Underground structure ground penetrating radar image automatic discrimination method and system based on KAN

Abstract

The invention relates to a KAN-based underground structure ground penetrating radar image automatic discrimination method and system. The method comprises the following steps: acquiring ground penetrating radar GPR images of underground structures containing different defect types by using a ground penetrating radar; performing defect type labeling on each GPR image, wherein the defect types are divided into intact structures, cracks, groundwater accumulation and leakage and abnormal structure deformation; performing data enhancement processing on the labeled data set, and finally randomly dividing the enhanced data set into a training set and a test set; and constructing a ResKANet discrimination model, training the ResKANet discrimination model applied to GPR image discrimination of the underground structure by using the obtained data set, obtaining image defect category discrimination by using the trained ResKANet discrimination model, and outputting a discrimination result. The method gives consideration to relatively high discrimination speed and relatively high discrimination precision, and is suitable for defect detection of an underground structure, especially for high-precision identification of abnormal deformation defects of the structure.

Description

Technical Field

[0001] The present invention relates to the field of underground structure construction, management and maintenance, and in particular to a KAN-based underground structure ground penetrating radar image automatic identification method and system. Background Art

[0002] In the field of underground structure construction, management, and maintenance, ground penetrating radar is widely used. It can achieve excellent performance in concrete structure detection, steel bar detection, pipeline and cable positioning, foundation and foundation detection, historical building assessment, unexploded ordnance and abandoned facility detection, etc. The analysis of GPR images generated by ground penetrating radar detection is generally divided into manual analysis and analysis using a discrimination system. Manual discrimination has the disadvantages of low work efficiency, inability to quickly cover large areas, and difficulty in realizing rapid data processing and analysis; and the quality of manual discrimination is unstable, and the results of discrimination are easily affected by personal experience, skills, and the working status of the day, resulting in low stability and reliability of the detection results. At the same time, manual discrimination also has the problem of low cost-effectiveness.

[0003] Compared with manual analysis, the use of discrimination system analysis can quickly process a large amount of data, complete the analysis and recognition of GPR images in a short time, and is not interfered by subjective factors, and can maintain consistency and objectivity. However, the current discrimination system mainly uses built-in image processing algorithms, morphological operations or threshold segmentation of pixel values, etc., although it can distinguish cracks, leakage, abnormal structural deformation and other defects to a certain extent, but the accuracy is not high. Some intelligent algorithms can only identify a certain defect, especially the detection accuracy of abnormal structural deformation is poor, and it is easy to make mistakes when it cannot cope with noise and irregular targets in complex environments.

[0004] Due to the complexity of the underground structure in this application, the existing identification system cannot meet the dual requirements of speed and accuracy. Therefore, an automatic identification system for underground structure ground penetrating radar images based on a neural network Kolmogorov-Arnold network (KAN) with a learnable activation function is proposed. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the above technologies and provide a KAN-based underground structure ground penetrating radar image automatic identification method and system, which takes into account both high identification speed and high identification accuracy, and is suitable for defect detection in underground structures, especially for high-precision identification of abnormal structural deformation defects.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] In a first aspect, the present invention provides a method for automatically identifying underground structures in ground penetrating radar images based on KAN, the method comprising the following steps:

[0008] Step 1: Use ground penetrating radar to obtain ground penetrating radar (GPR) images of underground structures with different defect types; annotate each GPR image with defect types, and the defect types are divided into intact structure, cracks, groundwater accumulation and leakage, and abnormal structural deformation;

[0009] Step 2: Perform data augmentation on the labeled data set, and finally randomly divide the augmented data set into training set and test set;

[0010] The data enhancement process is as follows: scaling the annotated GPR image to four different sizes, namely 256×256, 288×288, 320×320, and 352×352 pixels; for each size of the image, using a square area with a size larger than 224x224 and smaller than the actual size of the operated image to extract the left, middle, and right sides or the upper, middle, and lower parts of the operated image, and then extracting the four corners and the center of each extracted square area with a size of 224x224 squares, and scaling each extracted square area to a 224x224 square; finally, mirroring all the obtained 224x224 images, i.e., flipping left and right; a total of 144 cropping areas are obtained for each annotated GPR image;

[0011] Step 3: Build the ResKANet discriminant model

[0012] The obtained data set is used to train the ResKANet discriminant model applied to the GPR image discrimination of underground structures, and the trained ResKANet discriminant model is used to obtain the image defect category discrimination and output the discrimination result;

[0013] The ResKANet discriminant model includes a convolution layer composed of three 3*3 convolution kernels connected in sequence, a 3*3 pooling layer, four convolution groups with different numbers of channels connected to the KAN network, an average pooling layer, a fully connected layer, a regularization layer and a Softmax activation function; a KAN network is set after each convolution group, and the feature image after each convolution of a convolution group is input into the KAN network, and then enters the next convolution group; after the convolution of the four convolution groups, it is input into the average pooling layer to obtain a more representative feature vector, and the vector is then input into the fully connected layer of 1000 units to obtain a 1000-dimensional vector, and the obtained 1000-dimensional vector is used as the input of the regularization technology dropout, so that some neuron outputs are randomly set to zero to obtain a regularized vector, and then the regularized vector is input into the Softmax activation function, and finally the output of the model is converted into a probability distribution, so as to obtain the final prediction result, and then the defect type is discriminated;

[0014] Step 4: Scale the GPR image to be identified to a size of 224x224 and input it into the trained ResKANet discrimination model to identify the defect categories of the underground structure ground penetrating radar image.

[0015] Furthermore, the four convolution groups are connected in sequence: a low-channel 1*1 convolution kernel, a low-channel 3*3 convolution kernel, and a high-channel 1*1 convolution kernel. The number of repetitions and the number of channels of each convolution group are different. The first convolution group is repeated 3 times, the second convolution group is repeated 4 times, the third convolution group is repeated 6 times, and the fourth convolution group is repeated 3 times; the high channel number is twice the low channel number, the low channel number of the first convolution group is 128, and the high channel number is 256; the low channel number of the fourth convolution group is 1024, and the high channel number is 2048.

[0016] Furthermore, the accuracy of the bottom radar image is above 90%, and the floating point calculation is 3.4×10 11 ; The accuracy of abnormal structural deformation is over 90%.

[0017] According to a second aspect, an electronic device includes a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of the method when executing the computer program.

[0018] In a third aspect, a KAN-based automatic identification system for underground structure ground penetrating radar images is provided, the system comprising:

[0019] Ground Penetrating Radar, used to obtain GPR images of underground structures during the inspection process;

[0020] The image processing module is used to label the GPR images of the image acquisition module according to their categories and perform data enhancement processing to obtain a data set, which is divided into a training set and a test set;

[0021] The GPR images include four situations: intact structure, cracks, leakage, and abnormal structural deformation;

[0022] The data enhancement process is as follows: scaling the annotated GPR image to four different sizes, namely 256×256, 288×288, 320×320, and 352×352 pixels; for each size of the image, using a square area with a size larger than 224x224 and smaller than the actual size of the operated image to extract the left, middle, and right sides or the upper, middle, and lower parts of the operated image, and then extracting the four corners and the center of each extracted square area with a size of 224x224 squares, and scaling each extracted square area to a 224x224 square; finally, mirroring all the obtained 224x224 images, i.e., flipping left and right; a total of 144 cropping areas are obtained for each annotated GPR image;

[0023] The ResKANet discriminant model is used to discriminate GPR images. It is connected to the ground penetrating radar and image processing modules, and the obtained data set is used to train the ResKANet discriminant model for GPR image discrimination applied to underground structures. The trained ResKANet discriminant model is used to obtain image defect category discrimination and output the discrimination result.

[0024] Display the early warning module and obtain the output results of the ResKANet discrimination model. If there are cracks, groundwater accumulation and leakage, or abnormal structural deformation in the underground structure, the discrimination system will issue a corresponding alarm to remind the staff; if there are no related problems, the structure will be displayed as intact;

[0025] The ResKANet discriminant model includes a convolution layer composed of three 3*3 convolution kernels connected in sequence, a 3*3 pooling layer, four convolution groups with different numbers of channels connected to the KAN network, an average pooling layer, a fully connected layer, a regularization layer and a Softmax activation function; a KAN network is set after each convolution group, and the feature image after each convolution of a convolution group will be input into the KAN network, and then enter the next convolution group; after the convolution of the four convolution groups, it is input into the average pooling layer to obtain a more representative feature vector, and the vector is then input into the fully connected layer of 1000 units to obtain a 1000-dimensional vector. The obtained 1000-dimensional vector is used as the input of the regularization technology dropout, and some neuron outputs are randomly set to zero to obtain a regularized vector, and then the regularized vector is input into the Softmax activation function, and finally the output of the model is converted into a probability distribution, so as to obtain the final prediction result, and then the defect type is discriminated.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention creatively applies the KAN network to the automatic identification of ground penetrating radar images of underground structures, realizes the automatic identification of GPR images generated by ground penetrating radar in concrete structure detection, and can accurately and efficiently identify the defect types in GPR images. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the model training process of the KAN-based underground structure ground penetrating radar image automatic identification method of the present invention.

[0029] Figure 2 It is a structural diagram of the ResKANet discrimination model in the present invention.

[0030] Figure 3 It is a schematic diagram of the structure of the KAN network in the present invention. DETAILED DESCRIPTION

[0031] In order to more clearly describe the technical problems, technical solutions and advantages of the present invention, the following will be described in detail with reference to the drawings and embodiments. It should be noted that these embodiments are only used to illustrate the principles and application scope of the present invention and should not be regarded as limiting the present invention.

[0032] The present invention provides a method for automatically distinguishing underground structures from ground penetrating radar images based on KAN, and the steps of the method are:

[0033] Step 1: Use ground penetrating radar to obtain ground penetrating radar (GPR) images of underground structures with different defect types; label each GPR image with defect types, which are divided into intact structure, cracks, groundwater accumulation and leakage (there may be cavities), and abnormal structural deformation;

[0034] Step 2: Perform data augmentation on the labeled data set, and finally randomly divide the augmented data set into training set and test set;

[0035] The data enhancement process is as follows: scaling the annotated GPR image to four different sizes, namely 256×256, 288×288, 320×320, and 352×352 pixels; for each size of the image, using a square area with a size larger than 224x224 and smaller than the actual size of the operated image to extract the left, middle, and right sides or the upper, middle, and lower parts of the operated image, and then extracting the four corners and the center of each extracted square area with a size of 224x224 squares, and scaling each extracted square area to a 224x224 square; finally, mirroring all the obtained 224x224 images, i.e., flipping left and right; a total of 144 cropping areas are obtained for each annotated GPR image;

[0036] Step 3: Build the ResKANet discriminant model

[0037] The obtained data set is used to train the ResKANet discriminant model applied to the GPR image discrimination of underground structures, and the trained ResKANet discriminant model is used to obtain the image defect category discrimination and output the discrimination result;

[0038] The ResKANet discriminant model includes a convolution layer composed of three 3*3 convolution kernels connected in sequence, a 3*3 pooling layer, four convolution groups with different numbers of channels connected to the KAN network, an average pooling layer, a fully connected layer, a regularization layer and a Softmax activation function; a KAN network is set after each convolution group, and the feature image after each convolution of a convolution group is input into the KAN network, and then enters the next convolution group; after the convolution of the four convolution groups, it is input into the average pooling layer to obtain a more representative feature vector, and the vector is then input into the fully connected layer of 1000 units to obtain a 1000-dimensional vector, and the obtained 1000-dimensional vector is used as the input of the regularization technology dropout, so that some neuron outputs are randomly set to zero to obtain a regularized vector, and then the regularized vector is input into the Softmax activation function, and finally the output of the model is converted into a probability distribution, so as to obtain the final prediction result, and then the defect type is discriminated;

[0039] Step 4: Scale the GPR image to be identified to a size of 224x224 and input it into the trained ResKANet discrimination model to identify the defect categories of the underground structure ground penetrating radar image.

[0040] Example 1

[0041] The method for automatically identifying underground structures in ground penetrating radar images based on KAN in this embodiment has the following specific steps:

[0042] Step 1: Use ground penetrating radar to obtain ground penetrating radar (GPR) images of underground structures with different defect types; annotate each GPR image with defect types, and the defect types are divided into intact structure, cracks, groundwater accumulation and leakage, and abnormal structural deformation;

[0043] Step 2: Perform data augmentation on the labeled data set, and finally randomly divide the augmented data set into training set and test set;

[0044] The data enhancement process is as follows: scaling the annotated GPR image to four different sizes, namely 256×256, 288×288, 320×320, and 352×352 pixels; for each size of the image, using a square area with a size larger than 224x224 and smaller than the actual size of the operated image to extract the left, middle, and right sides or the upper, middle, and lower parts of the operated image, and then extracting the four corners and the center of each extracted square area with a size of 224x224 squares, and scaling each extracted square area to a 224x224 square; finally, mirroring all the obtained 224x224 images, i.e., flipping left and right; a total of 144 cropping areas are obtained for each annotated GPR image;

[0045] Step 3: Build the ResKANet discriminant model

[0046] The enhanced GPR image is used as the input of the ResKANet discriminant model. The GPR image is used as the input of the ResKANet discriminant model. It first passes through three convolution layers with 3*3 convolution kernels. The number of output channels of the convolution layer is 64, and a feature image with a height of H, a width of W, and a number of channels of 64 is obtained. The ResNeXt network using three 3*3 convolution kernels can effectively reduce the number of parameters used, and because the output of each position only depends on a 3x3 area instead of a 7x7 area, the amount of calculation is reduced while ensuring accuracy.

[0047] The H*W*64 feature image is used as the input of the 3*3 pooling layer. After pooling, the H / 2*W / 2*64 feature image is obtained. The H / 2*W / 2*64 feature image is used as the input and enters the four convolution groups in turn. The number of channels of these four convolution groups increases successively. They are all composed of two 1*1 convolution kernels and a 3*3 convolution kernel. The number of repetitions of each convolution group is different. The first convolution group is repeated 3 times, the second convolution group is repeated 4 times, the third convolution group is repeated 6 times, and the fourth convolution group is repeated 3 times. The feature image of H / 2*W / 2*64 is used as the first convolution group input. It is first processed by a 1x1 convolution layer to increase the number of channels from 64 to 128, and then processed by group convolution. Here, the convolution kernel of the group convolution is 3x3, the number of groups is 32, and the number of channels it outputs is also equal to 128. Then it is dimensionalized by 1x1 convolution to increase the number of channels from 128 to 256 to obtain the new image feature F(x). Finally, its output F(x) is added to the input x to obtain the final output H(x). The input passed to the next layer is:

[0048] H(x)=F(x)+x

[0049] H(x) is used as the second input of the first convolution group, and the convolution operation is repeated again. Unlike the first convolution operation, the first 1*1 convolution group in the second and third convolutions performs a dimensionality reduction operation, reducing the number of channels from 256 to 128, and then performs group convolution with a group number of 32 and 1*1 dimensionality increase, until the first convolution group is repeated three times, obtaining a feature image G1(x) with a dimension of H / 2*W / 2*256.

[0050] The feature image G1(x) is used as the input of the first KAN network. In the first KAN network, this feature image will be decomposed into 256 groups, each containing H / 2*W / 2 one-dimensional data. Then, an internal single variable function and an external single variable function are learned for each group of data. The internal single variable function Internal(x) maps the input real value x to an output value between 0 and 1 in the form of a logistic sigmoid function. The calculation method is as follows:

[0051] Internal(x)=x / (1+e -x )

[0052] The external univariate function External(x) connects a set of known data points through piecewise linear approximation to form a smooth curve. The calculation method is as follows:

[0053]

[0054] Where n is the number of internal univariate functions Internal(x), c i is the coefficient optimized during training, B i (x) is the B-spline basis function. The output of the KAN network is The calculation is as follows:

[0055]

[0056] Among them, ω is the weight, the internal univariate function is Internal(x), and the external univariate function is External(x).

[0057] The output of the first KAN network is used as the input of the second convolution group, and the convolution operation of the second convolution group is performed again. After repeating the convolution group processing four times, the second convolution group obtains a feature image of H / 2*W / 2*512. The feature image obtained by the second convolution group is input into the second KAN network to improve the classification ability of the model.

[0058] The output of the second KAN network is used as the input of the third convolution group, and the convolution operation of the third convolution group is performed again. After repeating the convolution process 6 times, the third convolution group obtains a feature image of H / 2*W / 2*1024. The feature image obtained by the third convolution group is input into the third KAN network to improve the classification ability of the model.

[0059] The output of the third KAN network is used as the input of the fourth convolution group, and the convolution operation of the fourth convolution group is performed again. After repeating the convolution process three times, the fourth convolution group obtains a feature image of H / 2*W / 2*2048. The feature image obtained by the fourth convolution group is input into the fourth KAN network to improve the classification ability of the model.

[0060] After being processed by the convolution group and the KAN network, the feature image of size H / 2*W / 2*2048 is used as the input of the global average pooling. For each channel c, the calculation formula of the global average pooling is:

[0061]

[0062] Where H / 2 represents the height of the feature map, W / 2 represents the width of the feature map, and Xc(i,j) represents the pixel value at the i-th row and j-th column on the c-th channel. Finally, through the average pooling layer, each channel will get an average value to obtain a 2048-dimensional feature vector. The vector is input into the fully connected layer of 1000 units to obtain a 1000-dimensional vector. The calculation method is as follows:

[0063] Y=WX+b

[0064] Where Y is a 1000-dimensional vector; W is the weight matrix of the fully connected layer, which contains the connection strength between each unit and the input feature; b is the bias vector used to adjust the output of each unit. The obtained vector Y is used as the input of the regularization technique dropout. By randomly generating a mask vector M, the elements of the mask vector are multiplied by the corresponding elements of the input vector Y, so as to achieve random zeroing of the output of some neurons. The calculation method is as follows:

[0065] dropout(Y)=(Y1M1,Y2M2,·····Y 1000 M 1000 ,)

[0066] Where Y is the output vector of the fully connected layer, M is the mask vector, and the regularized vector dropout(Y) is used as the input of the Softmax activation function to convert the output into a probability distribution to obtain the defect type. The calculation formula is:

[0067]

[0068] in, is the element in dropout(Y), e is the base of the natural logarithm, and the sum in the denominator is the sum of the exponential functions of all input elements.

[0069] The enhanced training set is read into the ResKANet discriminant model, and all images are iteratively trained in turn. After 40,000 iterations, the training model reaches convergence, that is, when the model training gradient is close to 0 (less than 0.01 can be considered close to 0), the training is stopped, and the optimal network parameters are extracted for prediction. If the training gradient is not close to 0, the error back propagation is performed to adjust the weight parameters. Model calibration stage: 1) Input the test set data to obtain the discrimination result; 2) Compare the actual parameters of the discrimination target used for the test with the preliminary discrimination result to obtain the calibrated ResKANet discriminant model; When predicting, the model first loads the trained parameters and the input image, and uses the trained ResKANet discriminant model to identify the category of each image, thereby realizing the discrimination of underground structure defects.

[0070] The GPR image to be identified is scaled to a size of 224x224 and then input into the trained ResKANet identification model to automatically identify the type of GPR image, such as cracks in the underground structure, groundwater accumulation and leakage, and abnormal structural deformation.

[0071] The ResKANet discrimination model of the present invention greatly improves the GPR image discrimination rate, enabling the system to achieve high-speed and high-precision GPR image discrimination.

[0072] The enhanced images are divided into a training set plus a validation set and a test set in a ratio of 7:3. The training set is used to train the model, and the test set is used to evaluate the model performance under different parameter settings, so as to select the best parameter settings and make the model have better generalization ability. The use of the k-fold cross-validation method in the training process can make full use of the limited data set, more accurately evaluate the model performance, select the best model and parameters, and provide estimates of the model variance and bias.

[0073] The method of the present invention is based on the training and testing of the KAN network, the number of parameters is reduced, and the accuracy is improved. The hardware device used is Intel (R) Core (TM) i9-13900FK CPU @ 3.00GHz, RAM 128GB, NVIDIA GeForce RTX 600048G, the performance of the method of the present invention and the existing model are compared on the hardware device, and the performance test is performed using the established data set. The test results are shown in Table 1 and Table 2.

[0074] Table 1 Comparison of overall recognition accuracy and speed on different models

[0075] Model Params Accuracy FLOPs AlexNet <![CDATA[6.2×10 7 ]]> 57.1％ <![CDATA[1.4×10 11 ]]> ResNet <![CDATA[2.55×10 7 ]]> 82.2％ <![CDATA[3.8×10 11 ]]> ResNeXt <![CDATA[2.50×10 7 ]]> 85.4％ <![CDATA[3.7×10 11 ]]> VGGNet <![CDATA[1.38×10 8 ]]> 87.9％ <![CDATA[1.5×10 12 ]]> ResNeXt-KAN <![CDATA[2.23×10 7 ]]> 84.8％ <![CDATA[3.4×10 11 ]]> ResKANet <![CDATA[2.38×10 7 ]]> 90.8％ <![CDATA[3.4×10 11 ]]>

[0076] Table 2 Comparison of recognition accuracy of different types of defects on different models

[0077]

[0078] Among them, params represents the number of parameters, Accuracy represents the accuracy, and FLOPs represents floating-point operations. The ResNeXt-KAN structure refers to Figure 2 After each convolution group in , the KAN network is not set, and the last fully connected layer is directly replaced by the KAN network. The Res models in the above table are all based on Resnet50.

[0079] Compared with the existing common automatic discrimination algorithms, the present invention significantly reduces the number of parameters while improving accuracy, so as to achieve both accuracy and resolution speed. In addition, the recognition accuracy of abnormal structural deformation is significantly improved by the method of the present invention. When the overall accuracy reaches more than 90%, the accuracy of abnormal structural deformation is increased to more than 94%.

[0080] Example 2

[0081] The underground structure ground penetrating radar image automatic identification system based on KAN in this embodiment includes:

[0082] Ground Penetrating Radar, used to obtain GPR images of underground structures during the inspection process;

[0083] The image processing module is used to label the GPR images of the image acquisition module according to their categories and perform data enhancement processing to obtain a data set, which is divided into a training set and a test set;

[0084] The GPR images include four situations: intact structure, cracks, leakage, and abnormal structural deformation;

[0085] The data enhancement process is as follows: scaling the annotated GPR image to four different sizes, namely 256×256, 288×288, 320×320, and 352×352 pixels; for each size of the image, using a square area with a size larger than 224x224 and smaller than the actual size of the operated image to extract the left, middle, and right sides or the upper, middle, and lower parts of the operated image, and then extracting the four corners and the center of each extracted square area with a size of 224x224 squares, and scaling each extracted square area to a 224x224 square; finally, mirroring all the obtained 224x224 images, i.e., flipping left and right; a total of 144 cropping areas are obtained for each annotated GPR image;

[0086] The ResKANet discriminant model is used to discriminate GPR images. It is connected to the ground penetrating radar and image processing modules, and the obtained data set is used to train the ResKANet discriminant model for GPR image discrimination applied to underground structures. The trained ResKANet discriminant model is used to obtain image defect category discrimination and output the discrimination result.

[0087] The warning module is displayed to obtain the output results of the ResKANet discrimination model. If there are cracks, groundwater accumulation and leakage, or abnormal structural deformation in the underground structure, the discrimination system will issue a corresponding alarm to remind the staff; if there are no related problems, the structure will be displayed as intact.

[0088] Example 3

[0089] Ground penetrating radar obtains GPR images of underground structures;

[0090] Input the GPR image of the underground structure to be detected into the trained ResKANet discriminant model;

[0091] If it is identified as structurally intact, it indicates that the underground structure is intact;

[0092] If it is identified as a crack, it indicates that there are cracks in the underground structure;

[0093] If it is identified as groundwater accumulation leakage, it indicates that there is groundwater in the underground structure and there is a cavity problem;

[0094] If it is identified that the underground structure is deformed, the deformation of the underground structure is displayed.

[0095] Example 4

[0096] The hardware devices used in the system of this embodiment include:

[0097] Processor: As the core component of the present invention, the processor is responsible for controlling and managing the operation of the entire system, including data acquisition, data processing, image recognition and other functions, and needs to have sufficient computing power and parallel processing capabilities to meet real-time requirements. The processor can be in different forms such as single-chip microcomputers, microprocessors, computers, etc. to meet the needs of different application scenarios;

[0098] Memory: Memory can be used to store collected data and historical data for subsequent processing and analysis. It has the characteristics of high speed, high reliability and scalability to meet the needs of long-term stable operation of the system;

[0099] Database: Use database to store and manage collected data, historical data, analysis results and other information;

[0100] Network interface: used for data exchange and communication, with the characteristics of high speed, high stability and high security, ensuring the reliability and security of data transmission.

[0101] The processor is configured to execute computer executable instructions. When the computer executable instructions are executed by the processor, the various steps of the KAN-based underground structure ground penetrating radar image automatic identification method are implemented.

[0102] The memory stores a computer program, which can be executed by the processor to implement the various steps of the above-mentioned KAN-based method for automatic identification of underground structure ground penetrating radar images.

[0103] The database is configured to store and manage computer application data, including various data types and structures, which are applied to the various steps of the above-mentioned KAN-based underground structure ground penetrating radar image automatic identification method.

[0104] The network interface realizes communication and data transmission between computers. The network interface can provide various communication protocols and data transmission methods to meet the communication and data transmission needs of different application scenarios and different requirements.

[0105] The present invention is mainly used for automatic identification of ground penetrating radar images of underground structures. The automatic identification system is used to automatically identify GPR images and determine the conditions of underground structures.

[0106] The present invention aims to solve the problems caused by the slow speed of GPR image recognition of underground structures, the low accuracy of recognition of defects such as cracks, leakage, and abnormal structural deformation (the overall recognition of the three things, bottom-penetrating radar, the accuracy is raised to more than 90%), and especially the low accuracy of recognition of abnormal structural deformation. The ResKANet discrimination model of the present invention reasonably embeds the Kolmogorov-Arnold network (KAN network) to automatically distinguish the type of GPR image of underground structures. For the GPR image of the underground structure, the system can quickly identify the GPR image of the underground structure, determine whether there are cracks, groundwater, and deformation of the underground structure, and automatically alarm to remind the staff. Compared with the prior art, the technical solution has the following advantages and application prospects: it improves the efficiency of GPR image recognition of underground structures, and provides the possibility of faster GPR image recognition under the premise of pursuing high precision. This is of great significance for the GPR image recognition of underground structures and has broad application prospects.

[0107] In the description of this specification, the described specific features, structures, or characteristics may be combined in an appropriate manner in any one or more embodiments or examples.

[0108] It should be understood that each part of the present invention can be implemented by hardware, software or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or hardware stored in a memory and executed by a suitable instruction execution system.

[0109] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0110] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A KAN-based method for automatic identification of underground structure ground penetrating radar images, characterized in that: The steps of the method are: Step 1: Use ground penetrating radar to obtain ground penetrating radar (GPR) images of underground structures with different defect types; annotate each GPR image with defect types, and the defect types are divided into intact structure, cracks, groundwater accumulation and leakage, and abnormal structural deformation; Step 2: Perform data augmentation on the labeled data set, and finally randomly divide the augmented data set into training set and test set; The data enhancement process is as follows: scaling the annotated GPR image to four different sizes, namely 256×256, 288×288, 320×320, and 352×352 pixels; for each size of the image, using a square area with a size larger than 224x224 and smaller than the actual size of the operated image to extract the left, middle, and right sides or the upper, middle, and lower parts of the operated image, and then extracting the four corners and the center of each extracted square area with a size of 224x224 squares, and scaling each extracted square area to a 224x224 square; finally, mirroring all the obtained 224x224 images, i.e., flipping left and right; a total of 144 cropping areas are obtained for each annotated GPR image; Step 3: Build the ResKANet discriminant model The obtained data set is used to train the ResKANet discriminant model applied to the GPR image discrimination of underground structures, and the trained ResKANet discriminant model is used to obtain the image defect category discrimination and output the discrimination result; The ResKANet discriminant model includes a convolution layer composed of three 3*3 convolution kernels connected in sequence, a 3*3 pooling layer, four convolution groups with different numbers of channels connected to the KAN network, an average pooling layer, a fully connected layer, a regularization layer and a Softmax activation function; a KAN network is set after each convolution group, and the feature image after each convolution of a convolution group is input into the KAN network, and then enters the next convolution group; after the convolution of the four convolution groups, it is input into the average pooling layer to obtain a more representative feature vector, and the vector is then input into the fully connected layer of 1000 units to obtain a 1000-dimensional vector, and the obtained 1000-dimensional vector is used as the input of the regularization technology dropout, so that some neuron outputs are randomly set to zero to obtain a regularized vector, and then the regularized vector is input into the Softmax activation function, and finally the output of the model is converted into a probability distribution, so as to obtain the final prediction result, and then the defect type is discriminated; Step 4: Scale the GPR image to be identified to a size of 224x224 and input it into the trained ResKANet discrimination model to identify the defect categories of the underground structure ground penetrating radar image.

2. The method according to claim 1, characterized in that: The four types of convolution groups are connected in sequence: a 1*1 convolution kernel with a low channel number, a 3*3 convolution kernel with a low channel number, and a 1*1 convolution kernel with a high channel number. The number of repetitions and the number of channels of each convolution group are different. The first convolution group is repeated 3 times, the second convolution group is repeated 4 times, the third convolution group is repeated 6 times, and the fourth convolution group is repeated 3 times; the number of high channels is twice the number of low channels. The low channel number of the first convolution group is 128, and the high channel number is 256; the low channel number of the fourth convolution group is 1024, and the high channel number is 2048.

3. The method according to claim 2, characterized in that The accuracy of the bottom radar image is over 90%, and the floating point calculation is 3.4×10 11 ; The accuracy of abnormal structural deformation is over 90%.

4. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to claim 1 or 2 are implemented.

5. A KAN-based automatic identification system for underground structure ground penetrating radar images, characterized in that: The system comprises: Ground Penetrating Radar, used to obtain GPR images of underground structures during the inspection process; The image processing module is used to label the GPR images of the image acquisition module according to their categories and perform data enhancement processing to obtain a data set, which is divided into a training set and a test set; The GPR images include four situations: intact structure, cracks, leakage, and abnormal structural deformation; The data enhancement process is as follows: scaling the annotated GPR image to four different sizes, namely 256×256, 288×288, 320×320, and 352×352 pixels; for each size of the image, using a square area with a size larger than 224x224 and smaller than the actual size of the operated image to extract the left, middle, and right sides or the upper, middle, and lower parts of the operated image, and then extracting the four corners and the center of each extracted square area with a size of 224x224 squares, and scaling each extracted square area to a 224x224 square; finally, mirroring all the obtained 224x224 images, i.e., flipping left and right; a total of 144 cropping areas are obtained for each annotated GPR image; The ResKANet discriminant model is used to discriminate GPR images. It is connected to the ground penetrating radar and image processing modules, and the obtained data set is used to train the ResKANet discriminant model for GPR image discrimination applied to underground structures. The trained ResKANet discriminant model is used to obtain image defect category discrimination and output the discrimination result. Display the early warning module and obtain the output results of the ResKANet discrimination model. If there are cracks, groundwater accumulation and leakage, or abnormal structural deformation in the underground structure, the discrimination system will issue a corresponding alarm to remind the staff; if there are no related problems, the structure will be displayed as intact; The ResKANet discriminant model includes a convolution layer composed of three 3*3 convolution kernels connected in sequence, a 3*3 pooling layer, four convolution groups with different numbers of channels connected to the KAN network, an average pooling layer, a fully connected layer, a regularization layer and a Softmax activation function; a KAN network is set after each convolution group, and the feature image after each convolution of a convolution group will be input into the KAN network, and then enter the next convolution group; after the convolution of the four convolution groups, it is input into the average pooling layer to obtain a more representative feature vector, and the vector is then input into the fully connected layer of 1000 units to obtain a 1000-dimensional vector. The obtained 1000-dimensional vector is used as the input of the regularization technology dropout, and some neuron outputs are randomly set to zero to obtain a regularized vector, and then the regularized vector is input into the Softmax activation function, and finally the output of the model is converted into a probability distribution, so as to obtain the final prediction result, and then the defect type is discriminated.

Images