KAN-based ground penetrating radar image automatic discrimination method and system for underground structure

Abstract

The application discloses a KAN-based automatic ground structure ground penetrating radar image discrimination method and system, and steps of the method are as follows: ground penetrating radar is used to obtain ground penetrating radar (GPR) images of underground structures containing different defect types; each GPR image is labeled with a defect type, and the defect type is divided into a structure intact, a crack, groundwater accumulation leakage and structural abnormal deformation; a data set after labeling is subjected to data enhancement processing, and finally, the enhanced data set is randomly divided into a training set and a test set; a ResKANet discrimination model is constructed, the obtained data set is used to train the ResKANet discrimination model applied to GPR image discrimination of underground structures, the trained ResKANet discrimination model is used to obtain image defect category discrimination, and a discrimination result is output. The method takes into account high discrimination speed and high discrimination precision, is suitable for defect detection of underground structures, and is especially suitable for high-precision identification of structural abnormal deformation defects.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of underground structure construction, management and maintenance, and in particular to a KAN-based automatic ground penetrating radar image discrimination method and system for underground structures. BACKGROUND

[0002] In the field of underground structure construction, management and maintenance, ground penetrating radar is widely used, and can achieve concrete structure detection, steel bar detection, pipeline and cable positioning, foundation and base detection, historical building evaluation, unexploded ordnance and abandoned facility detection, etc. The analysis of the GPR image generated by the 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 a large area, and difficulty in achieving rapid data processing and analysis. Moreover, the quality of manual discrimination is unstable, and the results of discrimination are easily affected by personal experience, skills and the working state of the day, resulting in low stability and reliability of the detection results, and low cost-effectiveness of manual discrimination.

[0003] Compared with manual analysis, analysis using a discrimination system can quickly process a large amount of data, complete the analysis and identification of the GPR image in a short time, and is not affected by subjective factors, thereby maintaining consistency and objectivity. However, the current discrimination system mainly uses built-in image processing algorithms, such as morphological operations or threshold segmentation of pixel values, to discriminate defects such as cracks, leaks and structural abnormal deformation. However, the accuracy is not high, and some intelligent algorithms can only identify one type of defect. In particular, the detection accuracy of structural abnormal deformation is poor, and it is difficult to deal with noise and irregular targets in complex environments, which may lead to discrimination errors.

[0004] Due to the complexity of underground structures in the present application, the existing discrimination system cannot meet the dual requirements of speed and accuracy. Therefore, a KAN-based automatic ground penetrating radar image discrimination system for underground structures is proposed, which has a learnable activation function. SUMMARY

[0005] The present application overcomes the shortcomings of the above-mentioned technologies and provides a KAN-based automatic ground penetrating radar image discrimination method and system for underground structures, which balances high discrimination speed and high discrimination accuracy, and is suitable for defect detection of underground structures, especially high-precision identification of structural abnormal deformation defects.

[0006] To achieve the above-mentioned purposes, the technical solution adopted by the present application is as follows:

[0007] In a first aspect, the present application provides a KAN-based automatic identification method for ground penetrating radar images of underground structures, the steps of the method being:

[0008] Step 1: Obtain ground penetrating radar (GPR) images of underground structures containing different defect types using a ground penetrating radar; label each GPR image with a defect type, and the defect types include 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 a training set and a test set;

[0010] The data augmentation process is as follows: scale the labeled GPR images to four different sizes, namely 256x256, 288x288, 320x320 and 352x352 pixels; for each size of the picture, extract the left, middle and right or upper, middle and lower parts of the picture with a square region whose size is greater than 224x224 and less than the actual size of the picture being operated, then extract the four corners and center of each extracted square region with a 224x224 square, and then scale each extracted square region to a 224x224 square; finally, mirror process all the obtained 224x224 pictures, i.e. left-right flipping; a total of 144 cropped regions are obtained for each labeled GPR image;

[0011] Step 3: Construct a ResKANet discrimination model

[0012] Use the obtained data set to train a ResKANet discrimination model for GPR image discrimination of underground structures, and use the trained ResKANet discrimination model to obtain image defect classification and output the discrimination results;

[0013] The ResKANet discriminant model comprises a convolution layer composed of three 3*3 convolution kernels connected in sequence, a 3*3 pooling layer, four convolution groups with different channel numbers connected to the KAN network, an average pooling layer, a full connection layer, a regularization layer and a Softmax activation function; each convolution group is provided with a KAN network, and the feature image after the convolution of each convolution group is input into the KAN network and then into the next convolution group; the features after the convolution of the four convolution groups are input into the average pooling layer to obtain more representative feature vectors, the vectors are input into the full connection layer with 1000 units to obtain a 1000-dimensional vector, the 1000-dimensional vector is input into the regularization technology dropout to randomly set zero for part of the neurons to obtain a vector after regularization processing, then the vector after regularization processing is input into the Softmax activation function, finally the output of the model is converted into a probability distribution, so as to obtain the final prediction result and further discriminate the defect type.

[0014] Step 4: input the GPR image to be discriminated into the trained ResKANet discriminant model after scaling to 224x224 size, to realize the recognition of the defect type of the ground structure ground penetrating radar image.

[0015] Further, the four convolution groups are connected in sequence and comprise 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 channel number 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] Further, the accuracy of the ground penetrating radar image is 90% or more, and the floating point operation is 3.4x10 11 The accuracy of the structural abnormal deformation is 90% or more.

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

[0018] In the third aspect, a KAN-based ground structure ground penetrating radar image automatic discrimination system comprises:

[0019] A ground penetrating radar is used to obtain the ground penetrating radar GPR image of the underground structure in the detection process.

[0020] An image processing module is configured to label the GPR image obtained by the image acquisition module with a corresponding category and perform data enhancement processing to obtain a data set, which is divided into a training set and a test set;

[0021] The GPR image includes four cases of intact structure, cracks, leakage, and abnormal structural deformation.

[0022] The data enhancement processing process is as follows: the labeled GPR image is scaled to four different sizes, i.e., 256x256, 288x288, 320x320, and 352x352 pixels; for each size of the picture, a square region with a size greater than 224x224 and less than the actual size of the operated picture is used to extract the left, middle, and right or upper, middle, and lower parts of the operated picture, and then a 224x224 square is extracted at the four corners and center of each extracted square region, and each extracted square region is scaled to a 224x224 square; finally, all the obtained 224x224 pictures are subjected to mirror processing, i.e., left-right flipping; a total of 144 cropped regions are obtained for each labeled GPR image.

[0023] A ResKANet discriminant model is configured to realize discrimination of the GPR image; the ResKANet discriminant model is connected with the ground penetrating radar and the image processing module, and is trained by using the obtained data set to apply to the discrimination of the GPR image of the underground structure; the ResKANet discriminant model is used to obtain the discrimination of the image defect category after being trained, and outputs the discrimination result.

[0024] A display warning module is configured to obtain the output result of the ResKANet discriminant model; if the underground structure has cracks, groundwater accumulation leakage, and abnormal structural deformation, the discrimination system issues a corresponding alarm to remind the staff; if there is no related problem, the structure is displayed as intact.

[0025] The ResKANet discriminant model comprises a convolution layer composed of three 3*3 convolution kernels connected in sequence, a 3*3 pooling layer, four convolution groups with different channel numbers connected with a KAN network, an average pooling layer, a full connection layer, a regularization layer and a Softmax activation function; the KAN network is arranged after each convolution group, and the feature image after the convolution of each convolution group is input into the KAN network and then into the next convolution group; the features after the convolution of the four convolution groups are input into the average pooling layer to obtain more representative feature vectors, the vectors are input into the full connection layer with 1000 units to obtain a 1000-dimensional vector, the 1000-dimensional vector is taken as the input of the regularization technology dropout, part of the neuron outputs is randomly set to zero to obtain a vector after regularization processing, then the vector after regularization processing is input into the Softmax activation function, finally the output of the model is converted into a probability distribution, so that the final prediction result is obtained, and then the defect type is discriminated.

[0026] Compared with the prior art, the beneficial effects of the present application are:

[0027] The KAN network is creatively applied to the automatic discrimination of the ground penetrating radar image of the underground structure in the present application, the automatic discrimination of the GPR image generated in the detection of the concrete structure by the ground penetrating radar is realized, and the defect type in the GPR image can be accurately and efficiently discriminated. BRIEF DESCRIPTION OF DRAWINGS

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

[0029] Figure 2 Fig. 2 is a structure schematic diagram of the ResKANet discriminant model in the present application.

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

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

[0032] The KAN-based automatic discrimination method of the ground penetrating radar image of the underground structure of the present application, the steps of the method are:

[0033] Step 1, obtain ground penetrating radar (GPR) images of underground structures containing different defect types by using ground penetrating radar; label each GPR image with a defect type, and the defect type is divided into the following categories: structure intact, presence of cracks, groundwater accumulation and leakage (may have 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 a training set and a test set;

[0035] The data augmentation process is as follows: scale the labeled GPR images to four different sizes, namely 256x256, 288x288, 320x320, and 352x352 pixels; for each size of the picture, extract the left, middle, and right or upper, middle, and lower parts of the picture with a square region whose size is greater than 224x224 and less than the actual size of the picture being operated, and then extract the four corners and center of each extracted square region with a 224x224 square, and then scale each extracted square region to a 224x224 square; finally, mirror process all the obtained 224x224 pictures, i.e., left-right flipping; a total of 144 cropped regions are obtained for each labeled GPR image;

[0036] Step 3, construct a ResKANet discriminant model

[0037] Use the obtained data set to train a ResKANet discriminant model for GPR image discrimination of underground structures, and use the trained ResKANet discriminant model to obtain image defect classification 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 channel numbers connected to the KAN network, an average pooling layer, a fully connected layer, a regularization layer, and a Softmax activation function; KAN network is set after each convolution group, and the feature image after each convolution group is input into the KAN network and then into the next convolution group; the convolution of the four convolution groups is input into the average pooling layer to obtain a more representative feature vector, the vector is input into the 1000-unit fully connected layer to obtain a 1000-dimensional vector, the 1000-dimensional vector is input into the regularization technique dropout, and part of the neuron output is randomly set to zero to obtain a regularized vector, and then the regularized vector is input into the Softmax activation function, finally the output of the model is converted into a probability distribution, thereby obtaining the final prediction result and further discriminating the defect type;

[0039] Step 4, the GPR image to be identified is scaled to 224x224 size and input into the trained ResKANet discrimination model to realize identification of the defect category of the ground structure ground penetrating radar image.

[0040] Embodiment 1

[0041] The embodiment is based on the KAN automatic discrimination method of the ground structure ground penetrating radar image, and the specific steps are as follows:

[0042] Step 1, using ground penetrating radar to obtain ground penetrating radar GPR images of underground structures containing different defect types; each GPR image is labeled with a defect type, and the defect type is divided into a structure in good condition, a crack, groundwater accumulation leakage, and structural abnormal deformation;

[0043] Step 2, the data set after labeling is processed by data enhancement, and finally the enhanced data set is randomly divided into a training set and a test set;

[0044] The data enhancement process is as follows: the labeled GPR image is scaled to four different sizes, which are 256x256, 288x288, 320x320 and 352x352 pixels; for each size of picture, a square region with a size greater than 224x224 and less than the actual size of the operated picture is used to extract the left, middle and right or upper, middle and lower parts of the operated picture, and then a 224x224 square is extracted in the four corners and center of each extracted square region, and each extracted square region is scaled to a 224x224 square; finally, all the obtained 224x224 pictures are mirror processed, that is, flipped left and right; each labeled GPR image obtains 144 cropped regions;

[0045] Step 3, constructing a ResKANet discrimination model

[0046] The enhanced GPR image is used as the input of the ResKANet discrimination model, and the GPR image is used as the input of the ResKANet discrimination model. First, the GPR image is input into a convolution layer with a 3*3 convolution kernel, and the output channel number of the convolution layer is 64, to obtain a feature image with a height of H, a width of W and a channel number of 64. The use of a stacked ResNeXt network with a 3*3 convolution kernel can effectively reduce the number of parameters used, and because the output of each position only depends on a 3x3 region, not a 7x7 region, the calculation amount is reduced under the premise of ensuring accuracy.

[0047] The H*W*64 feature image is input to a 3*3 pooling layer, and after the pooling operation, a H / 2*W / 2*64 feature image is obtained. The H / 2*W / 2*64 feature image is input into four convolution groups in turn, and the number of channels of the four convolution groups increases in turn, each of which is composed of two 1*1 convolution kernels and one 3*3 convolution kernel. Each convolution group has different repetition times, the first kind of convolution group repeats 3 times, the second kind of convolution group repeats 4 times, the third kind of convolution group repeats 6 times, and the fourth kind of convolution group repeats 3 times. The H / 2*W / 2*64 feature image is input into the first kind of convolution group, first passes through a 1x1 convolution layer for dimension increasing processing, and then passes through a grouped convolution for processing. The convolution kernel of the grouped convolution is 3x3, the group number is 32, and the output channel number is also equal to 128. Then, the 1x1 convolution is used for dimension increasing, and the channel number is increased from 128 to 256 to obtain a new image feature F(x). Finally, the output F(x) of the new image feature is added to the input x to obtain the final output H(x). The input of the next layer is:

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

[0049] H(x) is input into the first kind of convolution group for the second time, and the convolution operation is repeated again. Different from the first convolution operation, the first 1*1 convolution group in the second and third convolutions is a dimension reduction operation, which reduces the channel number from 256 to 128, and then performs grouped convolution with a group number of 32 and 1*1 dimension increasing, until the first kind of convolution group is repeated three times, to obtain a feature image G1(x) with a dimension of H / 2*W / 2*256.

[0050] The feature image G1(x) is input into the first KAN network, and in the first KAN network, the feature image is decomposed into 256 groups, each group 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 through a sigmoid function, and the calculation method is as follows:

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

[0052] The external single variable function External(x) is a smooth curve formed by connecting a group of known data points through piecewise linear approximation, and the calculation method is as follows:

[0053]

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

[0055]

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

[0057] The output of the first KAN network is taken 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 taken as the input of the third convolution group, and the convolution operation of the third convolution group is performed again. After repeating the convolution processing 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 taken as the input of the fourth convolution group, and the convolution operation of the fourth convolution group is performed again. After repeating the convolution processing 3 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 the processing of the convolution group and the KAN network, the feature image with a size of H / 2*W / 2*2048 is taken as the input of the global average pooling. For each channel c, the calculation formula of the global average pooling is as follows:

[0061]

[0062] where H / 2 represents the height of the feature map, W / 2 represents the width of the feature map, Xc(i,j) represents the pixel value of the cth channel located at the ith row and jth column, and finally, an average value is obtained for each channel through the average pooling layer, obtaining a 2048-dimensional feature vector. The vector is input into a fully connected layer with 1000 units to obtain a 1000-dimensional vector, and the calculation method is as follows:

[0063] Y = WX + b

[0064] Wherein Y is a 1000-dimensional vector; W is the weight matrix of the full connection layer, which contains the connection strength between each unit and the input feature; b is the bias vector, which is used to adjust the output of each unit. The obtained vector Y is input to the regularization technique dropout, and a mask vector M is randomly generated, and the elements of the mask vector are multiplied with the corresponding elements of the input vector Y, so that the random zeroing of part of the neuron output is realized. The calculation method is as follows:

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

[0066] Wherein Y is the output vector of the full connection layer, M is the mask vector, and the vector dropout(Y) after regularization processing is input to the Softmax activation function, and the output is converted into a probability distribution to obtain the defect type, and the calculation formula is as follows:

[0067]

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

[0069] The enhanced training set is read into the ResKANet discrimination model, and all images are iteratively trained in turn, and after 40000 iterations, the training model converges, that is, when the model training gradient is close to 0 (less than 0.01 can be considered as 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 adjusted. Weight parameters. Model calibration stage: 1) input the test set data to obtain the discrimination result; 2) compare the actual parameters of the test discrimination target with the preliminary discrimination result to obtain the calibrated ResKANet discrimination model; when predicting, the model first loads the trained parameters, loads the input image, and discriminates the category of each image through the trained ResKANet discrimination model, so as to realize the discrimination of underground structure defects.

[0070] The GPR image to be discriminated is scaled to 224x224 size and input into the trained ResKANet discrimination model, and the types of GPR images are automatically discriminated, such as cracks in underground structure, groundwater aggregation leakage and abnormal deformation of structure.

[0071] The ResKANet discrimination model of the application greatly improves the discrimination rate of GPR image, so that the system can realize high-speed and high-precision discrimination of GPR image.

[0072] The enhanced pictures are allocated 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 performance of the model under different parameter settings, so as to select the best parameter setting, and also to enable the model to have better generalization capability. In the training process, the k-fold cross-validation method 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 model variance and bias.

[0073] The method of the application 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 application and the existing model is 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 of different models

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

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

[0077]

[0078] Wherein, params represents the number of parameters, Accuracy represents the accuracy, FLOPs represents the floating point operation, and ResNeXt-KAN structure refers to not setting KAN network after each convolution group in Figure 2 The last fully connected layer is directly replaced with KAN network. The Res model in the above table is based on Resnet50.

[0079] Compared with the existing common automatic discrimination algorithm, the application significantly reduces the number of parameters under the premise of improving the accuracy, so that the accuracy and the recognition speed are both achieved. In addition, the recognition accuracy of the structural abnormal deformation of the method of the application is significantly improved, and when the overall accuracy reaches more than 90%, the accuracy of the structural abnormal deformation is improved to more than 94%.

[0080] Embodiment 2

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

[0082] Ground penetrating radar, used for acquiring ground penetrating radar GPR images of underground structures in the detection process;

[0083] An image processing module is configured to label the GPR image of the image acquisition module with a corresponding category and perform data enhancement processing to obtain a data set, which is divided into a training set and a test set;

[0084] The GPR image includes four cases of intact structure, crack, leakage, and abnormal structural deformation.

[0085] The data enhancement processing process is as follows: the labeled GPR image is scaled to four different sizes, i.e., 256x256, 288x288, 320x320, and 352x352 pixels; for each size of picture, a square region with a size greater than 224x224 and less than the actual size of the operated picture is used to extract the left, middle, and right or upper, middle, and lower parts of the operated picture, and then a 224x224 square is extracted at the four corners and center of each extracted square region, and each extracted square region is scaled to a 224x224 square; finally, all the obtained 224x224 pictures are mirror processed, i.e., flipped left and right; a total of 144 cropped regions are obtained for each labeled GPR image.

[0086] The ResKANet discriminant model is used to realize the discrimination of the GPR image; the ground penetrating radar and the image processing module are connected, the ResKANet discriminant model applied to the discrimination of the GPR image of the underground structure is trained by using the obtained data set, the image defect category discrimination is obtained by using the trained ResKANet discriminant model, and the discrimination result is output.

[0087] The display warning module obtains the output result of the ResKANet discriminant model, and if the underground structure has cracks, groundwater accumulation leakage, and abnormal structural deformation, the discrimination system issues a corresponding alarm to remind the staff; if there is no related problem, the structure is displayed as intact.

[0088] Example 3

[0089] The ground penetrating radar obtains the GPR image of the underground structure.

[0090] The GPR image of the underground structure to be detected is input into the trained ResKANet discriminant model.

[0091] If the structure is identified as intact, the underground structure is displayed as intact.

[0092] If the structure is identified as having cracks, the underground structure is displayed as having cracks.

[0093] If the structure is identified as having groundwater accumulation leakage, the underground structure is displayed as having groundwater and having a cavity problem.

[0094] If the structure is identified as having deformation, the underground structure is displayed as deformed.

[0095] Embodiment 4

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

[0097] Processor: As the core component of the present application, the processor is responsible for controlling and managing the operation of the entire system, including data acquisition, data processing, image discrimination, etc., and needs to have sufficient computing power and parallel processing capability to meet the real-time requirements. The processor can take different forms such as single-chip microcomputer, microprocessor, computer, etc. to meet the needs of different application scenarios;

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

[0099] Database: The database is used to store and manage the collected data, historical data and analysis results, etc.

[0100] Network interface: for data exchange and communication, with high speed, high stability and high security, etc. to ensure the reliability and security of data transmission.

[0101] The processor is configured to execute computer executable instructions, which when executed by the processor, implement each step of the KAN-based automatic discrimination method for ground penetrating radar images of underground structures.

[0102] The memory stores a computer program, which can be executed by the processor to implement each step of the KAN-based automatic discrimination method for ground penetrating radar images of underground structures.

[0103] The database is configured to store and manage the data of the computer application program, including various data types and structures, which are applied to each step of the KAN-based automatic discrimination method for ground penetrating radar images of underground structures.

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

[0105] The present application is mainly applied to automatic discrimination of ground penetrating radar images of underground structures, which uses an automatic discrimination system to automatically identify GPR images and determine the condition of underground structures.

[0106] The present application aims to solve the problem of low accuracy of current underground structure GPR image identification speed, identification of defects such as cracks, leakage, and abnormal deformation of structure (overall three things are recognized, ground penetrating radar, accuracy is more than 90%), especially the low accuracy of structure abnormal deformation identification. In the ResKANet identification model of the present application, the Kolmogorov-Arnold network (KAN network) is reasonably embedded to automatically identify the type of GPR image of underground structure. For the GPR image of underground structure, the system can quickly identify the GPR image of underground structure, judge whether there are cracks, underground water and deformation of 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 GPR image identification efficiency of underground structure, and provides the possibility of faster GPR image identification under the premise of pursuing high precision. This has great significance for the GPR image identification of underground structure and has broad application prospects.

[0107] In the description of the present application, the specific features, structures, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0108] It should be understood that parts of the present application can be realized by hardware, software or a combination thereof. In the above embodiments, a plurality of steps or methods can be realized by software or hardware stored in a memory and executed by a suitable instruction execution system.

[0109] Those skilled in the art in the technical field can understand that all or part of the steps carried out by the above-mentioned embodiment method can be completed by a program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0110] The present application is not described and is applicable to the prior art.

Claims

1. A KAN-based automatic discrimination method for ground penetrating radar images of underground structures, characterized by, The steps of the method are: Step 1, obtaining ground penetrating radar (GPR) images of underground structures containing different defect types by using ground penetrating radar; labeling each GPR image with a defect type, and the defect type is classified as structure intact, crack, groundwater accumulation leakage, and abnormal structure deformation; Step 2, performing data augmentation processing on the labeled data set, and finally randomly dividing the augmented data set into a training set and a test set; The data augmentation processing process is: scaling the labeled GPR image to four different sizes, 256x256, 288x288, 320x320, and 352x352 pixels; for each size of the picture, a square region with a size greater than 224x224 and less than the actual size of the operated picture is used to extract the left, middle and right or upper, middle and lower parts of the operated picture, and then a 224x224 square is extracted in the four corners and center of each extracted square region, and each extracted square region is scaled to a 224x224 square; finally, mirror processing is performed on all the obtained 224x224 pictures, that is, left and right flipping; a total of 144 cropped regions are obtained for each labeled GPR image; Step 3, constructing a ResKANet discriminant model The obtained data set is used to train a ResKANet discriminant model applied to GPR image discrimination of underground structures, and the trained ResKANet discriminant model is used to obtain 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 channel numbers connected to the KAN network, an average pooling layer, a full connection layer, a regularization layer, and a Softmax activation function; KAN network is set after each convolution group, and the feature image after each convolution group is input into the KAN network and then into the next convolution group; the convolution of the four convolution groups is input into the average pooling layer to obtain a more representative feature vector, the vector is input into the full connection layer with 1000 units to obtain a 1000-dimensional vector, the obtained 1000-dimensional vector is input into the regularization technology dropout, and part of the neuron output is randomly set to zero to obtain a vector after regularization processing, then the vector after regularization processing is input into the Softmax activation function, finally the output of the model is converted into a probability distribution, thereby obtaining the final prediction result and further discriminating the defect type; Step 4, scaling the GPR image to be discriminated to 224x224 size and inputting it into the trained ResKANet discriminant model to realize recognition of the defect category of the ground penetrating radar image of the underground structure.

2. The method of claim 1, wherein, The low channel number 1*1 convolution kernel, the low channel number 3*3 convolution kernel and the high channel number 1*1 of the four convolution groups are connected in turn, the repetition number and the channel number 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.

3. The method of claim 2, wherein, The accuracy of the ground penetrating radar image is over 90%, and the floating point is 3.4x10 11 -6; and the accuracy of the structural abnormal deformation is over 90%.

4. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor implements the steps of the method of claim 1 or 2 when executing the computer program.

5. A KAN-based automatic discrimination system for ground penetrating radar images of underground structures, characterized by, The system comprises: a ground penetrating radar for obtaining a ground penetrating radar (GPR) image of an underground structure during detection; an image processing module for labeling the GPR image of the image acquisition module with a corresponding category and performing data enhancement processing to obtain a data set and divide it into a training set and a test set; the GPR image includes four cases of intact structure, cracks, leakage and abnormal structural deformation; the data enhancement processing process is: scaling the labeled GPR image to four different sizes, 256x256, 288x288, 320x320 and 352x352 pixels; for each size of picture, a square region with a size greater than 224x224 and less than the actual size of the operated picture is used to extract the left, middle and right or upper, middle and lower parts of the operated picture, and then a 224x224 square is extracted at the four corners and center of each extracted square region, and then each extracted square region is scaled to a 224x224 square; finally, mirror processing is performed on all the obtained 224x224 pictures, that is, left and right flipping; a total of 144 cropped regions are obtained for each labeled GPR image; a ResKANet discriminant model for realizing discrimination of the GPR image; connected with the ground penetrating radar and the image processing module, the ResKANet discriminant model is trained using the obtained data set to be applied to discrimination of the GPR image of the underground structure, and the trained ResKANet discriminant model is used to obtain image defect category discrimination and output a discrimination result; a display warning module that obtains the output result of the ResKANet discriminant model, and if there are cracks, groundwater accumulation leakage and abnormal structural deformation problems in the underground structure, the discrimination system issues a corresponding alarm to remind the staff; if there are no related problems, the structure is intact. The ResKANet discriminant model comprises a convolution layer composed of three 3*3 convolution kernels connected in sequence, a 3*3 pooling layer, four convolution groups with different channel numbers connected with a KAN network, an average pooling layer, a full connection layer, a regularization layer and a Softmax activation function; each convolution group is provided with a KAN network, and the feature image after the convolution of each convolution group is input into the KAN network and then into the next convolution group; the convolution of the four convolution groups is input into the average pooling layer to obtain a more representative feature vector, the vector is input into the full connection layer with 1000 units to obtain a 1000-dimensional vector, the obtained 1000-dimensional vector is input into the regularization technology dropout, part of the neuron outputs is randomly set to zero to obtain a vector after regularization processing, the vector after regularization processing is input into the Softmax activation function, finally the output of the model is converted into a probability distribution, so as to obtain the final prediction result and further discriminate the defect type.

Images