Defect identification model training method, tunnel lining quality defect identification method and product

Abstract

The invention discloses a defect identification model training method, a tunnel lining quality defect identification method and a product, and relates to the technical field of data enhancement, and the method comprises the steps: obtaining scanning image data of a tunnel construction site; performing data enhancement on the scanned image data to obtain enhanced image data; constructing a simulation image of the scanned image data by using a finite difference time domain-based method to obtain a simulation numerical model; generating an image with the same standard as the scanned image data by using a progressive generative adversarial network to obtain generative adversarial data; mixing the enhanced image data, the simulation numerical model, the generative adversarial data and the scanned image data to obtain a mixed data set; and training a target detection network by using the mixed data set to obtain a defect identification model. According to the method, the data set is promoted by using three different data enhancement methods, so that more synthetic defect samples are generated, the training set is enriched, and the generalization ability of the model is improved.

Description

Technical Field

[0001] The present application relates to the field of data enhancement technology, and in particular to a defect recognition model training, tunnel lining quality defect recognition method and product. Background Art

[0002] In modern construction projects, the safety and durability of bridges and tunnels have always been important considerations in design and construction. In this process, the quality of concrete structures directly affects the overall safety and economy of the project. For example, during the construction of tunnels, defects in concrete linings, such as cracks, spalling and voids, are often caused by improper operations during construction, material quality problems, and environmental factors. These problems not only lead to reduced strength of the structure, but may even cause serious safety hazards and endanger the normal operation of the tunnel. Therefore, timely and effective detection and identification of defects in the lining is crucial to maintaining structural safety.

[0003] At present, the detection methods of concrete structures mainly include acoustic detection, infrared thermal imaging and non-destructive testing (NDT) technology. These technologies have achieved certain results in practical applications, but there are still some technical problems. For example, the acoustic detection method has shortcomings in detection accuracy and environmental influence, while infrared thermal imaging is easily disturbed by surface temperature fluctuations. In addition, traditional non-destructive testing technology often requires professional technicians to carry out complex detection processes and data analysis. Its efficiency and accuracy are heavily dependent on the operator's experience. When faced with a large amount of data, manual interpretation is time-consuming and difficult to ensure accuracy.

[0004] In order to improve the efficiency and accuracy of concrete structure inspection, deep learning technology has been gradually introduced into the automatic detection of structural defects in recent years. Studies have shown that the effectiveness of convolutional neural networks (CNNs) in image processing and pattern recognition has been widely recognized, and they can identify defects in images by automatically extracting features. However, since concrete inspection datasets are often small-scale and unbalanced, the training algorithm is prone to overfitting during the feature learning process. Summary of the invention

[0005] The purpose of this application is to provide a defect recognition model training, tunnel lining quality defect recognition method and product, which can generate more synthetic defect samples, enrich the training set, and improve the generalization ability of the model.

[0006] To achieve the above objectives, this application provides the following solutions:

[0007] In a first aspect, the present application provides a defect recognition model training method, comprising:

[0008] Obtaining scanned image data of the tunnel construction site;

[0009] Performing data enhancement on the scanned image data to obtain enhanced image data;

[0010] constructing a simulated image of the scanned image data using a finite difference time domain method to obtain a simulated numerical model;

[0011] Generate an image of the same standard as the scanned image data using a progressive generative adversarial network to obtain generative adversarial data;

[0012] Mixing the enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data to obtain a mixed data set;

[0013] The target detection network is trained using the mixed data set to obtain a defect recognition model; the backbone layer of the target detection network is a VanillaNet backbone network; the neck layer of the target detection network includes: a Deformable-LKA attention mechanism.

[0014] Optionally, the step of using a progressive generative adversarial network to generate an image of the same standard as the scanned image data to obtain generative adversarial data specifically includes:

[0015] Use scanned image data to train a progressive generative adversarial network to obtain an image generation model;

[0016] The image generation model is used to generate an image with the same standard as the scanned image data to obtain generative adversarial data.

[0017] Optionally, performing data enhancement on the scanned image data to obtain enhanced image data specifically includes:

[0018] The image data is rotated, translated, scaled, cropped and noise-added to obtain enhanced image data.

[0019] Optionally, the acquiring of scanning image data of the tunnel construction site specifically includes:

[0020] Obtain geological radar image data of the tunnel construction site;

[0021] The geological radar image data is subjected to zero-time correction, DC offset removal, energy gain, one-dimensional bandpass filtering and two-dimensional spatial filtering operations to obtain scanning image data of the tunnel construction site.

[0022] Optionally, mixing the enhanced image data, the simulated numerical model, the generative adversarial data and the scanned image data to obtain a mixed data set specifically includes:

[0023] Performing tunnel defect feature annotation on the enhanced image data to obtain enhanced annotation data;

[0024] Perform tunnel defect feature annotation on the simulated numerical model to obtain simulated annotation data;

[0025] Annotating the generated adversarial data with tunnel defect features to obtain adversarial annotated data;

[0026] Annotating the scanned image data with tunnel defect features to obtain original annotated data;

[0027] The enhanced labeled data, the simulated labeled data, the adversarial labeled data and the original labeled data are mixed to obtain a mixed data set.

[0028] Optionally, the enhanced image data, the simulated numerical model, the generative adversarial data and the scanned image data are mixed to obtain a mixed data set; and a target detection network is trained using the mixed data set to obtain a defect recognition model, specifically including:

[0029] Mixing the enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data in different proportions to obtain a plurality of mixed data sets;

[0030] Inputting each group of the mixed data sets into the target detection network respectively to obtain the recognition accuracy corresponding to each group of the mixed data sets;

[0031] Select the ratio corresponding to the mixed data set with the highest recognition accuracy as the training ratio;

[0032] The enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data are mixed according to the training ratio to obtain a mixed data set; and the target detection network is trained using the mixed data set to obtain a defect recognition model.

[0033] In a second aspect, the present application provides a method for identifying quality defects of a tunnel lining, comprising:

[0034] Acquire image data to be identified at the tunnel construction site;

[0035] The image data to be identified is input into a defect recognition model to obtain a defect recognition result; the defect recognition model is trained using any of the defect recognition model training methods described above.

[0036] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-mentioned defect recognition model training methods or tunnel lining quality defect recognition methods.

[0037] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned defect recognition model training methods or tunnel lining quality defect recognition methods.

[0038] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned defect recognition model training methods or tunnel lining quality defect recognition methods.

[0039] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0040] The present application provides a defect recognition model training method, which obtains scanned image data of a tunnel construction site; performs data enhancement on the scanned image data to obtain enhanced image data; uses a finite difference time domain method to construct a simulated image of the scanned image data to obtain a simulated numerical model; uses a progressive generative adversarial network to generate an image with the same standard as the scanned image data to obtain generative adversarial data; then mixes the enhanced image data, the simulated numerical model, the generative adversarial data and the scanned image data to obtain a mixed data set; uses the mixed data set to train a target detection network to obtain a defect recognition model; the backbone layer of the target detection network is the VanillaNet backbone network; the neck layer of the target detection network includes: Deformable-LKA attention mechanism. The present application uses 3 different data enhancement methods to enhance the data set, thereby generating more synthetic defect samples, enriching the training set, and improving the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present application 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 application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 This is an application environment diagram of a defect recognition model training method or a tunnel lining quality defect recognition method in an embodiment of the present application;

[0043] Figure 2 A flowchart of a defect recognition model training method provided in one embodiment of the present application;

[0044] Figure 3 A schematic diagram of a simulation numerical model provided in an embodiment of the present application;

[0045] Figure 4A schematic diagram of geological radar image data provided in an embodiment of the present application;

[0046] Figure 5 A schematic diagram of the YOLOv8-LQD network structure provided in an embodiment of the present application;

[0047] Figure 6 A schematic diagram of MAP and Loss-epoch curves of the YOLOv8-LQD model provided in an embodiment of the present application;

[0048] Figure 7 A schematic diagram of a detailed step flow of a defect recognition model training method provided in one embodiment of the present application;

[0049] Figure 8 An overall reference schematic diagram of a defect recognition model training method provided in one embodiment of the present application;

[0050] Fig. 9 A schematic diagram of a flow chart of a method for identifying quality defects of a tunnel lining provided in another embodiment of the present application;

[0051] Fig.10 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] This application designs a deep learning network specifically for abnormal state recognition of tunnel lining radar images, and proposes a method for implementing an enhanced data set. Its innovation is reflected in: ① In terms of intelligent recognition model, the YOLOv8 detection algorithm is used as the original model of intelligent recognition, and the VanillaNet backbone network and Deformable-LKA attention mechanism are introduced to improve the original YOLOv8 network and construct the YOLOv8-LQD detection algorithm. ② In terms of data enhancement, three data enhancement strategies, namely traditional data enhancement, finite difference time domain (FDTD) simulation and generative adversarial network (ProGAN), are used to enrich the data set, and the synthetic data generated by these three data enhancement strategies are combined with the original data to create a comprehensive enhanced data set.

[0054] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0055] The defect recognition model training method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the scanned image data to be processed to the server 104. After receiving the scanned image data, the server 104 performs data enhancement on the scanned image data to obtain enhanced image data; a simulated image of the scanned image data is constructed based on the finite difference time domain method to obtain a simulated numerical model; an image with the same standard as the scanned image data is generated using a progressive generative adversarial network to obtain generative adversarial data; the enhanced image data, the simulated numerical model, the generative adversarial data and the scanned image data are mixed to obtain a mixed data set; the target detection network is trained using the mixed data set to obtain a defect recognition model; the backbone layer of the target detection network is the VanillaNet backbone network; the neck layer of the target detection network includes: Deformable-LKA attention mechanism. The server 104 can feed back the obtained defect recognition model to the terminal 102. In addition, in some embodiments, the defect recognition model training method can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly process the scanned image data to be processed, or the server 104 can obtain the scanned image data to be processed from the data storage system and process it.

[0056] The tunnel lining quality defect identification method provided in the embodiment of the present application can also be applied to Figure 1 In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set up separately, or it can be integrated on the server 104, or it can be placed on the cloud or other servers. The terminal 102 can send the image data to be identified to the server 104. After the server 104 receives the image data to be identified, it inputs the image data to be identified into the defect recognition model to obtain a defect recognition result. The server 104 can feedback the obtained defect recognition result to the terminal 102. In addition, in some embodiments, the tunnel lining quality defect identification method can be implemented by the server 104 or the terminal 102 alone. For example, the terminal 102 can directly process the image data to be identified, or the server 104 can obtain the image data to be identified from the data storage system and process it.

[0057] The terminal 102 may be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart vehicle-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers, or may be a cloud server.

[0058] In an exemplary embodiment, Figure 2 As shown, a defect recognition model training method is provided. The method is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 1 The server 104 in the example is used as an example to illustrate the method, which includes the following steps S1 to S6. Among them:

[0059] S1. Obtaining scanning image data of a tunnel construction site.

[0060] First, geological radar image data of the tunnel construction site is obtained; then, zero-time correction, DC offset removal, energy gain, one-dimensional bandpass filtering and two-dimensional spatial filtering operations are performed on the geological radar image data to obtain scanning image data of the tunnel construction site.

[0061] In this embodiment, a MALA system equipped with an 800MHz shielded antenna is used to manually collect real GPR (Ground Penetrating Radar) image data at the tunnel construction site. The survey lines are arranged along the longitude direction of the tunnel, located at the vault, left and right shoulders, and left and right walls (the vault, left and right spandrels, and left and right side walls of the inner circle of the tunnel lining). A GPR trace (A scan) is collected every 0.02 meters, and the acquisition time window is set to 42 nanoseconds. Multiple GPR traces are combined into a grayscale GPR image (B scan), that is, a ground penetrating radar image data (GPR image).

[0062] The geological radar image data is then processed to suppress noise and highlight target reflections. The key steps include zero-time correction, removal of direct current (DC) offset, energy gain, one-dimensional bandpass filtering, and two-dimensional spatial filtering. In the zero-time correction, the start time of the GPR profile is adjusted to the lining surface to provide accurate positioning information. Removal of DC offset eliminates background noise and highlights effective signals by subtracting the average value of each trace. In addition, an energy gain with a scaling factor of 0.32 is performed to compensate for energy attenuation and thus enhance signal strength. A one-dimensional bandpass filter from 200 to 900 MHz is applied to filter noise and retain information within the target frequency range. Next, a 3×3 two-dimensional spatial filter is performed to smooth the GPR image, reduce image noise and improve readability. Finally, the time window is cropped to 15 nanoseconds to optimize data storage and processing efficiency and retain important information.

[0063] S2. Perform data enhancement on the scanned image data to obtain enhanced image data.

[0064] In this embodiment, the processed scanned image data is subjected to traditional data enhancement, FDTD (Finite-Difference Time-Domain Method) numerical simulation synthesis and progressive generative adversarial network data synthesis respectively.

[0065] Among them, traditional data enhancement is the data enhancement described in step S2. Traditional data enhancement uses a series of operations such as rotation, translation, scaling, cropping and adding noise to the scanned image data to obtain enhanced image data.

[0066] S3. Constructing a simulated image of the scanned image data using a finite difference time domain method to obtain a simulated numerical model.

[0067] FDTD numerical simulation synthesis refers to the use of the gprMax program based on the finite difference time domain (FDTD) method to build a numerical model for simulating tunnel lining geological radar detection. The model consists of an air layer, a lining layer and a surrounding rock layer. The I-beam ribs are located in the initial lining, the double-row steel bars are located in the secondary lining, and irregular holes are scattered on the interface between the initial lining and the rock.

[0068] The construction process mainly involves writing Python scripts. The main functions of the scripts include generating input files and calling the gprMAX program to generate GPR simulation images by running the gprMAX program. The written input files include defining the model domain, determining the medium properties, setting the source and receiver, creating a grid, and building the geometry.

[0069] The model constructed is a tunnel lining model. The geometry of the model is as follows: Figure 3 The obtained simulation numerical model is shown in Figure 4It should be noted that the model was constructed using a script, which controlled the location of the holes and the randomness of the lining thickness in the model. Figure 4 It is the unprocessed geological radar image data.

[0070] The model materials use built-in air and pec to represent voids and steel ribs. The relative dielectric constant εr of the lining is set to 9, and the electrical conductivity σ is 1e-5S / m; the relative dielectric constant εr of the surrounding rock is 6, and the electrical conductivity σ is 1e-3S / m.

[0071] S4. Generate an image with the same standard as the scanned image data using a progressive generative adversarial network to obtain generative adversarial data.

[0072] First, a progressive generative adversarial network is trained using scanned image data to obtain an image generation model; then, the image generation model is used to generate an image with the same standard as the scanned image data to obtain generative adversarial data.

[0073] In this embodiment, the progressive generative adversarial network (ProGAN) data synthesis refers to using scanned image data to train the ProGAN network, where the generator starts to generate images from the lowest resolution, and as the training progresses, the network layers are gradually added to generate higher resolution images until the discriminator cannot distinguish between the high-resolution synthesized images and the images in the training data set. Finally, a picture similar to the scanned image data is generated to obtain generative adversarial data.

[0074] It is worth noting that the execution order of steps S2, S3 and S4 in this embodiment is not fixed. All three steps will generate similar GPR images, which will be mixed with the scanned image data collected on site. By comparing the MAP values ​​of the data sets with different mixing ratios in the improved YOLOv8 model (YOLOv8-LQD), an optimal ratio is determined. Finally, the unknown GPR image is recognized according to the weight file of this optimal ratio.

[0075] S5. Mix the enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data to obtain a mixed data set.

[0076] First, the enhanced image data is annotated with tunnel defect features to obtain enhanced annotated data; the simulated numerical model is annotated with tunnel defect features to obtain simulated annotated data; the generated adversarial data is annotated with tunnel defect features to obtain adversarial annotated data; the scanned image data is annotated with tunnel defect features to obtain original annotated data; then, the enhanced annotated data, the simulated annotated data, the adversarial annotated data and the original annotated data are mixed to obtain a mixed data set.

[0077] In this example, the scanning image data and the GPR images after three data enhancements (enhanced image data, simulated numerical model, and generated adversarial data) are mixed to form a mixed data set. The mixed radar image is opened using LabelImg software under Windows 10 system. After finding the tunnel defect features in the image, the marking box in the software is used to select the corresponding position and indicate the defect category. The defect features include steel deficiency, voids, and thin lining.

[0078] S6. Use the mixed data set to train the target detection network to obtain a defect recognition model; the backbone layer of the target detection network is the VanillaNet backbone network; the neck layer of the target detection network includes: Deformable-LKA attention mechanism.

[0079] In this embodiment, the YOLOv8-LQD detection algorithm is first constructed.

[0080] The YOLOv8 architecture is mainly composed of Backbone, Neck, Head and Loss. Backbone enhances the learning ability of convolutional neural networks by introducing the CSP (Cross Stage Partial) module. The CSP module reduces the amount of calculation by segmenting the feature map while maintaining feature transfer. The Neck part uses the SPPF (Spatial Pyramid Pooling Fusion) module, combined with the FPN (Feature Pyramid Network) and PAN (Path Aggregation Network) structures to improve the inference speed of the model. The Head part adopts a decoupled design, separating the detection head from the classification head, and turning to an anchor-free method to improve the accuracy of target detection.

[0081] In order to cope with the complex scenes and low signal-to-noise ratio in tunnel lining detection, YOLOv8 has been improved to form the YOLOv8-LQD framework. Figure 5 ,The main improvements of this framework are at two levels: First, the backbone layer introduces the ,VanillaNet module, which only contains basic convolution and pooling layers. This design reduces the ,amount of computation and parameters, making it suitable for terminal devices with ,limited computing resources. The structure of VanillaNet simplifies the network and ,improves the perception of complex feature information.

[0082] Secondly, a Deformable-LKA (Deformable Local KeypointsAttention) attention mechanism is added to the neck layer, which can focus on important features and suppress complex background interference, thereby improving the overall performance of the model and data processing efficiency. With this improvement, YOLOv8-LQD is not only lighter, but also has significantly improved detection accuracy. The ability to dynamically adjust the position of the convolution kernel enhances the model's adaptability to complex geometric transformations and shape variations, making it more accurate and effective in industrial applications such as lining quality inspection. Figure 6 shown.

[0083] As an optional implementation, the training process in this embodiment may include the following steps:

[0084] Step 1: Input images: Input images from the training dataset into the model. The images are stored in the specified JPEG format and come with corresponding label files.

[0085] Step 2: Image preprocessing: Resize: Resize the input image to the size required by the model (e.g., 640x640 pixels) to maintain consistent input dimensions.

[0086] Normalization: Convert pixel values ​​to floating point numbers in the range [0, 1] to reduce numerical fluctuations during model training.

[0087] Step 3: Feature extraction: Pass the preprocessed input image to the backbone network, which is an important component of YOLOv8 and is responsible for extracting key features in the image. YOLOv8 supports multiple backbone network configurations.

[0088] Step 4: Feature Pyramid Construction: The feature map is processed by the Feature Pyramid Network (FPN) to aggregate features at different scales. This step is to capture the features of small and large objects at the same time and enhance multi-scale detection capabilities.

[0089] Step 5: Bounding box prediction: After being processed by the feature pyramid, the model generates multiple feature maps. These feature maps are used to generate candidate bounding boxes, and class prediction and box regression (position adjustment) are performed on each candidate box.

[0090] Step 6: Target classification and box regression: For each anchor box, the model calculates: category confidence and bounding box coordinates. Category confidence: the probability that an object belongs to a certain category; Bounding box coordinates: adjust the object position according to the anchor box, including the regression of the center coordinates and size.

[0091] Step 7: NMS (Non-Maximum Suppression): In order to remove duplicate detection results, the non-maximum suppression (NMS) algorithm is applied. By setting the confidence threshold, NMS will select the detection box with the highest confidence and remove other boxes that overlap too much with it. This ensures that the final output is the most accurate detection result.

[0092] Step 8: Output results: Output the final detection results, including the category, confidence score, and bounding box coordinates of each object.

[0093] In addition, this embodiment also provides a method for selecting a training ratio:

[0094] Mixing the enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data in different proportions to obtain a plurality of mixed data sets;

[0095] Inputting each group of the mixed data sets into the target detection network respectively to obtain the recognition accuracy corresponding to each group of the mixed data sets;

[0096] Select the ratio corresponding to the mixed data set with the highest recognition accuracy as the training ratio;

[0097] The enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data are mixed according to the training ratio to obtain a mixed data set; and the target detection network is trained using the mixed data set to obtain a defect recognition model.

[0098] In this embodiment, three kinds of data-enhanced GPR image data are added to the real GPR image (scanned image data) in different proportions (10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100%) to form 30 different data set combinations. The hyperparameters are kept unchanged during training, and the training is carried out on a workstation using the PyTorch1.12.1 framework, which is equipped with an NVIDIARTX 4090 graphics card. After the training is completed, the addition ratio of different data-enhanced synthetic GPR images and the corresponding recognition accuracy are obtained to obtain the optimal ratio of adding synthetic images.

[0099] The input for training is a GPR image dataset with different mixing ratios (a mixture of data collected on-site (scanned image data) and three types of data augmentation output GPR images). The output results mainly look at the model weight file and the evaluation result file. The different mixing ratios are compared based on the MAP value of the model in the evaluation result file (evaluation indicator MAP50). The higher the MAP value, the better the mixing ratio. The model weight to be used is determined based on the highest MAP value.

[0100] Add synthetic images according to the optimal ratio to establish a training dataset containing real GPR images and synthetic images for training the YOLOv8-LQD model. After the model training is completed, a weight file for optimal GPR image recognition will be generated, thereby realizing intelligent recognition of tunnel lining quality defects.

[0101] In summary, this embodiment aims to improve the accuracy and efficiency of tunnel lining defect detection and overcome the technical bottlenecks and data shortage problems in traditional methods. This embodiment proposes a new deep learning-based bridge and tunnel concrete lining defect detection framework. The framework uses three different data enhancement methods to enhance the data set, thereby generating more synthetic defect samples, enriching the training set, and improving the generalization ability of the model. At the same time, it combines the transfer learning method and pre-training on the existing high-quality natural image data set to help the model better learn key features. This method not only improves the accuracy of concrete defect detection, but also shows stronger robustness when dealing with defects in complex backgrounds. Please refer to the detailed step flow chart. Figure 7 .

[0102] Compared with the prior art, the advantages of the YOLOv8-LQD detection model algorithm provided in this embodiment are mainly reflected in improving detection efficiency and accuracy, while effectively reducing data collection costs. The overall reference schematic diagram is shown in FIG. Figure 8 shown.

[0103] First, the detection accuracy of YOLOv8-LQD has been significantly improved, reaching an average detection accuracy (MAP) of 93.4%, which is 7.4% higher than the original YOLOv8 network. This advantage is mainly due to the introduction of the VanillaNet backbone module and the C2f-DLKA attention module in the YOLOv8 network, which effectively enhances the performance of the model. By optimizing the network architecture, the feature extraction capability of the model is improved, enabling it to better identify and classify various defects in tunnel linings.

[0104] Secondly, the use of progressive generative adversarial networks (ProGANs) for data enhancement directly solves the problem of the scarcity of high-quality lining defect samples. Through simulation, a variety of tunnel lining quality defect images are generated, which significantly improves the richness and diversity of the data set. This technical point enables the detection model to be exposed to more training samples during training, thereby achieving faster model convergence speed and higher accuracy.

[0105] Finally, this method significantly reduces the cost of field data collection. Experiments show that the cost of data collection is reduced by 60% when synthetic data is fully utilized. This result is due to the strategy of synthesizing virtual data through GAN, which reduces the dependence on expensive experimental equipment and time while maintaining the quality of real data. By selecting the appropriate ratio of virtual data to real data, the model performance is further optimized without adding additional costs.

[0106] In another exemplary embodiment of the present application, Fig. 9 As shown, a method for identifying quality defects of tunnel lining is provided, comprising:

[0107] A1. Obtain image data to be identified at the tunnel construction site.

[0108] A2. Inputting the image data to be identified into a defect recognition model to obtain a defect recognition result; the defect recognition model is trained using the defect recognition model training method described above.

[0109] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig.10 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a defect recognition model training method or a tunnel lining quality defect recognition method is implemented.

[0110] Those skilled in the art will understand that Fig.10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0111] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0112] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0113] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0114] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0115] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0116] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0117] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0118] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, 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 this application.

Claims

1. A defect recognition model training method, characterized in that: The defect recognition model training method comprises: Acquire scanned image data of the tunnel construction site; Performing data enhancement on the scanned image data to obtain enhanced image data; constructing a simulated image of the scanned image data using a finite difference time domain method to obtain a simulated numerical model; Generate an image of the same standard as the scanned image data using a progressive generative adversarial network to obtain generative adversarial data; Mixing the enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data to obtain a mixed data set; The target detection network is trained using the mixed data set to obtain a defect recognition model; the backbone layer of the target detection network is a VanillaNet backbone network; the neck layer of the target detection network includes: a Deformable-LKA attention mechanism.

2. The defect recognition model training method according to claim 1, characterized in that: The step of using a progressive generative adversarial network to generate an image of the same standard as the scanned image data to obtain generative adversarial data specifically includes: Use scanned image data to train a progressive generative adversarial network to obtain an image generation model; The image generation model is used to generate an image with the same standard as the scanned image data to obtain generative adversarial data.

3. The defect recognition model training method according to claim 1, characterized in that: Performing data enhancement on the scanned image data to obtain enhanced image data specifically includes: The image data is rotated, translated, scaled, cropped and noise-added to obtain enhanced image data.

4. The defect recognition model training method according to claim 1, characterized in that: The obtaining of scanning image data of the tunnel construction site specifically includes: Obtain geological radar image data of the tunnel construction site; The geological radar image data is subjected to zero-time correction, DC offset removal, energy gain, one-dimensional bandpass filtering and two-dimensional spatial filtering operations to obtain scanning image data of the tunnel construction site.

5. The defect recognition model training method according to claim 1, characterized in that: Mixing the enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data to obtain a mixed data set specifically includes: Performing tunnel defect feature annotation on the enhanced image data to obtain enhanced annotation data; Perform tunnel defect feature annotation on the simulated numerical model to obtain simulated annotation data; Annotating the generated adversarial data with tunnel defect features to obtain adversarial annotated data; Annotating the scanned image data with tunnel defect features to obtain original annotated data; The enhanced labeled data, the simulated labeled data, the adversarial labeled data and the original labeled data are mixed to obtain a mixed data set.

6. The defect recognition model training method according to claim 1, characterized in that: The enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data are mixed to obtain a mixed data set; and the target detection network is trained using the mixed data set to obtain a defect recognition model, specifically including: Mixing the enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data in different proportions to obtain a plurality of mixed data sets; Inputting each group of the mixed data sets into the target detection network respectively to obtain the recognition accuracy corresponding to each group of the mixed data sets; Select the ratio corresponding to the mixed data set with the highest recognition accuracy as the training ratio; The enhanced image data, the simulated numerical model, the generated adversarial data and the scanned image data are mixed according to the training ratio to obtain a mixed data set; and the target detection network is trained using the mixed data set to obtain a defect recognition model.

7. A method for identifying quality defects of tunnel lining, characterized in that: The tunnel lining quality defect identification method comprises: Acquire image data to be identified at the tunnel construction site; The image data to be identified is input into a defect recognition model to obtain a defect recognition result; the defect recognition model is trained using the defect recognition model training method according to any one of claims 1 to 6.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the defect recognition model training method described in any one of claims 1 to 6 or the tunnel lining quality defect recognition method described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the defect recognition model training method described in any one of claims 1 to 6 or the tunnel lining quality defect recognition method described in claim 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the defect recognition model training method described in any one of claims 1 to 6 or the tunnel lining quality defect recognition method described in claim 7 is implemented.

Images