Tunnel lining steel frame recognition method and related equipment based on GAR-Net

Abstract

The present invention discloses a tunnel lining steel frame recognition method and related equipment based on GAR-Net, performs spatial domain enhancement on geological radar data to improve the robustness of the neural network structure, and constructs a GAR-Net network model, which includes a mileage recognition model, a lining thickness recognition model and a steel arch frame recognition model. The lining thickness recognition model and the steel arch frame recognition model respectively perform lining thickness recognition and steel arch frame recognition on the basis of the mileage recognition model. Whether the tunnel lining steel frame has defects can be judged through the positioning information of the lining thickness and the tunnel steel arch frame, which overcomes the shortcomings of traditional radar image interpretation, can realize automatic recognition of the tunnel steel arch frame and lining thickness, and significantly increases the accuracy and recognition efficiency of automatic identification of geological radar images.

Description

Technical Field

[0001] The present invention relates to the technical field of geological radar identification, and in particular to a tunnel lining steel frame identification method based on GAR-Net and related equipment. Background Art

[0002] In recent years, China has made remarkable progress in the construction of railway and highway tunnels, and its construction scale and speed have reached the world's leading level. However, with the rapid expansion of these infrastructures, the quality of tunnel engineering has gradually become the focus of attention. The insufficient thickness of tunnel lining concrete, the spacing of steel arch frames that does not meet the design requirements, and the voids in the tunnel lining will affect the safety and stability of the tunnel to varying degrees, thereby affecting the engineering quality of the tunnel. In order to ensure the safe operation of the tunnel after completion, special inspections of its structure are required during the construction of the tunnel. In current engineering projects, the internal defects of the tunnel structure are often detected and evaluated by arranging geological radar survey lines. The technicians manually identify the lining thickness, steel arch spacing, and voids behind the lining based on the collected radar images. The accuracy of the identification is heavily dependent on the experience level of the technicians, which often results in different degrees of misjudgment and missed judgment. In the process of radar image interpretation, the manual identification method not only consumes a lot of manpower, but also has relatively low efficiency, and is significantly affected by the subjective judgment of the identifier. In order to overcome the shortcomings of the manual identification method, the automatic identification technology of geological radar images has become a research hotspot in the field of tunnel detection in recent years.

[0003] In the early research on geological radar image recognition, scholars mainly used the "traditional" method of manually extracting radar wave signal features to identify simple signals. Hough transform combined with SVM (support vector machine) and other methods also became one of the mainstream machine learning algorithms at the time. During this period, Maas and Schmalzl et al. used hyperbolic signals in radar images for recognition, and then Sagnard and Tarel et al. proposed a template matching algorithm. Terrasse proposed a new dictionary learning algorithm, which has been widely used in the field of geological radar image defect identification. These studies have greatly promoted the development of geological radar image recognition technology. However, in the face of geological radar engineering application scenarios with complex environmental noise interference, the disadvantages of "traditional" machine learning algorithms such as weak model generalization ability, poor robustness, and insufficient computing power have gradually emerged.

[0004] Therefore, the traditional geological radar method of identifying the steel arch frame and thickness of the tunnel lining mainly uses manual interpretation for judgment. The identification speed is slow and the accuracy depends on the experience level of the technicians, which brings extremely complicated work to the technicians. Summary of the invention

[0005] Based on the problems raised by the above background technology, the purpose of the present invention is to provide a tunnel lining steel frame identification method based on GAR-Net and related equipment. A new network structure GAR-Net is defined in the method. The network consists of three parts: mileage identification, lining thickness identification, and steel arch frame identification. It includes multiple algorithms such as pattern matching, semantic segmentation, and target detection. A tunnel lining steel frame and thickness automatic positioning and identification process based on the GAR-Net algorithm is proposed, which overcomes the shortcomings of traditional radar image interpretation, realizes the automatic identification of tunnel steel arch frames and lining thickness, and significantly increases the accuracy and recognition efficiency of automatic identification of geological radar images. It solves the problem that the traditional geological radar method of tunnel lining steel arch frame and thickness identification mainly uses manual interpretation for identification, has slow identification speed and the identification accuracy depends on the experience level of technicians, thus leading to low efficiency.

[0006] The present invention is achieved through the following technical solutions:

[0007] The first aspect of the present invention provides a tunnel lining steel frame recognition method based on GAR-Net, comprising the following steps:

[0008] Acquire lining geological radar data, perform spatial domain enhancement on the geological radar data, and generate a lining geological radar data set;

[0009] Constructing a GAR-Net network model, and using the lining geological radar data set to train the GAR-Net network model; wherein the GAR-Net network model includes a mileage recognition model, a lining thickness recognition model and a steel arch recognition model, the mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, the lining thickness recognition model is used to identify the tunnel lining thickness, and the steel arch recognition model is used to locate the steel arch based on the mileage coordinate information;

[0010] A geological radar image is obtained, and the trained GAR-Net network model is used to automatically locate and identify the lining thickness and the tunnel steel arch frame of the geological radar image.

[0011] In the above technical scheme, firstly, in the field of deep learning, in order to effectively train the supervised network structure, it is usually necessary to rely on a huge data set, but the tunnel lining geological radar image data is difficult to obtain, so in this method, the geological radar data is spatially enhanced to improve the robustness of the neural network structure.

[0012] Secondly, the GAR-Net network model is constructed, which mainly consists of three parts. The first part is the mileage recognition model, which mainly provides mileage coordinate information for the lining thickness recognition model in the second part and the steel arch recognition model in the third part. The lining thickness recognition model in the second part extracts features from the lining geological radar image and recognizes the lining thickness based on the extracted features based on the mileage coordinate information. The steel arch recognition model in the third part extracts features from the lining geological radar data to realize automatic recognition of steel bars, count the number of steel bars, and extract steel arches from them. It locates the steel bar heads according to the mileage coordinate information and automatically calculates the spacing between steel arches.

[0013] By using the trained GAR-Net network model to identify geological radar images, the lining thickness and the positioning information of the tunnel steel arch frame can be obtained. The lining thickness and the positioning information of the tunnel steel arch frame can be used to determine whether there are defects in the tunnel lining steel frame. This overcomes the shortcomings of traditional radar image interpretation and can realize automatic recognition of the tunnel steel arch frame and lining thickness, significantly increasing the accuracy and recognition efficiency of automatic geological radar image recognition.

[0014] In an optional embodiment, the mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, including:

[0015] Performing black-and-white binary processing on the radar image in the lining geological radar data set to obtain a black-and-white binary radar image;

[0016] Sequentially performing horizontal digital scale separation and vertical digital scale separation on the black-and-white binary radar image to obtain a black-and-white binary radar image with digital scale separation;

[0017] The pattern recognition method is used to digitally recognize the black and white binary radar image with digital scale separation, and the radar image coordinate system is constructed using the digital recognition results.

[0018] In an optional embodiment, the lining thickness identification model includes a contraction path and an expansion path; the contraction path includes four compression modules, and the expansion path includes four expansion modules;

[0019] Using four compression modules to convolve the radar image in the lining geological radar data set, and using four expansion modules to upsample the convolved radar image to obtain a fused feature image;

[0020] Each pixel point in the fused feature image is classified to obtain a background area and a lining thickness area.

[0021] In an optional embodiment, the four compression modules include a first compression module, a second compression module, a third compression module and a fourth compression module connected in sequence; the convolution layer of the first compression module adopts two 64-channel 3x3 convolution kernels, the convolution layer of the second compression module adopts two 128-channel 3x3 convolution kernels, the convolution layer of the third compression module adopts three 256-channel 3x3 convolution kernels, and the convolution layer of the fourth compression module adopts three 512-channel 3x3 convolution kernels;

[0022] The four expansion modules include a first expansion module corresponding to the first compression module, a second expansion module corresponding to the second compression module, a third expansion module corresponding to the third compression module, and a fourth expansion module corresponding to the fourth compression module;

[0023] Among them, the number of channels of the corresponding compression module and expansion module remains consistent.

[0024] In an optional embodiment, the steel arch recognition model is used to locate the steel arch based on the mileage coordinate information, including:

[0025] Extracting features from the radar image in the lining geological radar data set to obtain a steel bar feature image;

[0026] Extracting the contour of the steel bar feature image to obtain a steel bar structure;

[0027] Counting the number of the steel bar structure to obtain the number of steel bars;

[0028] Using the steel bar structure and the number of steel bars to identify the steel arch frame, and obtain the steel arch frame structure;

[0029] The steel arch structure is analyzed using the mileage coordinate information to obtain the position information of the steel bar heads and the spacing information between the steel arches.

[0030] The second aspect of the present invention provides a tunnel lining steel frame recognition system based on GAR-Net, comprising:

[0031] A data set processing module, the data set processing module is used to obtain lining geological radar data, perform spatial domain enhancement on the geological radar data, and generate a lining geological radar data set;

[0032] A training module, wherein the training module is used to construct a GAR-Net network model, and the GAR-Net network model is trained using the lining geological radar data set; wherein the GAR-Net network model includes a mileage recognition model, a lining thickness recognition model, and a steel arch recognition model, wherein the mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, the lining thickness recognition model is used to recognize the tunnel lining thickness, and the steel arch recognition model is used to locate the steel arch based on the mileage coordinate information;

[0033] The recognition module is used to obtain geological radar images and use the trained GAR-Net network model to automatically locate and recognize the lining thickness and tunnel steel arch frame of the geological radar images.

[0034] In an optional embodiment, the training module includes:

[0035] A binarization processing unit, the binarization processing unit is used to perform black-and-white binarization processing on the radar image in the lining geological radar data set to obtain a black-and-white binarized radar image;

[0036] A scale separation unit, the scale separation unit is used to sequentially perform horizontal digital scale separation and vertical digital scale separation on the black-and-white binary radar image to obtain a black-and-white binary radar image with digital scale separation;

[0037] The pattern recognition unit is used to digitally recognize the black-and-white binary radar image with digital scale separation by using a pattern recognition method, and to construct a radar image coordinate system by using the digital recognition result.

[0038] In an optional embodiment, the training module further includes:

[0039] A steel bar feature extraction unit, the steel bar feature extraction unit is used to extract features from the radar image in the lining geological radar data set to obtain a steel bar feature image;

[0040] A contour extraction unit, the contour extraction unit is used to extract the contour of the steel bar feature image to obtain a steel bar structure;

[0041] A quantity counting unit, the quantity counting unit is used to count the quantity of the steel bar structure to obtain the quantity of steel bars;

[0042] A steel arch identification unit, wherein the steel arch identification unit is used to identify the steel arch using the steel bar structure and the number of steel bars to obtain a steel arch structure;

[0043] A positioning unit is used to analyze the steel arch structure using the mileage coordinate information to obtain the position information of the steel bar head and the spacing information between the steel arches.

[0044] A third aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a tunnel lining steel frame identification method based on GAR-NET is implemented.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a tunnel lining steel frame identification method based on GAR-NET.

[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0047] 1. Perform spatial domain enhancement on geological radar data to improve the robustness of the neural network structure;

[0048] 2. The trained GAR-Net network model is used to identify the lining thickness and tunnel steel arch frame of the geological radar image. The positioning information of the lining thickness and the tunnel steel arch frame can be used to determine whether there are defects in the tunnel lining steel frame. This overcomes the shortcomings of traditional radar image interpretation and can realize automatic recognition of the tunnel steel arch frame and lining thickness, significantly increasing the accuracy and recognition efficiency of automatic identification of geological radar images. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0050] Figure 1 A schematic diagram of a process flow of a tunnel lining steel frame identification method based on GAR-Net provided in Example 1 of the present invention;

[0051] Figure 2 A schematic diagram of the structure of the GAR-Net network model provided in Example 1 of the present invention;

[0052] Figure 3 A graph showing the change in loss of the lining thickness recognition model provided in Example 1 of the present invention during the training process;

[0053] Figure 4 The intersection-and-union ratio of the lining thickness recognition model provided in Example 1 of the present invention during the training process;

[0054] Figure 5 The mIOU histogram provided in Example 1 of the present invention;

[0055] Figure 6 This is a thickness recognition effect diagram provided in Example 1 of the present invention;

[0056] Figure 7 A training loss graph of the backbone feature extraction network provided in Example 1 of the present invention;

[0057] Figure 8 The precision-recall graph provided in Example 1 of the present invention;

[0058] Fig. 9 The tunnel lining steel arch recognition effect diagram provided in Example 1 of the present invention;

[0059] Fig.10 Steel arch profile extraction and arch vertex map provided in Example 1 of the present invention;

[0060] Fig.11 The lining thickness and arch spacing identification result diagram provided in Example 1 of the present invention;

[0061] Fig.12 A schematic diagram of the structure of an electronic device provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0062] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0063] Example 1

[0064] Figure 1 The schematic diagram of the process of the tunnel lining steel frame identification method based on GAR-Net provided in Example 1 of the present invention is as follows: Figure 1 As shown in the figure, the tunnel lining steel frame recognition method based on GAR-Net includes the following steps:

[0065] Acquire lining geological radar data, perform spatial domain enhancement on the geological radar data, and generate a lining geological radar data set;

[0066] Constructing a GAR-Net network model, and using the lining geological radar data set to train the GAR-Net network model; wherein the GAR-Net network model includes a mileage recognition model, a lining thickness recognition model and a steel arch recognition model, the mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, the lining thickness recognition model is used to identify the tunnel lining thickness, and the steel arch recognition model is used to locate the steel arch based on the mileage coordinate information;

[0067] A geological radar image is obtained, and the trained GAR-Net network model is used to automatically locate and identify the lining thickness and the tunnel steel arch frame of the geological radar image.

[0068] It should be noted that in recent years, with the leapfrog development of computing performance, shallow feature extraction technology can no longer meet the deep-level needs in the field of computer vision. At present, deep learning algorithms can be used to mine high-dimensional implicit features of images in training data sets to greatly enhance the deep feature extraction of visual images, abandoning the complex image processing process that relied on artificially constructed features in the past, and enhancing the efficiency and recognition accuracy of deep learning algorithms. At present, in the quality inspection of tunnel engineering, some scholars have learned to use deep learning algorithms for the identification and detection of defects such as cavities, water seepage, and cracks. The identification of tunnel lining steel frames requires the identification of both the thickness of the tunnel lining and the steel arch frame. Therefore, based on the above-mentioned technology, the invention proposes a tunnel lining steel frame recognition method based on GAR-Net to solve the identification of tunnel lining steel frames.

[0069] First, in the field of deep learning, in order to effectively train the supervised network structure, it is usually necessary to rely on a large data set, but the tunnel lining geological radar image data is difficult to obtain. Therefore, in this method, the geological radar data is enhanced in the spatial domain to improve the robustness of the neural network structure.

[0070] Secondly, the GAR-Net network model is constructed, which mainly consists of three parts. The first part is the mileage recognition model, which mainly provides mileage coordinate information for the lining thickness recognition model in the second part and the steel arch recognition model in the third part. The lining thickness recognition model in the second part extracts features from the lining geological radar image and recognizes the lining thickness based on the extracted features based on the mileage coordinate information. The steel arch recognition model in the third part extracts features from the lining geological radar data to realize automatic recognition of steel bars, count the number of steel bars, and extract steel arches from them. It locates the steel bar heads according to the mileage coordinate information and automatically calculates the spacing between steel arches.

[0071] By using the trained GAR-Net network model to identify geological radar images, the lining thickness and the positioning information of the tunnel steel arch frame can be obtained. The lining thickness and the positioning information of the tunnel steel arch frame can be used to determine whether there are defects in the tunnel lining steel frame. This overcomes the shortcomings of traditional radar image interpretation and can realize automatic recognition of the tunnel steel arch frame and lining thickness, significantly increasing the accuracy and recognition efficiency of automatic geological radar image recognition.

[0072] Specifically, the lining geological radar dataset is generated mainly by means of spatial domain enhancement, such as image rotation, cropping, grayscale conversion, brightness adjustment, etc. In the embodiment of the present invention, the thickness dataset is carefully manually labeled using the LabelImg software tool, and the labeled target object is the geological radar tunnel lining thickness image. The steel arch frame dataset is carefully manually labeled using the LabelImg software tool, and the labeled target object is the tunnel lining steel arch frame image. The training of the embodiment of the present invention uses 1080 measured tunnel lining thickness geological radar images. In order to avoid the overfitting effect caused by the small dataset, the image is expanded by taking spatial domain image enhancement means to expand the thickness dataset to 1 680 images, with a total target of 793,820, enable the deep learning network model to learn more comprehensive characteristics of tunnel lining steel arch frames; using 1,500 measured steel arch frame tunnel lining steel arch frame geological radar images, spatial domain image enhancement is used to expand the image capacity, and the steel arch frame data set is expanded to 2,500 images, with a total target of 58,000, enable the deep learning network model to learn more comprehensive characteristics of tunnel lining steel arch frames; this embodiment divides the data set, randomly selects 90% of the data for training, and the remaining 10% of the data is used for testing.

[0073] In an optional embodiment, the mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, including:

[0074] Performing black-and-white binary processing on the radar image in the lining geological radar data set to obtain a black-and-white binary radar image;

[0075] Sequentially performing horizontal digital scale separation and vertical digital scale separation on the black-and-white binary radar image to obtain a black-and-white binary radar image with digital scale separation;

[0076] The pattern recognition method is used to digitally recognize the black and white binary radar image with digital scale separation, and the radar image coordinate system is constructed using the digital recognition results.

[0077] It should be noted that in order to realize the automatic calculation of the arch spacing and the automatic location of the thickness defect, it is necessary to automatically identify the mileage and scale of the geological radar image. In the embodiment of the invention, the mileage recognition model is used to realize the mileage recognition. The mileage recognition process is as follows: Figure 2 As shown, the mileage recognition model is used to obtain mileage coordinate information, wherein the image is firstly processed into black and white binary, and then the image is separated into horizontal and vertical digital scales respectively, and the pattern recognition method is used to perform digital recognition in the mileage and depth directions, which provides mileage coordinate information for the positioning of the steel bar and thickness recognition results.

[0078] In an optional embodiment, the lining thickness identification model includes a contraction path and an expansion path; the contraction path includes four compression modules, and the expansion path includes four expansion modules;

[0079] Using four compression modules to convolve the radar image in the lining geological radar data set, and using four expansion modules to upsample the convolved radar image to obtain a fused feature image;

[0080] Each pixel point in the fused feature image is classified to obtain a background area and a lining thickness area.

[0081] It should be noted that the lining thickness identification model is Figure 2 As shown in the figure, the radar image needs to be processed by the lining thickness recognition model to obtain the image and thickness pixel coordinates, and the image and thickness pixel coordinates are used to distinguish the regions, and the image is divided into the background region and the lining thickness region. The lining thickness model includes two paths: the contraction path and the expansion path. For the input geological radar tunnel lining image of any size, a contraction path and an expansion path are combined with channels to extract features, and finally the pixels of the image are classified. The contraction path includes four compression modules, each of which includes a convolution layer, an activation function, and a maximum pooling layer. The convolution layer in the compression module is used to convolve the radar image, and then the expansion module is used to upsample the convolved radar image to obtain a fused feature image, so as to learn the deeper features and the shallower features in the radar image at the same time. Finally, each pixel in the fused feature image is classified, and the pixels are divided into pixels in the background region and pixels in the lining thickness region, thereby realizing the automatic recognition of the tunnel lining thickness.

[0082] In an optional embodiment, the four compression modules include a first compression module, a second compression module, a third compression module and a fourth compression module connected in sequence; the convolution layer of the first compression module adopts two 64-channel 3x3 convolution kernels, the convolution layer of the second compression module adopts two 128-channel 3x3 convolution kernels, the convolution layer of the third compression module adopts three 256-channel 3x3 convolution kernels, and the convolution layer of the fourth compression module adopts three 512-channel 3x3 convolution kernels;

[0083] The four expansion modules include a first expansion module corresponding to the first compression module, a second expansion module corresponding to the second compression module, a third expansion module corresponding to the third compression module, and a fourth expansion module corresponding to the fourth compression module;

[0084] Among them, the number of channels of the corresponding compression module and expansion module remains consistent.

[0085] It should be noted that if Figure 2 As shown in the figure, each compression module is composed of a series of network structures, including convolutional layers, activation functions and maximum pooling layers. The first two convolutional layers use two 64-channel 3x3 convolutional kernels and two 128-channel 3x3 convolutional kernels respectively, and the last two use three 256-channel 3x3 convolutional kernels and three 512-channel 3x3 convolutional kernels respectively. When constructing the expansion path, an upsampling module with a step size of 2 is used. The expansion path and the contraction path need to maintain consistency in the number of channels, and finally achieve effective fusion of features, so that the lining thickness recognition model can learn the features of deeper and shallower parts at the same time.

[0086] Furthermore, the fourth compression module is connected to an underlying module. The convolution layer of the underlying module uses three 512-channel 3x3 convolution kernels. After convolution of the underlying module, the convolved data is transmitted to the fourth expansion module. The fourth expansion module simultaneously receives data from the fourth contraction module and the underlying module for processing.

[0087] Specifically, in this embodiment, the lining thickness recognition model mainly uses VGG for backbone feature extraction. During the training process, the parameters of the model are adjusted by setting two different stages. In the initial stage, the number of iterations is set to 0 to 50 times, and the data batch size for each processing is 4; then, in the thawing stage, the number of iterations is increased to 51 to 100 times, and the data batch size is selected to be 2. The optimizer uses Adam, and its parameters are set as follows: the momentum is set to 0.9, the weight offset is set to 0, and the initial learning rate is set to 10 -4 , the minimum learning rate is set to 10 -6, where the learning rate descent method uses cosine annealing, the loss function uses Dice Loss, and the adaptive learning rate descent method uses LearningRateScheduler. The training loss diagram of the lining thickness recognition model is shown in Figure 3 As shown in the figure, in this lining thickness recognition model, the training loss and the validation loss decrease synchronously, indicating that the network learning effect is good and it is an ideal learning state. There is a large starting point for both the training loss and the validation loss in the 50th round. This is because this training adopts transfer learning. Part of the network is frozen in the first 50 rounds, and the frozen part of the network is unfrozen in the last 50 rounds. Therefore, the training of the new network starts in the 50th round. Since the training of the new network has a process of relearning the parameters, the training loss increases at this time. After 100 rounds of training cycles, the loss function values ​​of the training set and the validation set are almost consistent, and the specific value is shown as 0.015.

[0088] After each training unlock, all the validation set data is used to calculate the intersection and union ratio. Figure 4 As shown in the figure, it can be seen that when the training round is 10, the mIOU has reached 96.8%. In the subsequent training, the mIOU tends to be stable. When the training round is 100, the final mIOU is 97.1%, indicating that the prediction effect of the semantic segmentation algorithm is good. Figure 5 is the average intersection-over-union ratio of the final lining thickness recognition model. The network is divided into two categories, one is the background class (background), and the IOU of the background is 98.5%; the other is the thickness class (TK1), and the IOU of the thickness is 95.7%. The mIOU of the two categories is 97.1%. Since the recognition accuracy is 100% when the mIOU is 1, the mIOU is 97.1%, which means that the recognition accuracy reaches 97.1%.

[0089] The trained network model is used to perform semantic segmentation and thickness recognition on the geological radar tunnel lining image. The recognition effect is as follows: Figure 6 As shown in the figure, the green part is the recognition result of the lining thickness by the semantic segmentation network. It can be seen that the lower edge of the green part almost coincides with the junction of the arch frame and the concrete, indicating that the accuracy of lining thickness recognition is relatively high in practical application.

[0090] In an optional embodiment, the steel arch recognition model is used to locate the steel arch based on the mileage coordinate information, including:

[0091] Extracting features from the radar image in the lining geological radar data set to obtain a steel bar feature image;

[0092] Extracting the contour of the steel bar feature image to obtain a steel bar structure;

[0093] Counting the number of the steel bar structure to obtain the number of steel bars;

[0094] Using the steel bar structure and the number of steel bars to identify the steel arch frame, and obtain the steel arch frame structure;

[0095] The steel arch structure is analyzed using the mileage coordinate information to obtain the position information of the steel bar head and the spacing information between the steel arches.

[0096] It should be noted that if Figure 2 As shown, a convolutional neural network is used to extract features from radar images in the lining geological radar data set. In this embodiment, the backbone feature extraction network uses Darknet-53 to extract features from radar images, and then uses a feature pyramid network (FPN) to perform multi-scale fusion on the extracted features, and finally generates three sets of feature images. The three sets of feature images respectively realize the recognition of larger objects, medium objects and small objects, and the scales of their feature images are width: 52, height: 52, dimension: 255, width: 26, height: 26, dimension: 255, width: 13, height: 13, dimension: 255. Among them, a separate logistic regression algorithm is also used in the backbone feature extraction network for data classification.

[0097] Furthermore, a residual module is added to the backbone feature extraction network to realize the long-range link of image feature extraction, which can successfully avoid the gradient vanishing problem and significantly enhance the efficiency of image feature extraction.

[0098] The steel bar feature image is obtained through the backbone feature extraction network, and the contour of the steel bar feature image is recognized to obtain the contour information of the steel bar, and the contour information is counted to determine the number of steel bars. The structure of the steel arch can be determined by using the contour information of the steel bar structure and the number of steel bars. In this embodiment, the steel arch is extracted and identified from the contour information of the steel bar structure and the number of steel bars using traditional image processing methods.

[0099] The mileage coordinate information is used to locate the steel bar head of the steel arch structure and calculate the distance between the steel arches, thereby completing the steel arch identification.

[0100] Specifically, in this embodiment, the training loss diagram of the backbone feature extraction network in the steel arch recognition model is as follows: Figure 7As shown in the figure, a total of 1200 rounds of training were conducted. The training loss in the first 600 rounds was greater than the validation loss, and the validation loss in the last 600 rounds was greater than the validation loss. During the network model training phase, it was observed that the training loss value finally stabilized at 0.1, while the validation loss value finally stabilized at 0.12. There was a large starting point for both the training loss and the validation loss in the 200th round. This was because the training process used transfer weights, and the backbone feature extraction network was frozen in the first 200 rounds, and the frozen part of the network was unfrozen in the last 200 rounds. Therefore, the training of the new network started in the 200th round, and the training loss suddenly increased at this time. Although the final validation loss was greater than the training loss, both the training loss and the validation loss were steadily decreasing during the training process, indicating that the network has been in a learning state and the learning and training effect is good.

[0101] The precision-recall (PR) graph during training is as follows: Figure 8 As shown in the figure, the area enclosed by the ordinate P and the abscissa R is the AP value. The recognition effect is best when the accuracy is large and the recall rate is large. AP can be understood as the recognition accuracy in the true sense. When the AP value is 1, the recognition accuracy is 100%. The AP of this training is 96.09%, so the recognition accuracy is 96.09%.

[0102] The trained steel arch recognition model is used to identify the steel arch in the geological radar tunnel lining image. The recognition effect is as follows: Fig. 9 As shown, the green part is the recognition result of the semantic segmentation network for the lining thickness, and the red frame is the recognition result of the steel arch frame. In this figure, a total of 11 steel arch frames are all recognized, and the recognition effect is good.

[0103] After the steel arch recognition is completed, the outline of the steel arch is extracted in sequence using traditional image processing algorithms such as binarization and contour extraction. Some of the extraction results are shown in the figure below. Fig.10 As shown in the figure, the green part is the extracted arch frame outline. The pixel point at the vertex of the steel arch frame outline is extracted. This point is the precise position point of the steel arch frame, such as Fig.10 As shown in the red dots in the figure. Thus, the contour extraction of the steel arch frame and the position identification of the arch frame vertex are completed. Finally, the final results of the lining thickness identification and the steel arch frame spacing identification are shown in the figure below. Fig.11 shown.

[0104] The above method finally completed the automatic positioning and identification of the lining thickness and thickness defect positions. The automatic identification output results are shown in Table 1. After manual review and hole verification, it was found that the lining thickness in the mileage sections of 184.6-185.4 and 199.6-210.7 was insufficient and did not meet the design requirements, which was consistent with the identification results.

[0105] Table 1

[0106]

[0107] The GAR-Net network also completed the automatic positioning and identification of steel arches and arch spacing. The automatic identification results were compared with the results of manual identification, and the results are shown in Table 2. From Table 2, it can be concluded that the average arch spacing of manually identified arches is 1.685m, the average arch spacing of automatically identified arches is 1.695m, and the identification accuracy is 96.1%.

[0108] Table 2

[0109]

[0110] Example 2

[0111] Embodiment 2 of the present invention provides a tunnel lining steel frame recognition system based on GAR-Net, comprising:

[0112] A data set processing module, the data set processing module is used to obtain lining geological radar data, perform spatial domain enhancement on the geological radar data, and generate a lining geological radar data set;

[0113] A training module, wherein the training module is used to construct a GAR-Net network model, and the GAR-Net network model is trained using the lining geological radar data set; wherein the GAR-Net network model includes a mileage recognition model, a lining thickness recognition model, and a steel arch recognition model, wherein the mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, the lining thickness recognition model is used to recognize the tunnel lining thickness, and the steel arch recognition model is used to locate the steel arch based on the mileage coordinate information;

[0114] The recognition module is used to obtain geological radar images and use the trained GAR-Net network model to automatically locate and recognize the lining thickness and tunnel steel arch frame of the geological radar images.

[0115] In an optional embodiment, the training module includes:

[0116] A binarization processing unit, the binarization processing unit is used to perform black-and-white binarization processing on the radar image in the lining geological radar data set to obtain a black-and-white binarized radar image;

[0117] A scale separation unit, the scale separation unit is used to sequentially perform horizontal digital scale separation and vertical digital scale separation on the black-and-white binary radar image to obtain a black-and-white binary radar image with digital scale separation;

[0118] The pattern recognition unit is used to digitally recognize the black-and-white binary radar image with digital scale separation by using a pattern recognition method, and to construct a radar image coordinate system by using the digital recognition result.

[0119] In an optional embodiment, the training module further includes:

[0120] A steel bar feature extraction unit, the steel bar feature extraction unit is used to extract features from the radar image in the lining geological radar data set to obtain a steel bar feature image;

[0121] A contour extraction unit, the contour extraction unit is used to extract the contour of the steel bar feature image to obtain a steel bar structure;

[0122] A quantity counting unit, the quantity counting unit is used to count the quantity of the steel bar structure to obtain the quantity of steel bars;

[0123] A steel arch identification unit, wherein the steel arch identification unit is used to identify the steel arch using the steel bar structure and the number of steel bars to obtain a steel arch structure;

[0124] A positioning unit is used to analyze the steel arch structure using the mileage coordinate information to obtain the position information of the steel bar head and the spacing information between the steel arches.

[0125] Example 3

[0126] Fig.12 A schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention is shown in FIG. Fig.12 As shown, the electronic device includes a processor 21, a memory 22, an input device 23 and an output device 24; the number of processors 21 in the computer device can be one or more. Fig.12 A processor 21 is taken as an example; the processor 21, the memory 22, the input device 23 and the output device 24 in the electronic device can be connected by a bus or other means. Fig.12 The example of connecting through bus is taken in the following.

[0127] The memory 22 is a computer-readable storage medium that can be used to store software programs, computer executable programs and modules. The processor 21 executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the memory 22, that is, the tunnel lining steel frame identification method based on GAR-NET of Example 1 is implemented.

[0128] The memory 22 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory 22 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 22 may further include a memory remotely arranged relative to the processor 21, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0129] The input device 23 can be used to receive the ID and password input by the user, etc. The output device 24 is used to output the network configuration page.

[0130] Example 4

[0131] Embodiment 4 of the present invention further provides a computer-readable storage medium, wherein the computer-executable instructions, when executed by a computer processor, are used to implement the tunnel lining steel frame identification method based on GAR-NET as provided in Embodiment 1.

[0132] An embodiment of the present invention provides a storage medium containing computer executable instructions, and its computer executable instructions are not limited to the method operations provided in Example 1, and can also execute related operations in the tunnel lining steel frame identification method based on GAR-NET provided in any embodiment of the present invention.

[0133] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. The tunnel lining steel frame identification method based on GAR-NET is characterized by: The steps include: Acquire lining geological radar data, perform spatial domain enhancement on the geological radar data, and generate a lining geological radar data set; Constructing a GAR-Net network model, and using the lining geological radar data set to train the GAR-Net network model; wherein the GAR-Net network model includes a mileage recognition model, a lining thickness recognition model and a steel arch recognition model, wherein the mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, the lining thickness recognition model is used to identify the tunnel lining thickness, and the steel arch recognition model is used to locate the steel arch based on the mileage coordinate information; wherein the steel arch positioning includes: using the mileage coordinate information to analyze the steel arch structure to obtain the steel bar head position information and the spacing information between the steel arches; Acquire a geological radar image, and use the trained GAR-Net network model to automatically locate and identify the lining thickness and the tunnel steel arch frame of the geological radar image; The mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, including: Performing black-and-white binary processing on the radar image in the lining geological radar data set to obtain a black-and-white binary radar image; Sequentially performing horizontal digital scale separation and vertical digital scale separation on the black-and-white binary radar image to obtain a black-and-white binary radar image with digital scale separation; The pattern recognition method is used to digitally recognize the black and white binary radar image with digital scale separation, and the radar image coordinate system is constructed using the digital recognition results.

2. The tunnel lining steel frame identification method based on GAR-NET according to claim 1 is characterized in that: The lining thickness identification model includes a contraction path and an expansion path; the contraction path includes four compression modules, and the expansion path includes four expansion modules; Using four compression modules to convolve the radar image in the lining geological radar data set, and using four expansion modules to upsample the convolved radar image to obtain a fused feature image; Each pixel point in the fused feature image is classified to obtain a background area and a lining thickness area.

3. The tunnel lining steel frame identification method based on GAR-NET according to claim 2 is characterized in that: The four compression modules include a first compression module, a second compression module, a third compression module and a fourth compression module connected in sequence; the convolution layer of the first compression module adopts two 64-channel 3x3 convolution kernels, the convolution layer of the second compression module adopts two 128-channel 3x3 convolution kernels, the convolution layer of the third compression module adopts three 256-channel 3x3 convolution kernels, and the convolution layer of the fourth compression module adopts three 512-channel 3x3 convolution kernels; The four expansion modules include a first expansion module corresponding to the first compression module, a second expansion module corresponding to the second compression module, a third expansion module corresponding to the third compression module, and a fourth expansion module corresponding to the fourth compression module; Among them, the number of channels of the corresponding compression module and expansion module remains consistent.

4. The tunnel lining steel frame identification method based on GAR-NET according to claim 1 is characterized in that: The steel arch recognition model is used to locate the steel arch based on the mileage coordinate information and includes: Extracting features from the radar image in the lining geological radar data set to obtain a steel bar feature image; Extracting the contour of the steel bar feature image to obtain a steel bar structure; Counting the number of the steel bar structure to obtain the number of steel bars; Using the steel bar structure and the number of steel bars to identify the steel arch frame, and obtain the steel arch frame structure; The steel arch structure is analyzed using the mileage coordinate information to obtain the position information of the steel bar heads and the spacing information between the steel arches.

5. The tunnel lining steel frame identification system based on GAR-NET is characterized by: include: A data set processing module, the data set processing module is used to obtain lining geological radar data, perform spatial domain enhancement on the geological radar data, and generate a lining geological radar data set; A training module, wherein the training module is used to construct a GAR-Net network model, and the GAR-Net network model is trained using the lining geological radar data set; wherein the GAR-Net network model includes a mileage recognition model, a lining thickness recognition model and a steel arch recognition model, wherein the mileage recognition model is used to provide mileage coordinate information for lining thickness recognition and steel bar head recognition, wherein the lining thickness recognition model is used to recognize the tunnel lining thickness, and wherein the steel arch recognition model is used to locate the steel arch based on the mileage coordinate information; wherein the steel arch positioning includes: analyzing the steel arch structure using the mileage coordinate information to obtain the steel bar head position information and the spacing information between the steel arches; A recognition module, which is used to obtain a geological radar image and automatically locate and recognize the lining thickness and the tunnel steel arch frame of the geological radar image using a trained GAR-Net network model; Wherein, the training module includes: A binarization processing unit, the binarization processing unit is used to perform black-and-white binarization processing on the radar image in the lining geological radar data set to obtain a black-and-white binarized radar image; A scale separation unit, the scale separation unit is used to sequentially perform horizontal digital scale separation and vertical digital scale separation on the black-and-white binary radar image to obtain a black-and-white binary radar image with digital scale separation; The pattern recognition unit is used to digitally recognize the black-and-white binary radar image with digital scale separation by using a pattern recognition method, and to construct a radar image coordinate system by using the digital recognition result.

6. The tunnel lining steel frame identification system based on GAR-NET according to claim 5 is characterized in that: The training module also includes: A steel bar feature extraction unit, the steel bar feature extraction unit is used to extract features from the radar image in the lining geological radar data set to obtain a steel bar feature image; A contour extraction unit, the contour extraction unit is used to extract the contour of the steel bar feature image to obtain a steel bar structure; A quantity counting unit, the quantity counting unit is used to count the quantity of the steel bar structure to obtain the quantity of steel bars; A steel arch identification unit, wherein the steel arch identification unit is used to identify the steel arch using the steel bar structure and the number of steel bars to obtain a steel arch structure; A positioning unit is used to analyze the steel arch structure using the mileage coordinate information to obtain the position information of the steel bar head and the spacing information between the steel arches.

7. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the tunnel lining steel frame identification method based on GAR-NET as claimed in any one of claims 1 to 4 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, a tunnel lining steel frame identification method based on GAR-NET as described in any one of claims 1 to 4 is implemented.

Images