A three-dimensional ground penetrating radar crack disease identification method and system

Abstract

The present invention provides a three-dimensional ground penetrating radar crack disease recognition method and system, belonging to the field of road disease detection, and the recognition method includes: obtaining a training sample set; the training sample set includes multiple road sample images and label information of each pixel in each road sample image; the label information includes 0 and 1; 0 indicates that the corresponding pixel is not a crack, and 1 indicates that the corresponding pixel is a crack; based on the spatial attention mechanism, VGG16 and Unet neural network, a crack Unet neural network is constructed; according to the training sample set, the crack Unet neural network is trained to obtain a crack recognition model; the road to be identified is scanned by a three-dimensional ground penetrating radar to obtain a road image to be identified; according to the road image to be identified, based on the crack recognition model, according to the road image to be identified, the cracks in the road to be identified are determined. By classifying each pixel in the image, cracks in the road can be accurately identified.

Description

Technical Field

[0001] The present invention relates to the field of road damage detection, and in particular to a three-dimensional ground penetrating radar crack damage identification method and system. Background Art

[0002] Under the influence of vehicle load, temperature, ultraviolet rays, etc., the road pavement structure will inevitably suffer damage as the operation time increases, such as cracks, subsidence, and interlayer delamination. Quickly, effectively, and accurately find the location, type, and severity of pavement structure diseases, and predict the development of diseases, which is of great significance for scientifically and rationally determining the maintenance time and formulating maintenance plans. Among the various types of pavement structure diseases, crack diseases are the most numerous diseases, and three-dimensional ground penetrating radar is an effective means to detect crack damage inside the pavement structure. However, in the interpretation of three-dimensional ground penetrating radar crack disease images, there is currently a lack of automated identification methods, and manual interpretation is mainly used. There are many problems in manual interpretation of radar images, such as: the interpretation process requires high professionalism and lacks sufficient data interpreters; the interpretation process is subjective, and the same radar image is interpreted by different personnel, and the interpretation results obtained are often different; the manual interpretation process is time-consuming, the workload is huge, and the interpretation efficiency is low. In addition, in the process of identifying crack diseases in single-channel three-dimensional ground penetrating radar images, there is a lack of auxiliary judgment information, which is prone to misjudgment. Moreover, radar data is an unevenly distributed three-dimensional dot matrix, which can easily cause the displayed image to be deformed and inconsistent with the actual image, thereby increasing the difficulty of recognition.

[0003] Researchers have done a lot of research on the automatic interpretation of radar images. Methods such as Canny edge detection, Hough transform, SVM (Support Vector Machine), compressed sensing, and convolutional neural networks have been proposed. Among them, Canny edge detection can achieve accurate segmentation of cracks, but since edge detection only considers local gradient changes in the image, there are a large number of false alarm signals and the recognition accuracy is low; Hough transform is an effective method for detecting and locating straight lines and analytical curves. The algorithm accuracy is less affected by random noise and signal discontinuity, but since cracks do not have fixed signal morphological characteristics, Hough transform is not suitable for automatic recognition of crack signals; SVM requires a large number of training samples for image feature selection and model parameter optimization, and is mainly used for regional classification, and cannot achieve accurate segmentation of crack disease boundaries. The classifier based on convolutional neural networks that has emerged in recent years can complete the task of identifying similar features in other unlabeled images by learning correctly labeled images. However, in the recognition of crack diseases, the accuracy and error rate indicators of this method are much lower than those of other targets, and the recognition result is a rectangular area, which does not achieve accurate segmentation of crack diseases.

[0004] Based on the above problems, a new crack disease identification method is urgently needed to improve the accuracy of crack identification. Summary of the invention

[0005] The purpose of the present invention is to provide a three-dimensional ground penetrating radar crack disease identification method and system, which can improve the accuracy of road crack identification.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A three-dimensional ground penetrating radar crack disease identification method, comprising:

[0008] Acquire a training sample set; the training sample set includes a plurality of road sample images and label information of each pixel in each road sample image; the label information includes 0 and 1; 0 indicates that the corresponding pixel is not a crack, and 1 indicates that the corresponding pixel is a crack;

[0009] Based on the spatial attention mechanism, VGG16 and Unet neural network, a crack Unet neural network is constructed;

[0010] According to the training sample set, the crack Unet neural network is trained to obtain a crack recognition model;

[0011] Scanning the road to be identified by three-dimensional ground penetrating radar to obtain an image of the road to be identified;

[0012] According to the road image to be identified, based on the crack identification model, and according to the road image to be identified, cracks in the road to be identified are determined.

[0013] Optionally, scanning the road to be identified by three-dimensional ground penetrating radar to obtain the image of the road to be identified specifically includes:

[0014] The three-dimensional ground penetrating radar is used to scan the road to be identified through a multi-channel antenna array to obtain radar images of each channel;

[0015] For the radar image of each channel, a bilinear interpolation method is used to interpolate the pixel points in the radar image to obtain an interpolated radar image of the corresponding channel;

[0016] The interpolated radar images of each channel are spliced ​​to obtain the road image to be identified.

[0017] Optionally, the crack Unet neural network includes an encoding subnetwork and a decoding subnetwork; the encoding subnetwork is connected to the decoding subnetwork;

[0018] The method of training the crack Unet neural network according to the training sample set to obtain a crack recognition model specifically includes:

[0019] Standardizing the road sample images in the training sample set to obtain corresponding standardized images and label information of each pixel in each standardized image;

[0020] For each standardized image, encoding the standardized image through the encoding subnetwork to obtain an encoding feature map;

[0021] Decoding the encoded feature map through the decoding subnetwork to obtain a decoded feature map;

[0022] Classify each pixel in the decoding feature map based on a softmax activation function, and determine a prediction result of each pixel in the decoding feature map; the prediction result includes 0 and 1;

[0023] According to the prediction results of each pixel in the decoding feature map and the label information of each pixel in the standardized image, the loss function is determined, and the encoding subnetwork and the decoding subnetwork are iteratively trained according to the loss function until the loss function converges to obtain the optimal crack Unet neural network, and the optimal crack Unet neural network is the crack recognition model.

[0024] Optionally, the encoding feature map includes a first residual map, a second residual map, a third residual map, a fourth residual map and a final encoding feature map;

[0025] The encoding sub-network includes a first encoding layer, a first maximum pooling layer, a second encoding layer, a second maximum pooling layer, a third encoding layer, a third maximum pooling layer, a fourth encoding layer, a fourth maximum pooling layer and a fifth encoding layer connected in sequence; the first encoding layer, the second encoding layer, the third encoding layer, the fourth encoding layer and the fifth encoding layer are also connected to the decoding sub-network;

[0026] The encoding of the standardized image by the encoding subnetwork to obtain an encoding feature map specifically includes:

[0027] Encoding the standardized image through a first encoding layer to obtain a first encoding feature map;

[0028] Performing a pooling operation on the first encoding feature map through a first maximum pooling layer to obtain a first pooling feature map;

[0029] Performing residual compression on the first encoding feature map to obtain a first residual map;

[0030] Encoding the first pooled feature map through a second encoding layer to obtain a second encoded feature map;

[0031] Performing a pooling operation on the second encoded feature map through a second maximum pooling layer to obtain a second pooled feature map;

[0032] Performing residual compression on the second encoded feature map to obtain a second residual map;

[0033] Encoding the second pooled feature map through a third encoding layer to obtain a third encoded feature map;

[0034] Performing a pooling operation on the third encoding feature map through a third maximum pooling layer to obtain a third pooling feature map;

[0035] Performing residual compression on the third encoding feature map to obtain a third residual map;

[0036] Encoding the third pooled feature map through a fourth encoding layer to obtain a fourth encoded feature map;

[0037] Performing a pooling operation on the fourth encoding feature map through a fourth maximum pooling layer to obtain a fourth pooling feature map;

[0038] Performing residual compression on the fourth encoding feature map to obtain a fourth residual map;

[0039] The fourth pooling feature map is encoded through a fifth encoding layer to obtain a final encoded feature map.

[0040] Optionally, the decoding subnetwork includes a first upsampling layer, a first decoding layer, a second upsampling layer, a second decoding layer, a third upsampling layer, a third decoding layer, a fourth upsampling layer and a fourth decoding layer connected in sequence; the first encoding layer is jump-connected to the fourth decoding layer; the second encoding layer is jump-connected to the third decoding layer; the third encoding layer is jump-connected to the second decoding layer; the fourth encoding layer is jump-connected to the first decoding layer; the fifth encoding layer is connected to the first upsampling layer;

[0041] The decoding of the encoded feature map by the decoding subnetwork to obtain a decoded feature map specifically includes:

[0042] Upsampling the encoded feature map through a first upsampling layer to obtain a first upsampling feature map;

[0043] Splicing the first up-sampled feature map and the fourth residual map to obtain a first spliced ​​map;

[0044] Decoding the first spliced ​​graph through a first decoding layer to obtain a first decoded graph;

[0045] Upsampling the first decoding image through a second upsampling layer to obtain a second upsampling feature image;

[0046] Splicing the second up-sampled feature map and the third residual map to obtain a second spliced ​​map;

[0047] Decoding the second spliced ​​graph through a second decoding layer to obtain a second decoded graph;

[0048] Upsampling the second decoding image through a third upsampling layer to obtain a third upsampling feature image;

[0049] Splicing the third up-sampled feature map with the second residual map to obtain a third spliced ​​map;

[0050] Decoding the third spliced ​​graph through a third decoding layer to obtain a third decoded graph;

[0051] Upsampling the third decoding image through a fourth upsampling layer to obtain a fourth upsampling feature image;

[0052] Splicing the fourth up-sampled feature map with the first residual map to obtain a fourth spliced ​​map;

[0053] The fourth spliced ​​image is decoded by a fourth decoding layer to obtain a decoded feature image.

[0054] Optionally, the first encoding layer includes a first convolution-feature selection module and a first parallel multi-directional attention mechanism module connected in sequence; the first parallel multi-directional attention mechanism module is also connected to the first maximum pooling layer and the decoding subnetwork;

[0055] The encoding of the standardized image through the first encoding layer to obtain a first encoding feature map specifically includes:

[0056] Performing feature extraction on the standardized image through a first convolution-feature selection module to obtain a first feature map;

[0057] A first parallel multi-directional attention mechanism module is used to establish an association relationship between pixels in the first feature map to obtain a first encoded feature map.

[0058] Optionally, the first convolution-feature selection module includes four two-dimensional convolution kernels and a channel attention module; the kernel sizes of the four two-dimensional convolution kernels are 3×3, 3×3, 2×2, and 2×2 respectively;

[0059] The step of extracting features from the standardized image by a first convolution-feature selection module to obtain a first feature map specifically includes:

[0060] Performing convolution operations on the standardized image through four two-dimensional convolution kernels respectively to obtain four corresponding convolution feature maps;

[0061] The four convolution feature maps are concatenated to obtain a concatenated convolution feature map;

[0062] The concatenated convolutional feature map is weighted by a channel attention module, and feature selection is performed to obtain a first feature map.

[0063] Optionally, the first parallel multi-directional attention mechanism module includes a 2×1 convolution kernel, a 1×2 convolution kernel, a 2×2 convolution kernel, and a 1×1 convolution kernel;

[0064] The step of establishing an association relationship between pixels in the first feature map by using a first parallel multi-directional attention mechanism module to obtain a first encoded feature map specifically includes:

[0065] Establishing an association relationship between each pixel point in the first feature map and the pixel point to its left through a 2×1 convolution kernel to obtain a left pixel association map;

[0066] Establishing an association relationship between each pixel in the first feature map and the pixel below it through a 1×2 convolution kernel to obtain a lower pixel association map;

[0067] Establishing an association relationship between each pixel point in the first feature map and four adjacent pixels on its two diagonal lines through a 2×2 convolution kernel and a 1×1 convolution kernel to obtain a diagonal pixel association map;

[0068] Connecting the left pixel association map, the lower pixel association map and the diagonal pixel association map to obtain a connected pixel association map;

[0069] A first association feature map is determined according to the connected pixel association map and the first feature map.

[0070] To achieve the above object, the present invention also provides the following solution:

[0071] A three-dimensional ground penetrating radar crack disease identification system, comprising:

[0072] A sample acquisition unit, used to acquire a training sample set; the training sample set includes a plurality of road sample images and label information of each pixel in each road sample image; the label information includes 0 and 1; 0 indicates that the corresponding pixel is not a crack, and 1 indicates that the corresponding pixel is a crack;

[0073] The network construction unit is used to construct a crack Unet neural network based on the spatial attention mechanism, VGG16 and Unet neural network;

[0074] A training unit, connected to the sample acquisition unit and the network construction unit respectively, for training the crack Unet neural network according to the training sample set to obtain a crack recognition model;

[0075] A scanning unit, used for scanning the road to be identified by using a three-dimensional ground penetrating radar to obtain an image of the road to be identified;

[0076] The recognition unit is connected to the training unit and the scanning unit respectively, and is used to determine the cracks in the road to be identified according to the road image to be identified, based on the crack recognition model and the road image to be identified.

[0077] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: based on the spatial attention mechanism, VGG16 and Unet neural networks, a crack Unet neural network is constructed, and it is trained according to the training sample set to obtain a crack recognition model. In actual application, according to the road image to be identified, based on the crack recognition model, by classifying each pixel in the image, cracks in the road can be accurately identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0079] Figure 1 It is a flow chart of the method for identifying crack damage by three-dimensional ground penetrating radar of the present invention;

[0080] Figure 2 This is a schematic diagram of the structure of the crack Unet neural network;

[0081] Figure 3 It is a structural diagram of the residual compression module;

[0082] Figure 4 It is a structural diagram of the multi-way convolution-feature selection module;

[0083] Figure 5 It is a structural diagram of the parallel multi-directional attention mechanism module;

[0084] Figure 6 This is a schematic diagram of full-section coverage detection of 3D ground penetrating radar;

[0085] Figure 7 It is a schematic diagram of bilinear interpolation;

[0086] Figure 8(a) is a schematic diagram of the training set loss reduction curve;

[0087] Figure 8(b) is a schematic diagram of the validation set loss decline curve;

[0088] Fig. 9 It is a schematic diagram of the module structure of the three-dimensional ground penetrating radar crack disease identification system of the present invention.

[0089] Explanation of symbols:

[0090] Sample acquisition unit-1, network construction unit-2, training unit-3, scanning unit-4, recognition unit-5. DETAILED DESCRIPTION

[0091] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0092] The purpose of the present invention is to provide a three-dimensional ground penetrating radar crack disease recognition method and system, by constructing a crack Unet neural network based on the spatial attention mechanism, VGG16 and Unet neural network, and by classifying each pixel in the image, cracks in the road can be accurately identified.

[0093] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0094] like Figure 1 As shown, the three-dimensional ground penetrating radar crack disease identification method of the present invention includes:

[0095] S1: Obtain a training sample set. The training sample set includes multiple road sample images and label information of each pixel in each road sample image. The label information includes 0 and 1; 0 indicates that the corresponding pixel is not a crack, and 1 indicates that the corresponding pixel is a crack.

[0096] S2: Construct a crack Unet neural network based on the spatial attention mechanism, VGG16 and Unet neural network.

[0097] S3: According to the training sample set, the crack Unet neural network is trained to obtain a crack recognition model.

[0098] S4: Scan the road to be identified by three-dimensional ground penetrating radar to obtain an image of the road to be identified.

[0099] S5: Determine cracks in the road to be identified according to the road image to be identified and based on the crack identification model and the road image to be identified.

[0100] Furthermore, the cracked Unet neural network includes an encoding subnetwork and a decoding subnetwork; the encoding subnetwork is connected to the decoding subnetwork. Specifically, the cracked Unet neural network (CrackUnet) used in the present invention is an improved Unet semantic segmentation network based on VGG16, and its network structure is as follows: Figure 2 As shown, the gray rectangle in the figure is the feature map output at each step.

[0101] Step S3 specifically includes:

[0102] S31: Standardize the road sample images in the training sample set to obtain the corresponding standardized images and label information of each pixel in each standardized image. In order to make the trained model have better generalization ability and improve training efficiency, the present invention standardizes the road sample images in the training sample set and centralizes the images through the mean. The data after centralization is more in line with the distribution law. Specifically, the following formula is used to standardize the road sample images:

[0103]

[0104] Among them, X′ is the standardized image matrix, X is the road sample image matrix, μ is the mean of the road sample image matrix, and σ is the standard deviation of the road sample image matrix.

[0105] S32: For each standardized image, the standardized image is encoded by the encoding subnetwork to obtain a coding feature map. Specifically, the coding feature map includes a first residual map, a second residual map, a third residual map, a fourth residual map and a final coding feature map. The encoding subnetwork includes a first coding layer, a first maximum pooling layer, a second coding layer, a second maximum pooling layer, a third coding layer, a third maximum pooling layer, a fourth coding layer, a fourth maximum pooling layer and a fifth coding layer connected in sequence; the first coding layer, the second coding layer, the third coding layer, the fourth coding layer and the fifth coding layer are also connected to the decoding subnetwork.

[0106] Step S32 specifically includes: encoding the standardized image through a first coding layer to obtain a first coding feature map. Pooling the first coding feature map through a first maximum pooling layer to obtain a first pooled feature map. Residual compression is performed on the first coding feature map to obtain a first residual map. In this embodiment, the image size of the standardized image is 512×512, which is convenient for downsampling by a multiple of 2, and the size is as close to the original image size as possible to prevent feature loss caused by image reduction.

[0107] The first pooled feature map is encoded by a second encoding layer to obtain a second encoded feature map. The second encoded feature map is pooled by a second maximum pooling layer to obtain a second pooled feature map. The second encoded feature map is residually compressed to obtain a second residual map.

[0108] The second pooled feature map is encoded by a third encoding layer to obtain a third encoded feature map. The third encoded feature map is pooled by a third maximum pooling layer to obtain a third pooled feature map. The third encoded feature map is residually compressed to obtain a third residual map.

[0109] The third pooling feature map is encoded by a fourth encoding layer to obtain a fourth encoding feature map. The fourth encoding feature map is pooled by a fourth maximum pooling layer to obtain a fourth pooling feature map. The fourth encoding feature map is residual compressed to obtain a fourth residual map.

[0110] The fourth pooling feature map is encoded through a fifth encoding layer to obtain a final encoded feature map.

[0111] In this embodiment, RS (Residual Squeeze Block) is used to perform residual compression on each coding feature map, compress feature channels, and reduce the number of parameters of the network model. In the crack Unet neural network, the decoding layer will connect the features of the corresponding coding layer as a supplement, but since the coding layer is shallower than the decoding layer and contains a large amount of "rough" information, connecting all these features will require the decoding layer to "refine" these features again when supplementing the information, and because not all information is valid, this is tantamount to adding a lot of noise to the deep features. Therefore, it is only necessary to supplement the decoding layer with appropriate information, and there is no need to connect all the features to the decoding layer. Therefore, the RS module is used to remove redundant channels to achieve the purpose of reducing the number of parameters. The structure of the residual compression module used in the present invention is very simple, such as Figure 3 As shown. After the feature map enters the residual compression module, it will be pooled with a step size of 2 to retain the eigenvalue with the largest response, and then the resolution will be restored using unpooling to retain the most effective features in space for connection. Finally, a 2×2 ordinary convolution is used to change the feature channel. The compression rate in this embodiment is 0.25, that is, the number of channels is compressed to 1 / 4 of the original.

[0112] In this embodiment, the first coding layer includes a first convolution-feature selection module and a first parallel multi-directional attention mechanism module connected in sequence. The first parallel multi-directional attention mechanism module is also connected to the first maximum pooling layer and the decoding subnetwork. The standardized image is encoded by the first coding layer to obtain a first encoded feature map, which specifically includes:

[0113] The first convolution-feature selection module is used to extract features from the standardized image to obtain a first feature map. Specifically, the first convolution-feature selection module includes four two-dimensional convolution kernels and a SENet (Sequeeze-Excitation, channel attention) module. The kernel sizes of the four two-dimensional convolution kernels are 3×3, 3×3, 2×2, and 2×2, respectively. The standardized image is convolved by four two-dimensional convolution kernels to obtain four corresponding convolution feature maps. The four convolution feature maps are spliced ​​to obtain a spliced ​​convolution feature map. The spliced ​​convolution feature map is weighted by the channel attention module, and feature selection is performed to obtain a first feature map.

[0114] Specifically, the multi-channel convolution-feature selection module can provide diverse features and give higher weights to effective features from the channel dimension. The MC-FS (Multi Convolution-Feature Selection Block) module consists of parallel convolution, dilated convolution, and channel attention modules. Its structure is as follows: Figure 4 As shown in the figure, w, h, and c represent the width, height, and number of channels of the input features, respectively. The input features are first subjected to four parallel convolutions, two of which are 3×3 and 2×2 dilated convolutions with a dilation rate of 2. The purpose is to obtain a larger receptive field with a small number of parameters. The addition of 3×3 and 2×2 standard convolutions avoids the chessboard effect of dilated convolutions and eliminates the discontinuity of features extracted by dilated convolutions and the problem of local feature loss. Therefore, for one input feature, the MC-FS module generates 4 features, and the number of channels of each feature is 1 / 4 of the output feature. After connecting these four features, the number of output channels is restored. Finally, a channel attention module is added to weight each channel to achieve feature selection and give effective channels higher weights. For the standardized image U∈R H×W×C A compression operation is performed, usually global pooling, to compress the feature shape to 1×1×C, and then an activation operation is performed, usually using a normal convolution or fully connected layer to learn the channel weights. Finally, the feature map U is weighted to obtain the first feature map Similarly, the shape of the feature does not change before and after the channel attention is used. The combination of dilated convolution and standard convolution captures both large and small receptive fields, avoiding the checkerboard effect while reducing the number of parameters. Four-way parallel convolution makes the extracted features more diverse, and the feature weighting of the diversity of the channel attention module can make the extracted features more accurate and effective.

[0115] The first parallel multi-directional attention mechanism module establishes an association relationship between the pixels in the first feature map to obtain a first encoded feature map. Specifically, the first parallel multi-directional attention mechanism module includes a 2×1 convolution kernel, a 1×2 convolution kernel, a 2×2 convolution kernel, and a 1×1 convolution kernel. The 2×1 convolution kernel is used to establish an association relationship between each pixel point in the first feature map and the pixel point to its left, and a left pixel association map is obtained. The 1×2 convolution kernel is used to establish an association relationship between each pixel point in the first feature map and the pixel point below it, and a lower pixel association map is obtained. The 2×2 convolution kernel and the 1×1 convolution kernel are used to establish an association relationship between each pixel point in the first feature map and the four adjacent pixels on its two diagonals, and a diagonal pixel association map is obtained. The left pixel association map, the lower pixel association map, and the diagonal pixel association map are connected to obtain a connected pixel association map. According to the connected pixel association map and the first feature map, a first associated feature map is determined.

[0116] The attention mechanism is a commonly used lightweight module to improve network performance. Its principle is to weight the specific dimensions of the features. This weighting is self-learned by the neural network and does not require human intervention. The PMDA (Parallel Multi-Directional Attention) used in the present invention is a spatial attention mechanism. Its principle is to decompose a 3×3 convolution kernel into 4 parts, and artificially establish an association relationship between a pixel and its adjacent pixels from a structural perspective. Its structure is as follows: Figure 5 shown.

[0117] PMDA can be divided into four paths. From top to bottom, for the input feature A of shape H×W×C, the first two paths use a 2×1 and a 1×2 convolution kernel to establish an association relationship between each pixel in feature A (the first feature map) and the pixels to its left and below, and the two convolutions obtain features B and C respectively. The third and fourth paths use a 2×2 convolution kernel with an expansion rate of 2 and a 1×1 convolution kernel. The two paths work together to establish an association relationship between each pixel and the four adjacent pixels on its two diagonals. Therefore, the results of these two paths are added to obtain the fused feature D. Then, the connection result F of features B, C, and D is combined with a 2×2 convolution kernel to merge the channels of feature F to obtain feature G, and the sigmoid activation function is used to convert its value into a weight in the (0,1) interval. Finally, G is multiplied by the original feature map A to complete the pixel weighting in the feature map to obtain A′ (the first associated feature map). Since only 5 convolution kernels are used and the convolution kernels are very small, the attention module only adds a few parameters to the model and is easy to implement.

[0118] Because cracks are usually long and continuous, if a pixel is a crack pixel, there is a high probability that there is a crack pixel in its adjacent pixels; if this pixel is the background, its adjacent area is also likely to be the background, so establishing an association relationship between pixels is more conducive to crack feature extraction. A PMDA module is added to each layer of the encoding subnetwork in the crack Unet neural network of the present invention to improve the crack feature extraction capability of the encoder.

[0119] Further, the second coding layer includes a second convolution-feature selection module and a second parallel multi-directional attention mechanism module connected in sequence. The second convolution-feature selection module is also connected to the first maximum pooling layer, and the second parallel multi-directional attention mechanism module is also connected to the second maximum pooling layer and the third decoding layer. The first pooling feature map is feature extracted by the second convolution-feature selection module to obtain a second feature map. The pixels in the second feature map are associated with each other by the second parallel multi-directional attention mechanism module to obtain a second coding feature map.

[0120] The third coding layer includes a third convolution-feature selection module, a first two-dimensional convolution module and a third parallel multi-directional attention mechanism module connected in sequence. The third convolution-feature selection module is also connected to the second maximum pooling layer. The third parallel multi-directional attention mechanism module is also connected to the third maximum pooling layer and the second decoding layer. The second pooling feature map is subjected to feature extraction by the third convolution-feature selection module to obtain a third feature map. The third feature map is subjected to feature extraction by the first two-dimensional convolution module to obtain a first two-dimensional feature map. The pixels in the first two-dimensional feature map are associated with each other by the third parallel multi-directional attention mechanism module to obtain a third coding feature map.

[0121] The fourth coding layer includes a fourth convolution-feature selection module, a second two-dimensional convolution module and a fourth parallel multi-directional attention mechanism module connected in sequence. The fourth convolution-feature selection module is also connected to the third maximum pooling layer. The fourth parallel multi-directional attention mechanism module is also connected to the fourth maximum pooling layer and the first decoding layer. The third pooling feature map is extracted by the fourth convolution-feature selection module to obtain a fourth feature map. The fourth feature map is extracted by the second two-dimensional convolution module to obtain a second two-dimensional feature map. The pixels in the second two-dimensional feature map are associated with each other by the fourth parallel multi-directional attention mechanism module to obtain a fourth coding feature map.

[0122] The fifth coding layer includes a fifth convolution-feature selection module, a third two-dimensional convolution module and a fifth parallel multi-directional attention mechanism module connected in sequence. The fifth convolution-feature selection module is also connected to the fourth maximum pooling layer. The fifth parallel multi-directional attention mechanism module is also connected to the first upsampling layer. The fourth pooling feature map is feature extracted by the fifth convolution-feature selection module to obtain a fifth feature map. The fifth feature map is feature extracted by the third two-dimensional convolution module to obtain a third two-dimensional feature map. The pixels in the third two-dimensional feature map are associated with each other by the fifth parallel multi-directional attention mechanism module to obtain a final coding feature map.

[0123] Among them, the first two-dimensional convolution module, the second two-dimensional convolution module and the third two-dimensional convolution module are all 3×3 convolution kernels. The processing process of the second convolution-feature selection module, the third convolution-feature selection module, the fourth convolution-feature selection module and the fifth convolution-feature selection module is the same as the processing process of the first convolution-feature selection module, and the processing process of the second parallel multi-directional attention mechanism module, the third parallel multi-directional attention mechanism module, the fourth parallel multi-directional attention mechanism module and the fifth parallel multi-directional attention mechanism module is the same as the processing process of the first parallel multi-directional attention mechanism module, which will not be repeated here.

[0124] The multi-way convolution-feature selection module in the network is placed at the first position of each layer to increase the number of channels while extracting multiple features, and then select these features to ensure the validity of the input features in subsequent operations, thereby improving the encoding ability of the encoding subnetwork. The network chooses to use a 3×3 convolution kernel for further feature extraction because the MC-FS module contains dilated convolutions and already has a large receptive field. At this time, using a small convolution kernel can reduce the overall number of parameters of the model. Then the parallel multi-directional attention mechanism module will establish an association relationship between the pixels of the features, thereby weighting the crack pixels and enhancing the network's ability to extract crack features. Finally, the output features perform maximum pooling to reduce the resolution of the features.

[0125] S33: Decode the encoding feature map through the decoding subnetwork to obtain a decoding feature map. Specifically, the decoding subnetwork includes a first upsampling layer, a first decoding layer, a second upsampling layer, a second decoding layer, a third upsampling layer, a third decoding layer, a fourth upsampling layer and a fourth decoding layer connected in sequence; the first encoding layer is jump-connected with the fourth decoding layer; the second encoding layer is jump-connected with the third decoding layer; the third encoding layer is jump-connected with the second decoding layer; the fourth encoding layer is jump-connected with the first decoding layer; and the fifth encoding layer is connected with the first upsampling layer.

[0126] Step S33 specifically includes: upsampling the encoded feature map through a first upsampling layer to obtain a first upsampling feature map, and concatenating the first upsampling feature map with the fourth residual map to obtain a first concatenated map.

[0127] The first spliced ​​image is decoded by a first decoding layer to obtain a first decoded image. The first decoded image is upsampled by a second upsampling layer to obtain a second upsampled feature image. The second upsampled feature image is spliced ​​with the third residual image to obtain a second spliced ​​image.

[0128] The second spliced ​​image is decoded by a second decoding layer to obtain a second decoded image. The second decoded image is upsampled by a third upsampling layer to obtain a third upsampled feature image. The third upsampled feature image is spliced ​​with the second residual image to obtain a third spliced ​​image.

[0129] The third spliced ​​image is decoded by a third decoding layer to obtain a third decoded image. The third decoded image is upsampled by a fourth upsampling layer to obtain a fourth upsampled feature image. The fourth upsampled feature image is spliced ​​with the first residual image to obtain a fourth spliced ​​image.

[0130] The fourth spliced ​​image is decoded by a fourth decoding layer to obtain a decoded feature image.

[0131] The last layer of the encoding subnetwork begins to enter the decoding subnetwork. The output features of the last layer of the encoding subnetwork are upsampled, and the resolution becomes twice the original to form the features of the decoding subnetwork part, and connected with the residual map output by the corresponding encoding layer. The features of the encoding layer first enter the compression module for channel compression, remove unnecessary shallow features and reduce the network size before connecting to prevent information loss, and then perform two feature extractions. The above steps are performed four times to complete the decoding. The resolution of the decoded feature map is the same as the original. Figure 1 The number of channels is the number of categories, and then softmax activation is performed on all channels of each pixel to complete pixel-level classification.

[0132] S34: Classify each pixel in the decoding feature map based on the softmax activation function to determine the prediction result of each pixel in the decoding feature map; the prediction result includes 0 and 1.

[0133] S35: Determine a loss function according to the prediction results of each pixel in the decoding feature map and the label information of each pixel in the standardized image, and iteratively train the encoding subnetwork and the decoding subnetwork according to the loss function until the loss function converges to obtain an optimal crack Unet neural network, and the optimal crack Unet neural network is a crack recognition model.

[0134] Since cracks are long and thin features, their pixels account for a small proportion of the image. In this paper, Dice Loss and Focal Loss are used as loss functions:

[0135]

[0136] Among them, Loss is the loss function value, N is the total number of pixels of the standardized image, and p i is the prediction result of the i-th pixel in the standardized image, y i is the label information of the i-th pixel in the standardized image, γ is a constant 2, p t is the predicted probability of the model predicting that the prediction result is a positive sample,

[0137] Furthermore, step S4 specifically includes:

[0138] S41: Use 3D ground penetrating radar to scan the road to be identified through a multi-channel antenna array to obtain radar images of each channel. The multi-channel antenna array has a scanning width of 1.5 meters each time. By setting up multiple detection channels and combining multiple data with accurate positioning data in the later stage, it can achieve full-section coverage scanning of any road, such as Figure 6 In the process of radar data recognition and processing, the three-dimensional radar data is converted into two-dimensional image data in the form of slices, and an XYZ three-dimensional coordinate system is established. The direction of the radar antenna is the x-axis, the depth direction perpendicular to the earth is the z-axis, and the direction perpendicular to the XZ plane is the y-axis. Three mutually perpendicular slices can be obtained, namely xy slices, yz slices and xz slices.

[0139] S42: For the radar image of each channel, use a bilinear interpolation method to interpolate pixel points in the radar image to obtain an interpolated radar image of the corresponding channel.

[0140] Specifically, the data matrix collected by the 3D radar has a horizontal (y) spacing of 0.071m, a forward (x) spacing of 0.025m, and a depth (z) spacing of approximately 0.009m. In order to make the data matrix uniform, it is necessary to interpolate the horizontal (y) and forward (x) directions to make the spacing close to the depth direction, so that the spacing between the x, y, and z dimensions is 0.009m. The interpolation calculation method is as follows: Figure 7 As shown. For any adjacent detection data point P1(x 1 ,y 1 ,f1),P2(x 1 ,y 2 ,f 2 )、P3(x 2 ,y 2 ,f 3 )、P4(x 2 ,y1 ,f 4 ), f 1 、f 2 、f 3 、f 4 are the radar data values ​​of the four points. Insert a point Q(x, y, f) at any position in the rectangular area with the four points as corner points. Use the following formula to determine the radar data value f of the random point:

[0141]

[0142] Based on the above formula, the radar data value at any coordinate can be calculated. The crack signal is not obvious on the YZ plane and XZ plane, but is more obvious on the XY plane. This is related to the fact that the crack disease is long, but the width and height are generally small. Therefore, the XY plane of the three-dimensional radar data is mainly intercepted to identify the crack disease.

[0143] S43: splicing the interpolated radar images of each channel to obtain a road image to be identified.

[0144] Since some areas may be missed during the lane scanning process, black areas may be left in the image. Each radar signal area divided by the black area corresponds to a lane. Since obstacles may be encountered during driving, the radar image of each lane may be curved or overlapped, but most of the effective radar information will not be affected and crack segmentation can still be achieved.

[0145] The method for obtaining the training sample set in step S1 is the same as the method for obtaining the road image to be identified, and will not be repeated here. The data set used by the training model consists of the ground penetrating radar XY plane. After scanning each track, it is necessary to splice the pictures of each track to obtain a complete road radar picture. The above method is used to obtain a large number of radar images of the horizontal surface of the road, and an appropriate number of crack images and disease-free images are selected from the image to form a data set. The crack signal of the radar image is the same as the crack shape seen on the road surface, and the color is darker or brighter than the surrounding area. When selecting, it should be ensured that the data set contains cracks of various common shapes, including transverse cracks, longitudinal cracks, and cross cracks.

[0146] The data set used in the present invention contains 4915 pictures, each with a resolution of 604×604. After the selection is completed, the crack area of ​​the image is annotated using the Labelme tool. Because radar images are different from common RGB images, there may be more interference signals, or they may be affected by different working conditions, and the edge pixels of the cracks cannot be accurately determined. And because the black area in the middle is a vacant area during connection and does not belong to the crack part, it is divided into two parts for annotation. Finally, 3915 pictures are used as training sets and verification sets, with a ratio of 8:2. The verification set does not participate in training and is used to verify the results of each round of training. The remaining pictures are used as a test set. The test set is divided into two parts. The first part is a crack test set. Each picture contains cracks, a total of 500 pictures, which are used to test the model's ability to segment cracks. The remaining 500 pictures are a crack-free test set, which is used to test whether the model is prone to misjudgment.

[0147] In order to prove the recognition accuracy of the Crack Unet provided by the present invention for cracks, the effects of different algorithms are compared and described below.

[0148] The training configuration used in the experiment is a computing platform with a graphics processor model of NVIDIA 1080Ti, 11GB video memory, and an operating system of CenterOS7. The deep learning framework used is Keras with TensorFlow-gpu as the backend. During the training process, random horizontal flipping, translation, and scaling are used to enhance the data set. The optimizer selected is Adam, the initial learning rate is set to 10-5, the batch is set to 3, the training round is set to 200, and the mechanism of automatic reduction of learning rate and early termination is used, that is, the loss of the validation set is monitored. If the validation loss does not decrease within 6 rounds, the learning rate is halved. If it continues for 10 rounds, the training is terminated. The decline curves of the training set loss and the validation set loss are shown in Figure 8(a) and Figure 8(b). Due to the automatic stop strategy, the model converged at the 63rd iteration.

[0149] The present invention uses MPA (Mean Pixel Accuracy), MIoU (Mean Intersection over Union), FLOPs (Floating-point operations) and FPS (Frames Per Second) as evaluation indicators for the model. The average pixel accuracy is the ratio of the number of correctly classified pixels to all pixels, which can be used to evaluate the accuracy of pixel-level classification. Each pixel in the segmentation result corresponds to one of the following four categories:

[0150] (1) True Positive (TP): The model predicts that the example is a positive example, and the label is also a positive example;

[0151] (2) False Positive (FP): The model predicts a positive example, but the label is a negative example;

[0152] (3) False Negative (FN): The model predicts a negative example, but the label is a positive example;

[0153] (4) True Negative (TN): The model predicts a negative example and the label is also a negative example.

[0154] For a category in semantic segmentation, PA (Pixel Accuracy) is:

[0155]

[0156] MPA is the mean of the sum of PAs of all categories. In addition to the target category to be segmented, semantic segmentation also has a default category of background. Therefore, the segmentation categories of the present invention include crack and background. IoU (Intersection over Union) is the intersection of the label area and the prediction area divided by their union. That is, for the label area T a and prediction area C a , IoU is expressed as:

[0157]

[0158] MIoU is the average of the sum of IoUs of all categories. IoU can evaluate the expected overlap between the segmented area and the label. FLOPs represents the number of floating point operations of the neural network, which can be used to evaluate the computational complexity or complexity of the model. The FLOPs of a single convolutional layer are calculated as follows:

[0159] FLOPs = (2 × C in ×K 2 -1)×H×W×C out ;

[0160] Among them, C in is the number of input channels, K is the convolution kernel size, H and W are the height and width of the output feature map, respectively, and C out is the number of output channels.

[0161] The present invention uses the current mainstream model to compare the semantic segmentation indicators with Crack UNet in the test set. The test is carried out in the crack data set. The comparison model includes deepLabv3, PSPNet, and Unet. All the test environments of the present invention are the same as the training environment. The comparison results are shown in Table 1:

[0162] Table 1 Comparison of indicators of different models

[0163]

[0164]

[0165] It can be seen that the MPA and MIoU of the Crack Unet provided by the present invention are higher than those of other models. The MPA is 7.35%, 2.27%, and 3.6% higher than deeplabv3, PSPNet, and Unet, respectively, and the MIoU is 1.4%, 0.99%, and 2.27% higher, respectively, which are all greatly improved, and the FLOPs of the model is also the lowest. Although the FPS of CrackNet is still low and the segmentation speed is slow, it has reached the level of real-time detection. In the actual engineering application of road detection, more attention is paid to the ability to identify diseases. Therefore, compared with the current mainstream algorithm, the present invention has better segmentation ability in the crack segmentation task of radar images.

[0166] In order to test the accuracy of the model under crack-free conditions, the present invention tests all models in crack-free data sets and calculates the accuracy of each model in crack-free data sets at different false alarm percentage thresholds. The false alarm percentage here is the percentage of misclassified pixels in a picture to the total pixels. The experiment uses three thresholds, which are <1%, <0.5% and <0.1%. The test results are shown in Table 2.

[0167] Table 2 Segmentation accuracy of each model under different false positive ratio thresholds in the crack-free test set

[0168] Proportion DeepLabv3 Unet PSPNet Crack Unet <0.1％ 92.80％ 53.80％ 90.80％ 88.80％ <0.5％ 98.20％ 80.40％ 95.80％ 94.40％ <1％ 99.40％ 93.40％ 98.53％ 97.20％

[0169] In the test of the crack-free dataset, Unet has the lowest accuracy at each threshold. Although the accuracy of Crack UNet is not the highest at 88.80%, it is not much different from DeepLabsv3 and PSPNet. Combining the accuracy of the two datasets, it can be seen that the performance of DeepLabv3 and Unet is not stable. DeepLabv3 performs poorly in the crack test set, but performs best in the crack-free test set. Therefore, DeepLabv3 has low segmentation ability. On the contrary, Unet performs well in the crack test set, but produces more false positives in the crack-free test set. Therefore, Unet has good segmentation ability, but it is easy to misclassify and has low robustness. PSPNet is more stable and has higher accuracy, but PSPNet has lower accuracy in the crack test set. Although Crack UNet does not perform best in the crack-free test set, the gap with DeepLabv3 and PSPNet is small.

[0170] like Fig. 9 As shown, the three-dimensional ground penetrating radar crack disease identification system of the present invention includes: a sample acquisition unit 1, a network construction unit 2, a training unit 3, a scanning unit 4 and an identification unit 5.

[0171] The sample acquisition unit 1 is used to acquire a training sample set. The training sample set includes multiple road sample images and label information of each pixel in each road sample image. The label information includes 0 and 1; 0 indicates that the corresponding pixel is not a crack, and 1 indicates that the corresponding pixel is a crack.

[0172] The network construction unit 2 is used to construct a crack Unet neural network based on the spatial attention mechanism, VGG16 and Unet neural network.

[0173] The training unit 3 is connected to the sample acquisition unit 1 and the network construction unit 2 respectively. The training unit 3 is used to train the crack Unet neural network according to the training sample set to obtain a crack recognition model.

[0174] The scanning unit 4 is used to scan the road to be identified by using a three-dimensional ground penetrating radar to obtain an image of the road to be identified.

[0175] The recognition unit 5 is connected to the training unit 3 and the scanning unit 4 respectively, and is used to determine cracks in the road to be identified according to the road image to be identified, based on the crack recognition model and the road image to be identified.

[0176] Specifically, the scanning unit 4 includes: a channel scanning module, an interpolation module and a splicing module.

[0177] The channel scanning module is used to use a three-dimensional ground penetrating radar to scan the road to be identified through a multi-channel antenna array to obtain a radar image of each channel.

[0178] The interpolation module is connected to the channel scanning module, and is used for interpolating the pixel points in the radar image of each channel by using a bilinear interpolation method to obtain an interpolated radar image of the corresponding channel.

[0179] The stitching module is connected to the interpolation module, and is used to stitch the interpolated radar images of each channel to obtain a road image to be identified.

[0180] Compared with the prior art, the three-dimensional ground penetrating radar crack disease identification system of the present invention has the same beneficial effects as the above-mentioned three-dimensional ground penetrating radar crack disease identification method, which will not be repeated here.

[0181] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0182] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A three-dimensional ground penetrating radar crack disease identification method, characterized in that: The three-dimensional ground penetrating radar crack disease identification method comprises: Acquire a training sample set; the training sample set includes a plurality of road sample images and label information of each pixel in each road sample image; the label information includes 0 and 1; 0 indicates that the corresponding pixel is not a crack, and 1 indicates that the corresponding pixel is a crack; Based on the spatial attention mechanism, VGG16 and Unet neural network, a crack Unet neural network is constructed; the crack Unet neural network includes an encoding subnetwork and a decoding subnetwork; the encoding subnetwork is connected to the decoding subnetwork; The encoding sub-network includes a first encoding layer, a first maximum pooling layer, a second encoding layer, a second maximum pooling layer, a third encoding layer, a third maximum pooling layer, a fourth encoding layer, a fourth maximum pooling layer and a fifth encoding layer connected in sequence; the first encoding layer, the second encoding layer, the third encoding layer, the fourth encoding layer and the fifth encoding layer are also connected to the decoding sub-network; The decoding subnetwork includes a first upsampling layer, a first decoding layer, a second upsampling layer, a second decoding layer, a third upsampling layer, a third decoding layer, a fourth upsampling layer and a fourth decoding layer connected in sequence; the first coding layer is jump-connected to the fourth decoding layer; the second coding layer is jump-connected to the third decoding layer; the third coding layer is jump-connected to the second decoding layer; the fourth coding layer is jump-connected to the first decoding layer; the fifth coding layer is connected to the first upsampling layer; The first encoding layer includes a first convolution-feature selection module and a first parallel multi-directional attention mechanism module connected in sequence; the first parallel multi-directional attention mechanism module is also connected to the first maximum pooling layer and the decoding sub-network; The second encoding layer includes a second convolution-feature selection module and a second parallel multi-directional attention mechanism module connected in sequence; the second convolution-feature selection module is also connected to the first maximum pooling layer, and the second parallel multi-directional attention mechanism module is also connected to the second maximum pooling layer and the third decoding layer; The third encoding layer includes a third convolution-feature selection module, a first two-dimensional convolution module and a third parallel multi-directional attention mechanism module connected in sequence; the third convolution-feature selection module is also connected to the second maximum pooling layer; the third parallel multi-directional attention mechanism module is also connected to the third maximum pooling layer and the second decoding layer; The fourth encoding layer includes a fourth convolution-feature selection module, a second two-dimensional convolution module and a fourth parallel multi-directional attention mechanism module connected in sequence; the fourth convolution-feature selection module is also connected to the third maximum pooling layer; the fourth parallel multi-directional attention mechanism module is also connected to the fourth maximum pooling layer and the first decoding layer; The fifth encoding layer includes a fifth convolution-feature selection module, a third two-dimensional convolution module and a fifth parallel multi-directional attention mechanism module connected in sequence; the fifth convolution-feature selection module is also connected to the fourth maximum pooling layer; the fifth parallel multi-directional attention mechanism module is also connected to the first upsampling layer; According to the training sample set, the crack Unet neural network is trained to obtain a crack recognition model; Scanning the road to be identified by three-dimensional ground penetrating radar to obtain an image of the road to be identified; According to the road image to be identified, based on the crack identification model, and according to the road image to be identified, cracks in the road to be identified are determined.

2. The three-dimensional ground penetrating radar crack disease identification method according to claim 1 is characterized in that: The method of scanning the road to be identified by three-dimensional ground penetrating radar to obtain the image of the road to be identified specifically includes: The three-dimensional ground penetrating radar is used to scan the road to be identified through a multi-channel antenna array to obtain radar images of each channel; For the radar image of each channel, a bilinear interpolation method is used to interpolate the pixel points in the radar image to obtain an interpolated radar image of the corresponding channel; The interpolated radar images of each channel are spliced ​​to obtain the road image to be identified.

3. The three-dimensional ground penetrating radar crack disease identification method according to claim 1 is characterized in that: According to the training sample set, the crack Unet neural network is trained to obtain a crack recognition model, which specifically includes: Standardizing the road sample images in the training sample set to obtain corresponding standardized images and label information of each pixel in each standardized image; For each standardized image, encoding the standardized image through the encoding subnetwork to obtain an encoding feature map; Decoding the encoded feature map through the decoding subnetwork to obtain a decoded feature map; Classify each pixel in the decoding feature map based on a softmax activation function, and determine a prediction result of each pixel in the decoding feature map; the prediction result includes 0 and 1; According to the prediction results of each pixel in the decoding feature map and the label information of each pixel in the standardized image, the loss function is determined, and the encoding subnetwork and the decoding subnetwork are iteratively trained according to the loss function until the loss function converges to obtain the optimal crack Unet neural network, and the optimal crack Unet neural network is the crack recognition model.

4. The three-dimensional ground penetrating radar crack disease identification method according to claim 3 is characterized in that: The coding feature map includes a first residual map, a second residual map, a third residual map, a fourth residual map and a final coding feature map; The encoding of the standardized image by the encoding subnetwork to obtain an encoding feature map specifically includes: Encoding the standardized image through a first encoding layer to obtain a first encoding feature map; Performing a pooling operation on the first encoding feature map through a first maximum pooling layer to obtain a first pooling feature map; Performing residual compression on the first encoding feature map to obtain a first residual map; Encoding the first pooled feature map through a second encoding layer to obtain a second encoded feature map; Performing a pooling operation on the second encoded feature map through a second maximum pooling layer to obtain a second pooled feature map; Performing residual compression on the second encoded feature map to obtain a second residual map; Encoding the second pooled feature map through a third encoding layer to obtain a third encoded feature map; Performing a pooling operation on the third encoding feature map through a third maximum pooling layer to obtain a third pooling feature map; Performing residual compression on the third encoding feature map to obtain a third residual map; Encoding the third pooled feature map through a fourth encoding layer to obtain a fourth encoded feature map; Performing a pooling operation on the fourth encoding feature map through a fourth maximum pooling layer to obtain a fourth pooling feature map; Performing residual compression on the fourth encoding feature map to obtain a fourth residual map; The fourth pooling feature map is encoded through a fifth encoding layer to obtain a final encoded feature map.

5. The three-dimensional ground penetrating radar crack disease identification method according to claim 4 is characterized in that: Decoding the encoded feature map through the decoding subnetwork to obtain a decoded feature map specifically includes: Upsampling the encoded feature map through a first upsampling layer to obtain a first upsampling feature map; Splicing the first up-sampled feature map and the fourth residual map to obtain a first spliced ​​map; Decoding the first spliced ​​graph through a first decoding layer to obtain a first decoded graph; Upsampling the first decoding image through a second upsampling layer to obtain a second upsampling feature image; Splicing the second up-sampled feature map and the third residual map to obtain a second spliced ​​map; Decoding the second spliced ​​graph through a second decoding layer to obtain a second decoded graph; Upsampling the second decoding image through a third upsampling layer to obtain a third upsampling feature image; Splicing the third up-sampled feature map with the second residual map to obtain a third spliced ​​map; Decoding the third spliced ​​graph through a third decoding layer to obtain a third decoded graph; Upsampling the third decoding image through a fourth upsampling layer to obtain a fourth upsampling feature image; Splicing the fourth up-sampled feature map with the first residual map to obtain a fourth spliced ​​map; The fourth spliced ​​image is decoded by a fourth decoding layer to obtain a decoded feature image.

6. The three-dimensional ground penetrating radar crack disease identification method according to claim 4 is characterized in that: The standardized image is encoded by a first encoding layer to obtain a first encoding feature map, specifically comprising: Performing feature extraction on the standardized image through a first convolution-feature selection module to obtain a first feature map; A first parallel multi-directional attention mechanism module is used to establish an association relationship between pixels in the first feature map to obtain a first encoded feature map.

7. The three-dimensional ground penetrating radar crack disease identification method according to claim 6 is characterized in that: The first convolution-feature selection module includes four two-dimensional convolution kernels and a channel attention module; the kernel sizes of the four two-dimensional convolution kernels are 3×3, 3×3, 2×2, and 2×2 respectively; The step of extracting features from the standardized image by a first convolution-feature selection module to obtain a first feature map specifically includes: Performing convolution operations on the standardized image through four two-dimensional convolution kernels respectively to obtain four corresponding convolution feature maps; The four convolution feature maps are concatenated to obtain a concatenated convolution feature map; The concatenated convolutional feature map is weighted by a channel attention module, and feature selection is performed to obtain a first feature map.

8. The three-dimensional ground penetrating radar crack disease identification method according to claim 6 is characterized in that: The first parallel multi-directional attention mechanism module includes a 2×1 convolution kernel, a 1×2 convolution kernel, a 2×2 convolution kernel, and a 1×1 convolution kernel; The step of establishing an association relationship between pixels in the first feature map by using a first parallel multi-directional attention mechanism module to obtain a first encoded feature map specifically includes: Establishing an association relationship between each pixel point in the first feature map and the pixel point to its left through a 2×1 convolution kernel to obtain a left pixel association map; Establishing an association relationship between each pixel in the first feature map and the pixel below it through a 1×2 convolution kernel to obtain a lower pixel association map; Establishing an association relationship between each pixel point in the first feature map and four adjacent pixels on its two diagonal lines through a 2×2 convolution kernel and a 1×1 convolution kernel to obtain a diagonal pixel association map; Connecting the left pixel association map, the lower pixel association map and the diagonal pixel association map to obtain a connected pixel association map; A first association feature map is determined according to the connected pixel association map and the first feature map.

9. A three-dimensional ground penetrating radar crack disease identification system, characterized in that: The three-dimensional ground penetrating radar crack disease identification system comprises: A sample acquisition unit, used to acquire a training sample set; the training sample set includes a plurality of road sample images and label information of each pixel in each road sample image; the label information includes 0 and 1; 0 indicates that the corresponding pixel is not a crack, and 1 indicates that the corresponding pixel is a crack; A network construction unit, used for constructing a crack Unet neural network based on a spatial attention mechanism, VGG16 and a Unet neural network; the crack Unet neural network includes an encoding subnetwork and a decoding subnetwork; the encoding subnetwork is connected to the decoding subnetwork; The encoding sub-network includes a first encoding layer, a first maximum pooling layer, a second encoding layer, a second maximum pooling layer, a third encoding layer, a third maximum pooling layer, a fourth encoding layer, a fourth maximum pooling layer and a fifth encoding layer connected in sequence; the first encoding layer, the second encoding layer, the third encoding layer, the fourth encoding layer and the fifth encoding layer are also connected to the decoding sub-network; The decoding subnetwork includes a first upsampling layer, a first decoding layer, a second upsampling layer, a second decoding layer, a third upsampling layer, a third decoding layer, a fourth upsampling layer and a fourth decoding layer connected in sequence; the first coding layer is jump-connected to the fourth decoding layer; the second coding layer is jump-connected to the third decoding layer; the third coding layer is jump-connected to the second decoding layer; the fourth coding layer is jump-connected to the first decoding layer; the fifth coding layer is connected to the first upsampling layer; The first encoding layer includes a first convolution-feature selection module and a first parallel multi-directional attention mechanism module connected in sequence; the first parallel multi-directional attention mechanism module is also connected to the first maximum pooling layer and the decoding sub-network; The second encoding layer includes a second convolution-feature selection module and a second parallel multi-directional attention mechanism module connected in sequence; the second convolution-feature selection module is also connected to the first maximum pooling layer, and the second parallel multi-directional attention mechanism module is also connected to the second maximum pooling layer and the third decoding layer; The third encoding layer includes a third convolution-feature selection module, a first two-dimensional convolution module and a third parallel multi-directional attention mechanism module connected in sequence; the third convolution-feature selection module is also connected to the second maximum pooling layer; the third parallel multi-directional attention mechanism module is also connected to the third maximum pooling layer and the second decoding layer; The fourth encoding layer includes a fourth convolution-feature selection module, a second two-dimensional convolution module and a fourth parallel multi-directional attention mechanism module connected in sequence; the fourth convolution-feature selection module is also connected to the third maximum pooling layer; the fourth parallel multi-directional attention mechanism module is also connected to the fourth maximum pooling layer and the first decoding layer; The fifth encoding layer includes a fifth convolution-feature selection module, a third two-dimensional convolution module and a fifth parallel multi-directional attention mechanism module connected in sequence; the fifth convolution-feature selection module is also connected to the fourth maximum pooling layer; the fifth parallel multi-directional attention mechanism module is also connected to the first upsampling layer; A training unit, connected to the sample acquisition unit and the network construction unit respectively, for training the crack Unet neural network according to the training sample set to obtain a crack recognition model; A scanning unit, used for scanning the road to be identified by using a three-dimensional ground penetrating radar to obtain an image of the road to be identified; The recognition unit is connected to the training unit and the scanning unit respectively, and is used to determine the cracks in the road to be identified according to the road image to be identified, based on the crack recognition model and the road image to be identified.

10. The three-dimensional ground penetrating radar crack disease identification system according to claim 9, characterized in that: The scanning unit comprises: A channel scanning module is used to use a three-dimensional ground penetrating radar to scan the road to be identified through a multi-channel antenna array to obtain a radar image of each channel; An interpolation module, connected to the channel scanning module, for interpolating the pixel points in the radar image of each channel by using a bilinear interpolation method to obtain an interpolated radar image of the corresponding channel; The splicing module is connected to the interpolation module and is used to splice the interpolated radar images of each channel to obtain the road image to be identified.

Images