A deep learning road surface segregation identification method based on ground penetrating radar

Abstract

The application discloses a kind of deep learning pavement segregation identification methods based on ground penetrating radar, it belongs to intelligent transportation field.The application solves the problem of poor identification effect of existing pavement segregation identification methods based on deep learning network.The method of the application is:1, obtaining the ground penetrating radar waveform data of the pavement with known segregation condition, then calculating radar wave amplitude, radar wave phase and phase gradient;2, calculating the segregation area attention score of the obtained ground penetrating radar waveform data;3, the radar wave amplitude, radar wave phase and phase gradient obtained are spliced and used as the input of the constructed deep learning network, and the constructed deep learning network is trained by fusing the segregation area attention score;4, obtaining the radar wave amplitude, radar wave phase, phase gradient and segregation area attention score of the ground penetrating radar waveform data of the pavement to be detected;then input the trained deep learning network to obtain the segregation identification result.The application can be applied to the field of intelligent transportation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent transportation, and particularly relates to a deep learning pavement segregation identification method based on ground penetrating radar. BACKGROUND

[0002] As a main inducement of early pavement diseases, segregation is easy to cause early pavement diseases such as rut, pothole and crack, and further shorten the service life of pavement, increase the road maintenance cost and affect the driving comfort. Therefore, how to quickly detect and locate the segregation in the whole space of pavement has become the current research focus.

[0003] At present, the pavement segregation detection problem can be divided into two categories: destructive testing and non-destructive testing. The destructive testing includes core sampling, but it has the disadvantages of damaging the pavement and low efficiency. For the non-destructive testing method, it includes automatic detection of segregation by using nuclear density meter, non-nuclear density meter, laser instrument and digital camera image, but these methods can only sample in a small range or can only detect the segregation on the road surface, and cannot quickly detect the whole space of the road. Ground penetrating radar provides a detection means under the road surface, but the current segregation detection based on ground penetrating radar is mainly carried out by manual visual observation of ground penetrating radar images, which consumes manpower, has low efficiency and strong subjectivity. Although a small amount of research directly uses ground penetrating radar images as the input of deep learning network to identify underground targets, the process of converting ground penetrating radar data into ground penetrating radar images causes the loss of radar waveform information, and due to the uncertain shape, size and pattern of segregation, the effect of these deep learning methods in segregation identification is poor.

[0004] In summary, the existing pavement segregation identification method based on deep learning network has poor recognition effect, and there is currently a lack of efficient and stable automatic detection means for segregation in the whole space of pavement. SUMMARY

[0005] The purpose of the present application is to solve the problem of poor recognition effect of the existing pavement segregation identification method based on deep learning network, and a deep learning pavement segregation identification method based on ground penetrating radar is proposed.

[0006] The technical scheme adopted by the present application to solve the above technical problems is:

[0007] A deep learning pavement segregation identification method based on ground penetrating radar, which specifically comprises the following steps:

[0008] Step one, obtaining ground penetrating radar waveform data of a pavement with known segregation conditions, and then performing amplitude recovery processing on the obtained ground penetrating radar waveform data to obtain radar signals after amplitude recovery;

[0009] The radar signal after amplitude recovery is decomposed in amplitude and phase to obtain radar wave amplitude and radar wave phase; and the gradient of the radar wave phase in the depth direction is taken as the phase gradient;

[0010] Step two, calculating the dissociation area attention score of the ground penetrating radar waveform data obtained in step one at each sampling point;

[0011] The specific process of step two is:

[0012] S(x, y) = [nor(E(x, y)) nor(V(x, y))] · W

[0013]

[0014]

[0015] Wherein, S(x, y) is the dissociation area attention score at the sampling point (x, y), E(x, y) is the information entropy at the sampling point (x, y), V(x, y) is the coefficient of variation at the sampling point (x, y), nor(·) is the maximum minimum normalization, W is the weight vector of the weighted sum operation, is a set of phase gradient data of neighbor points of the sampling point (x, y), N is the discretization level of the phase gradient data, is The probability that the data in falls into the i-th discrete level, i = 1, 2, …, N, std(·) is the standard deviation operator, and mean(·) is the mean operator.

[0016] Step three, constructing a deep learning network, splicing the radar wave amplitude, radar wave phase and phase gradient obtained in step one as the input of the constructed deep learning network, and training the constructed deep learning network by fusing the dissociation area attention score of each sampling point;

[0017] Step four, collecting ground penetrating radar waveform data of the road to be detected, and then processing the collected data by the method of step one to obtain the radar wave amplitude, radar wave phase and phase gradient of the ground penetrating radar waveform data of the road to be detected, and splicing the radar wave amplitude, radar wave phase and phase gradient to obtain the spliced data;

[0018] Calculating the dissociation area attention score of the ground penetrating radar waveform data of the road to be detected at each sampling point;

[0019] The spliced data and the dissociation area attention score are input into the trained deep learning network, and the dissociation recognition result of the road to be detected is output by the trained deep learning network.

[0020] The beneficial effects of the present application are:

[0021] The present application provides a ground penetrating radar based deep learning pavement segregation identification method, based on ground penetrating radar data, the present application uses signal processing technology to enhance the radar signal, and designs a segregation area attention score, and combines a deep learning network to realize automatic detection of segregation in the full space range of the pavement, compared with the previous ground penetrating radar data identification method using artificial naked eye observation, the present application solves the problems of low efficiency and strong subjectivity of artificial identification, at the same time, compared with the segregation identification technology based on nuclear density meter, non-nuclear density meter, laser instrument and digital camera image, the present application solves the problem that these sensors cannot detect segregation under the road surface, realizes full space identification and positioning of pavement segregation, realizes detection of segregation size and shape, in addition, the deep learning network designed by the present application takes radar wave amplitude, radar wave phase and phase gradient as input, and combines the characteristics of waveform disorder in the segregation area, uses the segregation area attention score to improve the performance of the deep learning network in segregation identification, and finally realizes accurate detection and positioning of segregation, and improves the segregation identification effect. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a flowchart of a ground penetrating radar based deep learning pavement segregation identification method of the present application;

[0023] Fig. 2(a) is radar waveform original data;

[0024] Fig. 2(b) is radar data after amplitude recovery;

[0025] Fig. 2(c) is amplitude data;

[0026] Fig. 2(d) is phase data;

[0027] Fig. 2(e) is phase gradient data;

[0028] Figure 3 is an example diagram of the segregation area attention score of the present application;

[0029] Figure 4 is a structure diagram of the deep learning network constructed by the present application;

[0030] Fig. 5(a) is a pavement structure schematic diagram of Xieruan line;

[0031] Fig. 5(b) is a pavement structure schematic diagram of Longshuang road;

[0032] Figure 6 is a segregation identification result diagram of the ground penetrating radar based deep learning pavement segregation identification method of the present application. DETAILED DESCRIPTION

[0033] Specifically, the method comprises the following steps:

[0034] Step one, obtaining the ground penetrating radar waveform data of the road surface with known segregation, and then performing amplitude recovery processing on the obtained ground penetrating radar waveform data to obtain radar signals after amplitude recovery;

[0035] The radar signals after amplitude recovery are decomposed in amplitude and phase to obtain radar wave amplitude and radar wave phase; and the gradient of the radar wave phase in the depth direction is taken as the phase gradient;

[0036] Step two, calculating the segregation area attention score of the ground penetrating radar waveform data obtained in step one at each sampling point;

[0037] The specific process of step two is as follows:

[0038] S(x,y)=[nor(E(x,y))nor(V(x,y))]·W

[0039]

[0040]

[0041] wherein S(x,y) is the segregation area attention score at the sampling point (x,y), E(x,y) is the information entropy at the sampling point (x,y), V(x,y) is the coefficient of variation at the sampling point (x,y), nor(·) is the maximum and minimum normalization, W is the weight vector of the weighted sum operation, is a set of phase gradient data of neighbor points of the sampling point (x,y), N is the discretization level of the phase gradient data, is the probability that the data in the set is M, wherein the number of data falling into the i-th discrete level is M i then i=1,2,…,N, std(·) is the standard deviation operator, and mean(·) is the mean operator;

[0042] The neighbor point set may be selected according to actual conditions, and in the present application, the phase gradient data of the sampling points in the rectangular window range from (x-5, y-1) to (x+5, y+1) with the sampling point (x,y) as the center is taken as the neighbor point set

[0043] Step three, constructing a deep learning network, splicing the radar wave amplitude, radar wave phase and phase gradient obtained in step one as the input of the constructed deep learning network, and training the constructed deep learning network by fusing the segregation region attention scores of each sampling point;

[0044] Step four, collecting ground penetrating radar waveform data of the road to be detected, and then processing the collected data by the method of step one to obtain radar wave amplitude, radar wave phase and phase gradient of the ground penetrating radar waveform data of the road to be detected, and splicing the radar wave amplitude, radar wave phase and phase gradient to obtain the spliced data;

[0045] Calculating the segregation region attention scores of the ground penetrating radar waveform data of the road to be detected at each sampling point;

[0046] Inputting the spliced data and the segregation region attention scores into the trained deep learning network, and outputting the segregation identification result of the road to be detected by the trained deep learning network.

[0047] Specific implementation method two, different from the specific implementation method one, the amplitude restored radar signal is:

[0048] I=I0r 2 e αd

[0049] Wherein, I represents the amplitude restored radar signal, I0 represents the obtained ground penetrating radar waveform data, r is the one-way distance of radar wave propagation, e is the base of natural logarithm, a is the absorption coefficient of road material to electromagnetic wave, and d is the propagation distance of electromagnetic wave in road material.

[0050] The amplitude restoration processing is performed according to the characteristics of energy attenuation in the process of electromagnetic wave propagation and the absorption characteristics of the propagation medium to electromagnetic wave, so as to enhance the radar wave data of the deep layer of the road.

[0051] The other steps and parameters are the same as those of the specific implementation method one.

[0052] Specific implementation method three, different from the specific implementation method one or two, the radar wave amplitude is:

[0053]

[0054] Wherein, real(·) is the real part operator, imag(·) is the imaginary part operator, and A is the radar wave amplitude.

[0055] The other steps and parameters are the same as those of the specific implementation method one or two.

[0056] Specific embodiment four, different from one of the specific embodiments one to three, the radar wave phase is:

[0057]

[0058] Wherein, arctan(·) is the inverse tangent function, and φ is the radar wave phase.

[0059] Other steps and parameters are the same as one of the specific embodiments one to three.

[0060] Specific embodiment five, different from one of the specific embodiments one to four, the phase gradient is:

[0061] G(x,y)=g(φ(x,y+1)-φ(x,y))

[0062] Wherein, G(x,y) is the phase gradient at the sampling point (x,y), x is the sampling point coordinate of the radar in the horizontal direction, y is the sampling point coordinate of the radar in the depth direction, φ(x,y+1) represents the radar wave phase at the sampling point (x,y+1), φ(x,y) represents the radar wave phase at the sampling point (x,y), and g(·) represents normalizing the variable to the interval [0,2π), that is, Is an integer.

[0063] Other steps and parameters are the same as one of the specific embodiments one to four.

[0064] Specific embodiment six, different from one of the specific embodiments one to five, the structure of the deep learning network constructed in step three specifically includes: first convolutional layer to sixteenth convolutional layer, first average pooling layer to eighth average pooling layer, first batch normalization layer to fifth batch normalization layer, maximum pooling layer, first deconvolutional layer to third deconvolutional layer and softmax classification layer.

[0065] Other steps and parameters are the same as one of the specific embodiments one to five.

[0066] Specific embodiment seven, different from one of the specific embodiments one to six, the working process of the deep learning network is:

[0067] The spliced data corresponding to the road surface to be detected and the dissociation region attention score are taken as the input of the deep learning network, and the spliced data sequentially passes through the first convolutional layer, the second convolutional layer, the first batch normalization layer and the first average pooling layer; the dissociation region attention score passes through the second average pooling layer;

[0068] The output of the first average pooling layer and the output of the second average pooling layer are fused in a channel-by-channel bitwise multiplication manner to obtain a fusion result A;

[0069] The fusion result A sequentially passes through a third convolutional layer, a fourth convolutional layer, a second batch normalization layer and a third average pooling layer; the output of the second average pooling layer is fused with the output of the third average pooling layer and the output of the fourth average pooling layer in a manner of channel-by-channel bitwise multiplication after passing through a fourth average pooling layer, to obtain a fusion result B;

[0070] The fusion result B sequentially passes through a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, a third batch normalization layer and a fifth average pooling layer; the output of the fourth average pooling layer is fused with the output of the fifth average pooling layer and the output of the sixth average pooling layer in a manner of channel-by-channel bitwise multiplication after passing through a sixth average pooling layer, to obtain a fusion result C;

[0071] The fusion result C sequentially passes through an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, a fourth batch normalization layer and a seventh average pooling layer; the output of the sixth average pooling layer is fused with the output of the seventh average pooling layer and the output of the eighth average pooling layer in a manner of channel-by-channel bitwise multiplication after passing through an eighth average pooling layer, to obtain a fusion result D;

[0072] The fusion result D sequentially passes through an eleventh convolutional layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fifth batch normalization layer, a max pooling layer, a fourteenth convolutional layer, a fifteenth convolutional layer and a sixteenth convolutional layer, and the output of the sixteenth convolutional layer is taken as the input of a first deconvolutional layer;

[0073] The output of the first deconvolutional layer is bitwise added with the output of the seventh average pooling layer, to obtain a bitwise addition result A';

[0074] The bitwise addition result A' is taken as the input of a second deconvolutional layer, and the output of the second deconvolutional layer is bitwise added with the output of the fifth average pooling layer, to obtain a bitwise addition result B';

[0075] The bitwise addition result B' is taken as the input of a third deconvolutional layer, and the output of the third deconvolutional layer passes through a softmax classification layer, and the softmax classification layer outputs a segregation recognition result of the road surface to be detected.

[0076] The other steps and parameters are the same as one of the first to sixth embodiments.

[0077] The eighth embodiment is different from one of the first to seventh embodiments in that the loss function of the deep learning network is:

[0078]

[0079] Wherein, L is the loss function; l x,y is the segregation label of the radar sampling point (x, y), if the radar sampling point (x, y) is segregation, then l x,y= 1, otherwise l x,y = 0; p x,y is the probability that radar sample point (x, y) is identified as segregation; n is the number of sample points in the horizontal direction; m is the number of sample points in the depth direction; log(·) is the logarithm operation with e as the base.

[0080] The other steps and parameters are the same as one of embodiments one to seven.

[0081] Embodiment nine, different from one of embodiments one to eight, is that the segregation identification result of the road surface to be detected is output by the softmax classification layer, and the specific process is as follows:

[0082] The softmax classification layer outputs the probability that each radar sample point belongs to segregation, and the sample point with a segregation probability greater than 0.5 is identified as segregation, and the sample point with a segregation probability less than or equal to 0.5 is identified as no segregation.

[0083] The other steps and parameters are the same as one of embodiments one to eight.

[0084] The method of the present application will be described in detail below in conjunction with the accompanying drawings:

[0085] As shown in Figure 1 , the method of the present application is a deep learning road surface segregation identification method based on ground penetrating radar, and the specific steps of the method are as follows:

[0086] 1. Radar waveform data processing

[0087] As shown in Fig. 2(a), the radar waveform raw data is processed by the present application to obtain radar wave amplitude, phase and phase gradient data as the input of the deep learning network, which includes the following steps:

[0088] 1) According to the characteristics that the energy gradually attenuates with the propagation distance in the electromagnetic wave propagation process, specifically, the electromagnetic wave energy is inversely proportional to the square of the propagation distance, and the electromagnetic wave is affected by the absorption of the propagation medium material, the radar signal is processed by amplitude restoration to enhance the radar data at the deep position of the road surface. The radar data after amplitude restoration is shown in Fig. 2(b), and specifically, the radar signal after amplitude restoration can be represented as I:

[0089] I = I0r 2 e αd

[0090] Where I0 represents the original radar data, r is the one-way distance of radar wave propagation, a is the absorption coefficient of road surface material to electromagnetic wave, and d is the propagation distance of electromagnetic wave in road surface material.

[0091] 2) Perform amplitude and phase decomposition on the recovered radar signal. Figure 2(c) shows the amplitude data and Figure 2(d) shows the phase data. Let the amplitude be represented by A and the phase by φ. Then the expressions for amplitude and phase can be written as:

[0092]

[0093]

[0094] Where real(·) is the real part operator, imag(·) is the imaginary part operator, and arctan(·) is the arctangent function.

[0095] 3) Take the gradient of the phase in the depth direction as the phase gradient. Figure 2(e) shows the phase gradient data. The phase gradient is represented as G:

[0096] G(x,y)=g(φ(x,y+1)-φ(x,y))

[0097] Where x represents the coordinates of the radar sampling point in the horizontal direction, y represents the coordinates of the radar sampling point in the depth direction, and the g(·) transformation normalizes the variables to the interval [0, 2π).

[0098] 4) The amplitude, phase, and phase gradient are spliced ​​together as input to the deep learning network. The spliced ​​data can be represented as [A φ G];

[0099] 2. Attention score extraction from the separated regions

[0100] like Figure 3 The image shows the attention score map for the segregated region. Based on the chaotic and disordered characteristics of the radar waveform in the segregated region, the attention score is designed and fused with the feature map extracted by the deep learning network through a channel-by-channel bitwise multiplication. This improves the deep learning network's focus on the features of the chaotic waveform region. Specifically, the attention score is obtained by calculating the weighted sum of the information entropy and coefficient of variation of the phase gradient within the region, and is represented as S:

[0101] S = [nor(E) nor(V)]·W

[0102]

[0103]

[0104] Where E is the information entropy, V is the coefficient of variation, nor(·) is the max-min normalization, and W is the weight vector of the weighted sum operation. N is the set of phase gradient data of the neighbor points of the point (x, y), and N is the discretization level of the phase gradient data, is the probability that the data in falls into the i-th discrete level, std(·) is the standard deviation operator, and mean(·) is the mean operator.

[0105] 3. Deep learning network construction and training

[0106] The deep learning network is constructed and trained, specifically including the following steps:

[0107] 1) Network structure

[0108] As shown in Figure 4 , the structure of the deep learning network described in the present application is introduced, taking radar wave amplitude, phase and phase gradient as input; batch normalization is added after the convolution layer in the down-sampling network; in addition, an isolated region attention score is added, and the isolated region attention score is multiplied with the feature map extracted in the down-sampling process channel by channel and bit by bit to improve the attention of the network to the waveform chaotic region; finally, the softmax layer is used to output the probability that each radar sampling point belongs to isolation, and the sampling points with an isolation probability greater than 0.5 are marked as isolated, otherwise they are marked as non-isolated, and thus the isolation identification result is obtained.

[0109] 2) Loss function

[0110] The deep learning network calculates an isolation probability for each radar sampling point, and uses a threshold of 0.5 to divide the radar sampling points into two categories of isolated and non-isolated. In the training process, the loss function L of the deep learning network is defined as:

[0111]

[0112] Wherein:

[0113] l x,y is the isolation label of the radar sampling point (x, y). If the point is isolated, l x,y = 1, otherwise l x,y = 0;

[0114] p x,y is the probability that the radar sampling point (x, y) is identified as isolated;

[0115] n is the number of sampling points in the horizontal direction;

[0116] m is the number of sampling points in the depth direction;

[0117] log(·) is the logarithm operation with base e.

[0118] Embodiment

[0119] To verify the effectiveness of the present application, data was collected on two newly built roads using three-dimensional ground penetrating radar, and the dataset was labeled. The deep learning network was trained and tested to verify the effectiveness of the segregation identification method. The specific experimental settings are as follows:

[0120] 1) Ground penetrating radar data

[0121] (1) The ground penetrating radar data comes from the West Outer Ring Line and Longshuang Road in Changchun City, Jilin Province. The West Outer Ring Line pavement includes a 5 cm thick asphalt layer and a 25 cm thick cement stabilized gravel layer. The Longshuang Road pavement includes a 7.5 cm thick asphalt layer and a 25 cm thick cement stabilized gravel layer. Figures 5(a) and 5(b) show the pavement structure schematic diagram of the two roads;

[0122] (2) The data collection work was carried out using a step frequency ground penetrating radar. The radar host model is GeoScope MK IV, the antenna array model is DXG1820, the radar frequency bandwidth is 100-3000 MHz, and a 25 ns scan window is used for data collection;

[0123] (3) The collected radar data was cropped into B-scan slices with a size of 256x128, and 200 slices with serious pavement segregation were selected for labeling. The B-scan slices were divided into training set and test set according to the ratio of 7:3;

[0124] 2) Segregation identification performance analysis

[0125] The neural network was trained and tested using the dataset. During the training process, the learning rate was set to 0.001, the Adam optimizer was used, and the existing target recognition deep learning network (SegNet, U-Net and DeepLab v3+) was compared. The recognition results were evaluated using class pixel accuracy (CPA) and intersection over union (IoU). The calculation formulas of CPA and IoU can be written as:

[0126]

[0127]

[0128] where P TP , P FP and P FN are the number of true pixels, false positive pixels and false negative pixels, respectively;

[0129] The class pixel accuracy and intersection over union results of segregation identification are shown in Table 1. At the same time Figure 6 shows some segregation identification results;

[0130] Table 1 dissociation recognition result statistics table

[0131] Method CAP IoU SegNet 0.5076 0.4142 U-Net 0.6166 0.4808 DeepLab v3+ 0.7182 0.5179 Inventive method 0.7422 0.6083

[0132] As can be seen from Table 1, the method of the present application can realize accurate recognition of road surface segregation, and the pixel accuracy rate for segregation recognition can reach 0.7422, and the intersection-over-union can reach 0.6083, which is higher than other target recognition algorithms; at the same time, from the perspective of the recognition result contour of the segregation area, the method of the present application is closer to the segregation label than other algorithms, and the detection and positioning effect of the present application is better than other algorithms. Figure 6 It can also be seen that the recognition result contour of the segregation area of the present application is closer to the segregation label than other algorithms, and the detection and positioning effect of the present application is better than other algorithms.

[0133] 3) The ablation experiment of the method of the present application was carried out, the network recognition performance in the case of using radar image as input and using [A φ G] as input was compared, and the influence of whether to add the segregation area attention score S on the network recognition performance was compared, and Table 2 is the comparison result of the ablation experiment;

[0134] Table 2 comparison result of ablation experiment

[0135] Method CAP IoU Radar image 0.6562 0.4673 [A φ G] 0.6845 0.5554 [A φ G]+S 0.7422 0.6083

[0136] As can be seen from Table 2, compared with directly using radar image for segregation recognition, using processed radar data [A φ G] for segregation recognition has better segregation recognition effect, and after fusing the segregation area attention score S in the deep learning network, the recognition performance of the deep learning network for segregation will be further strengthened, which proves the effectiveness of the radar waveform data processing method and the segregation area attention score proposed in the method of the present application in segregation recognition.

[0137] Compared with the method of manually observing ground penetrating radar data for segregation recognition, the obvious advantage is that the segregation can be automatically detected, the segregation recognition efficiency is improved, and the manpower is saved; compared with the segregation recognition technology based on kernel density meter, non-kernel density meter, laser instrument and digital camera image, the obvious advantage is that the segregation hidden under the road surface can be detected, and the shape and size of the segregation in the whole road surface space can be determined; in addition, compared with other target recognition algorithms, the radar waveform data is processed in the present application to obtain input data more conducive to deep network learning, and the segregation area attention score is introduced to improve the attention of the network to the waveform chaotic area features, and the segregation recognition performance is more excellent.

[0138] The above calculation examples of the present application are only used to illustrate the calculation model and calculation process of the present application, and are not used to limit the embodiments of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art, and all the embodiments cannot be exhausted here. Any obvious changes or variations derived from the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A ground penetrating radar-based deep learning pavement segregation identification method, characterized in that, The method specifically comprises the following steps: Step one, obtaining ground penetrating radar waveform data of a road with known segregation condition, and then performing amplitude recovery processing on the obtained ground penetrating radar waveform data to obtain radar signals after amplitude recovery; Performing amplitude and phase decomposition on the radar signals after amplitude recovery to obtain radar wave amplitude and radar wave phase; and taking the gradient of the radar wave phase in the depth direction as the phase gradient; Step two, calculating segregation area attention scores of the ground penetrating radar waveform data obtained in step one at each sampling point; The specific process of step two is as follows: wherein, is a local region attention score at sampling point is an information entropy at sampling point is a coefficient of variation at sampling point is a min-max normalization, is a weight vector for the weighted sum operation, is a set of phase gradient data of neighboring points of sampling point is a discretization level of the phase gradient data, is a probability that data in falls into the discrete level, is a standard deviation operator, is a mean operator;​​​​​ Step three, constructing a deep learning network, splicing the radar wave amplitude, radar wave phase and phase gradient obtained in step one as the input of the constructed deep learning network, and training the constructed deep learning network by fusing the segregation area attention scores of each sampling point; Step four, collecting ground penetrating radar waveform data of a road to be detected, and then processing the collected data by the method of step one to obtain radar wave amplitude, radar wave phase and phase gradient of the ground penetrating radar waveform data of the road to be detected, splicing the radar wave amplitude, radar wave phase and phase gradient to obtain spliced data; Calculating segregation area attention scores of the ground penetrating radar waveform data of the road to be detected at each sampling point; Inputting the spliced data and the segregation area attention scores into the trained deep learning network, and outputting segregation identification results of the road to be detected through the trained deep learning network; The structure of the deep learning network constructed in step three specifically comprises: a first convolutional layer to a sixteenth convolutional layer, a first average pooling layer to an eighth average pooling layer, a first batch normalization layer to a fifth batch normalization layer, a maximum pooling layer, a first deconvolutional layer to a third deconvolutional layer, and a softmax classification layer; The working process of the deep learning network is as follows: Taking the spliced data and the segregation area attention scores corresponding to the road to be detected as the input of the deep learning network, the spliced data sequentially passes through the first convolutional layer, the second convolutional layer, the first batch normalization layer and the first average pooling layer; the segregation area attention scores pass through the second average pooling layer; The output of the first average pooling layer and the output of the second average pooling layer are fused in a way of multiplying each channel by bit by bit, to obtain a fusion result A; The fusion result A sequentially passes through a third convolutional layer, a fourth convolutional layer, a second batch normalization layer and a third average pooling layer; the output of the second average pooling layer passes through a fourth average pooling layer, and the output of the third average pooling layer and the output of the fourth average pooling layer are fused in a way of multiplying each channel by bit by bit, to obtain a fusion result B; The fusion result B sequentially passes through a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, a third batch normalization layer and a fifth average pooling layer; the output of the fourth average pooling layer passes through a sixth average pooling layer, and the output of the fifth average pooling layer and the output of the sixth average pooling layer are fused in a way of multiplying each channel by bit by bit, to obtain a fusion result C; The fusion result C sequentially passes through an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, a fourth batch normalization layer and a seventh average pooling layer; and an output of the sixth average pooling layer is fused with an output of the seventh average pooling layer and an output of the eighth average pooling layer in a channel-by-channel bitwise multiplication manner after the output of the sixth average pooling layer passes through the eighth average pooling layer, to obtain a fusion result D; The fusion result D sequentially passes through an eleventh convolutional layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fifth batch normalization layer, a maximum pooling layer, a fourteenth convolutional layer, a fifteenth convolutional layer and a sixteenth convolutional layer, and an output of the sixteenth convolutional layer is taken as an input of a first deconvolutional layer; An output of the first deconvolutional layer is bitwise added with an output of the seventh average pooling layer, to obtain a bitwise addition result A'; The bitwise addition result A' is taken as an input of a second deconvolutional layer, and an output of the second deconvolutional layer is bitwise added with an output of the fifth average pooling layer, to obtain a bitwise addition result B'; The bitwise addition result B' is taken as an input of a third deconvolutional layer, and an output of the third deconvolutional layer passes through a softmax classification layer again, and an segregation identification result of the road surface to be detected is output through the softmax classification layer. 2.The ground penetrating radar-based deep learning method for road surface segregation identification according to claim 1, wherein, The radar signal after amplitude recovery is: wherein, represents the radar signal after amplitude restoration, represents the acquired ground penetrating radar waveform data, is the one-way distance of radar wave propagation, is the base of the natural logarithm, is the absorption coefficient of the electromagnetic wave by the road surface material, is the propagation distance of the electromagnetic wave in the road surface material. 3.The ground penetrating radar-based deep learning method for road surface segregation identification according to claim 2, characterized in that, The radar wave amplitude is: wherein, is a real part operator, is an imaginary part operator, is a radar wave amplitude. 4.The method of claim 3, wherein, The radar wave phase is: wherein, is the arctangent function, is the radar wave phase. 5.The ground penetrating radar-based deep learning method for pavement segregation identification according to claim 4, wherein, The phase gradient is: wherein is the phase gradient at the sampling point is the sampling point coordinate of the radar in the horizontal direction, is the sampling point coordinate of the radar in the depth direction, denotes the radar wave phase at the sampling point denotes the radar wave phase at the sampling point denotes the normalization of a variable to the interval .​​​ 6. The deep learning-based road surface segregation identification method based on ground-penetrating radar according to claim 5, characterized in that, The loss function of the deep learning network is: wherein: is a loss function; is a radar sample point is a categorical label for the radar sample point is categorical, then , otherwise ; is a probability that the radar sample point is identified as categorical; is a number of sample points in the horizontal direction; is a number of sample points in the depth direction. 7.The method of claim 6, wherein, The segregation identification result of the road surface to be detected is output through the softmax classification layer, and the specific process is as follows: The softmax classification layer outputs a probability that each radar sampling point belongs to segregation, and a sampling point with a segregation probability greater than 0.5 is identified as segregation, and a sampling point with a segregation probability less than or equal to 0.5 is identified as non-segregation.

Images