentation processing on the acquired ground penetrating radar echo image of the underground cavity target to obtain a processed ground penetrating radar echo image with similar distribution; Step 2: Annotate the ground penetrating radar echo image generated in step 1 to identify the relevant target pixel positions; Step 3: Randomly assign the labeled data in step 2 to the training set and the validation set; Step 4: Use the training set and validation set in step 3 to train the deep learning target recognition algorithm to obtain a weight model; Step 5: Input the weight model obtained in step 4 into the existing deep learning algorithm model to perform target recognition and detection on the ground penetrating radar echo image of the underground cavity target; The step 1 uses the SinGAN algorithm to perform image augmentation processing on the acquired ground penetrating radar echo image of the underground cavity target. Specifically, the SinGAN algorithm includes n GANs of different scales, namely _Gn ~ _G0 , with the scales increasing successively. The input of the _Gnth generator is a noise image, and the output result of the training is close to the original image _xn after downsampling. The input of each subsequent GAN is the output result of Gn ₊₁ plus a noise image. Except for _Gn , each remaining GAN learns the detailed information that supplements the output result of the previous GAN. From the above, we can see that the SinGAN algorithm consists of a pyramid structure {G ₀ ,...,G _N }, and is trained on an image pyramid of x:{x0,...,xN}, where x _n is a downsampled version of x, with a multiplier of r ⁿ . For some r＞1, each generator G _n is responsible for generating a real image sample wrt the color patch distribution in the corresponding image x _n ; As can be seen above, all generators and discriminators have the same receptive field, so the size of the structure captured during the generation process is decreasing; _Gn maps spatial Gaussian white noise _zN to image samples Right now In addition to the spatial noise _zn , each generator _Gn also accepts an upsampled version of the coarser-scale image, i.e. G _nExecute operation Where ψ _n is a convolutional block with 5 Conv(3x3)-BatchNorm-LeakyReLu; The training loss for the nth GAN consists of an adversarial formula, namely 2. According to the method for automatic identification of underground cavity targets by ground penetrating radar based on SinGAN algorithm in claim 1, it is characterized in that the step 2 of marking and clarifying the relevant target pixel positions is specifically to obtain the processed ground penetrating radar echo image with similar distribution, specifically by marking the relevant pixel positions of the generated ground penetrating radar echo image by using the labelimg tool, clarifying whether each sub-pixel belongs to the cavity label and performing global labeling. 3. According to the method for automatic recognition of underground cavity targets by ground penetrating radar based on SinGAN algorithm in claim 1, it is characterized in that the step 3 trains the deep learning target recognition algorithm to obtain the weight model, specifically, the deep learning target recognition algorithm YOLOv5 is trained with training parameters of 16 batches and 1000 epochs, and finally a trained weight model is obtained. 4. According to the method for automatic identification of underground cavity targets by ground penetrating radar based on SinGAN algorithm in claim 1, it is characterized in that the step 5 performs target identification and detection on the ground penetrating radar echo image of the underground cavity target, specifically, according to the trained deep learning algorithm model, the newly acquired ground penetrating radar echo image of the underground cavity target is input into the SinGAN algorithm model, and the ground penetrating radar echo image of the underground cavity target is automatically identified and detected, and finally a processed image is output, which includes the suspected underground cavity target selected in the box and its confidence.