Improved method for detecting internal diseases of yolov9 highway tunnel lining

Abstract

The invention provides an improved method for detecting internal diseases of a yalov9 highway tunnel lining, and the process of improving a yalov9 model comprises the following steps: step 201, adding an EMA attention module in Backbone; step 202, a GELAN in the Neck layer is replaced by a GELAN-EMA, and a convolutional layer in the middle of the GELAN is replaced by an EMA attention module; step 203, frame regression is carried out by using a Focaler-SIOU loss function; and step 204, a Center Loss loss function is added. According to the invention, the deep learning uses the deep convolutional neural network to automatically extract high-dimensional features from a large amount of radar data and carry out classification or regression analysis, so that tedious steps of manually designing a feature extractor and a classifier in a traditional method can be eliminated, and an effective solution is provided for geological radar waveform analysis. And the dependence on manpower is greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel disease detection, relates to the detection of the civil engineering structure of highway tunnels, and specifically relates to an improved yolov9 method for detecting internal diseases of highway tunnel linings. Background Art

[0002] Traditional tunnel disease detection methods mostly rely on manual inspection or simple physical detection techniques. Although some surface problems can be found, it is difficult to comprehensively and accurately detect potential hazards inside the tunnel, and the efficiency is low.

[0003] Ground Penetrating Radar (GPR) is a non-destructive detection technique that detects underground structures by emitting electromagnetic waves. It can obtain the reflected wave data inside the tunnel without damaging the tunnel structure, thereby judging the diseases or abnormalities inside the tunnel. These reflected waves contain rich underground information, such as cracks, cavities, moisture content, rock layer distribution, etc. However, how to extract effective information from complex radar waveforms and conduct accurate analysis has long been a difficult problem.

[0004] The challenges faced by traditional ground penetrating radar detection are as follows:

[0005] First, data complexity: The signals collected by ground penetrating radar are reflected waves of high-frequency electromagnetic waves, and usually a large amount of raw data needs to be processed. These data may be interfered by noise, resulting in blurred signals, which poses challenges to subsequent analysis.

[0006] Second, difficult feature extraction: Extracting features related to tunnel diseases from radar waveforms requires strong domain knowledge and manual experience, and different types of diseases may have similar radar reflection features, which is prone to misjudgment.

[0007] Third, low efficiency: Manual or traditional methods often rely on manual analysis of each section of radar data and cannot efficiently process a large amount of waveform information.

[0008] Contents of the Invention

[0009] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide an improved yolov9 method for detecting internal diseases of highway tunnel linings, and solve the technical problem of the large dependence on manpower in the existing disease detection methods.

[0010] To solve the above technical problems, the present invention adopts the following technical solutions to achieve:

[0011] An improved yolov9 method for detecting internal diseases of highway tunnel linings, the method comprising the following steps:

[0012] Step 1, data collection:

[0013] Ground Penetrating Radar is used for data collection to obtain GPR waveform diagrams.

[0014] Step 2, improvement of the yolov9 model:

[0015] Based on the yolov9 model, the yolov9 model is improved to obtain an improved yolov9 model.

[0016] The process of improving the yolov9 model includes:

[0017] Step 201, an EMA attention module is added to the Backbone.

[0018] Step 202, replace GELAN in the Neck layer with GELAN-EMA, where GELAN-EMA is obtained by adding an EMA module after the first convolutional layer of GELAN.

[0019] Step 203, use the Focaler-SIOU loss function for bounding box regression.

[0020] Step 204, add the Center Loss loss function, perform weighted summation with the Focal Loss for classification judgment.

[0021] Step 3, training of the improved yolov9 model:

[0022] Train the improved yolov9 model in Step 2 to obtain the trained improved yolov9 model, which is the disease detection model.

[0023] Step 4, disease detection:

[0024] Input the GPR waveform diagram obtained in Step 1 into the disease detection model trained in Step 3, extract features through the BackBone, then perform feature fusion through the Neck network, and finally obtain the disease detection results inside the highway tunnel lining through the head layer.

[0025] The present invention also has the following technical features:

[0026] In Step 3, use labelme to label the disease data for the GPR waveform diagram collected in Step 1. After the labeling is completed, perform label conversion to convert the json file labeled by labelme into a yolo label format file; divide the disease data in the yolo label format file into a training set and a validation set, enhance the training data, and then train the improved yolov9 model in Step 2 to obtain the trained improved yolov9 model, which is the disease detection model.

[0027] In step 4, the disease detection model includes the following steps:

[0028] Step 401: The Backbone is composed of Conv, GELAN, ADOWN, and EMA attention modules. First, two convolutional operations are performed on the input model file, and then the convolutional result is input into the first GELAN module. Subsequently, ADOWN is used for pooling. After pooling, two GELAN and ADOWN operations are performed again. Finally, the third downsampled feature is passed into the EMA attention module to learn efficient multi-scale attention features across space.

[0029] Step 402: The EMA attention module first performs feature grouping, and each group learns different semantic features. Then, the grouped features are input into parallel sub-networks, and the encoded channel information and multi-scale feature information in two spatial directions are learned through two 1*1 global average pooling branches and a 3*3 convolutional branch respectively. Finally, the cross-space information learned in different spatial dimension directions is aggregated, and these outputs are aggregated through matrix dot product operations to generate the first spatial attention map. Then, the output feature maps within each group are aggregated through the Sigmoid function of the spatial attention weight values generated by the parallel sub-networks to enhance the learning of attention features; the final output of the EMA attention module is the same size as the input.

[0030] Step 403: The output of the EMA attention module is used as the input of the Neck layer. The Neck layer consists of a main branch, an auxiliary reversible branch, and multi-level auxiliary information.

[0031] Compared with the prior art, the present invention has the following technical effects:

[0032] (Ⅰ) The present invention uses a deep convolutional neural network to automatically extract high-dimensional features from a large amount of radar data for classification or regression analysis, which can eliminate the cumbersome steps of manually designing feature extractors and classifiers in traditional methods, provides an effective solution for ground penetrating radar waveform analysis, and greatly reduces the dependence on human resources.

[0033] (Ⅱ) The present invention can achieve automatic feature learning: the deep learning model can automatically learn key features from radar waveforms without manual intervention, avoiding the complexity and subjectivity of manually extracting features.

[0034] (Ⅲ) The present invention can achieve efficient processing: through an end-to-end deep neural network model, radar data can be directly input, and the results of tunnel disease detection can be quickly obtained, greatly improving the detection efficiency.

[0035] (Ⅳ) The present invention can achieve high-precision detection: the deep learning model can be trained with a large amount of data to accurately identify different types of diseases. At the same time, by adding the attention feature module EMA, improving the bounding box regression loss function, and adding the center loss function, the accuracy of the algorithm is improved, which is 2 percentage points higher than the original yolov9 model. Description of the Drawings

[0036] Figure 1 It is a schematic diagram of the detection principle of the ground penetrating radar.

[0037] Figure 2 It is a schematic diagram of the training process of the improved yolov9.

[0038] Figure 3 It is a schematic diagram of the process of the improved yolov9 method for detecting internal diseases of highway tunnel linings.

[0039] Figure 4 It is a schematic diagram of the angular loss of the Focaler-SIOU loss function.

[0040] Figure 5 It is a schematic diagram of the distance loss of the Focaler-SIOU loss function.

[0041] Figure 6 It is a schematic diagram of the GPR waveform diagram collected by the ground penetrating radar.

[0042] Figure 7 It is a schematic diagram of the model structure of the improved yolov9.

[0043] The following further elaborates on the specific content of the present invention in conjunction with the embodiments. Detailed Embodiment

[0044] It should be noted that all the devices, data sets, network models, layers, modules, and loss functions in the present invention, unless otherwise specified, all adopt the devices, data sets, network models, layers, modules, and loss functions known in the prior art.

[0045] The following provides specific embodiments of the present invention. It should be noted that the present invention is not limited to the following specific embodiments, and any equivalent transformation based on the technical solution of this application falls within the protection scope of the present invention.

[0046] Embodiment:

[0047] This embodiment provides an improved yolov9 method for detecting internal diseases of highway tunnel linings, which includes the following steps:

[0048] Step 1, data collection:

[0049] Data collection is carried out using a ground penetrating radar to obtain a GPR waveform diagram.

[0050] Specifically, in this embodiment, the ground penetrating radar is installed on an integrated detection vehicle, and the integrated detection vehicle moves along the tunnel. As Figure 1 shown, the ground penetrating radar is used to detect the inside of the tunnel, and the reflected wave data is collected in real time. Information on the internal structure and diseases of the tunnel can be analyzed based on the strength, arrival time, and reflected waveform of the reflected signal.

[0051] Figure 1 In, a 0 is the incident angle of the transmitted signal I 0 on the interface R 0 a 1 is the exit angle of the reflected signal A 1 on the interface R 0 a 2 is the exit angle of the reflected signal A 2 on the interface R 0 a 3 is the exit angle of the reflected signal A 3 on the interface R 0 b 0 is the incident angle of the transmitted signal I 0 on the interface R 1 b 1 is the exit angle of the reflected signal A 2 on the interface R 1 the exit angle on.

[0052] Specifically, in this embodiment, a schematic diagram of the GPR waveform diagram collected by the ground penetrating radar is as Figure 6 shown, Figure 6 The disease waveform diagram collected by the radar is marked by the red circle in.

[0053] Step two, improvement of the yolov9 model:

[0054] Based on the yolov9 model, the yolov9 model is improved to obtain an improved yolov9 model.

[0055] The process of improving the yolov9 model includes:

[0056] Step 201, an EMA attention module is added to the Backbone.

[0057] Step 202, replace GELAN in the Neck layer with GELAN-EMA, and add the EMA module after the first convolutional layer of GELAN to improve the accuracy and generalization of the model.

[0058] Step 203: Perform bounding box regression using the Focaler-SIOU loss function. For the problem of high similarity of GPR waveform diagrams,

[0059] Step 204: Add the Center Loss loss function to reduce the intra-class distance and increase the inter-class distance, thereby improving the accuracy.

[0060] Specifically in this embodiment, the schematic diagram of the improved yolov9 model structure is as Figure 7 shown. Figure 7 Only the model structure diagram during inference is described in [reference], and there is no auxiliary branch.

[0061] Step 3: Training of the improved yolov9 model:

[0062] As Figure 2 shown, use labelme to label the disease data for the GPR waveform diagrams collected in Step 1. After the labeling is completed, perform label conversion to convert the json file labeled by labelme into a yolo label format file; divide the disease data in the yolo label format file into a training set and a validation set, perform data augmentation, and then train the improved yolov9 model in Step 2 to obtain the improved yolov9 model after training, which is the disease detection model.

[0063] In this embodiment, the characteristics of the disease data are mainly weak signals and uneven oscillations.

[0064] In this embodiment, the improved yolov9 model performs transfer learning using the disease dataset under the official model pre-trained on the COCO dataset.

[0065] In this embodiment, the data augmentation includes mosaic, affine transformation, random cropping, and / or color space transformation. Mosaic randomly arranges, splices, and scales 4 pictures. Affine transformation performs rotation, translation, and scaling operations on the pictures. Color space transformation adjusts the brightness and saturation of the pictures. Data augmentation can improve the generalization of the model.

[0066] Step 4: Disease detection:

[0067] As Figure 3 shown, input the GPR waveform diagram obtained in Step 1 into the disease detection model after training in Step 3, extract features through BackBone, then perform feature fusion through the Neck network, and finally obtain the detection result of the internal diseases of the highway tunnel lining through the head layer.

[0068] As Figure 3As shown, in this embodiment, further, the disease location is obtained through the Focaler-SIOU loss function for bounding box regression, and the disease category is obtained through the Focal Loss function.

[0069] In step four, the disease detection model includes the following steps:

[0070] Step 401, the Backbone is composed of Conv, GELAN, ADOWN, and EMA attention modules. First, two convolutional operations are performed on the input model file, and then the convolutional result is input into the first GELAN module. This module has rich gradient flow information, which is beneficial to improving the accuracy of the model. Subsequently, ADOWN is used for pooling, and after pooling, two GELAN and ADOWN operations are performed again. Finally, the third downsampling feature is passed into the EMA attention module to learn efficient multi-scale attention features across spaces.

[0071] In this embodiment, Conv, GELAN, and ADOWN are all modules known in the art.

[0072] Step 402, the EMA attention module first performs feature grouping, and each group learns different semantic features. Then, the grouped features are input into the parallel sub-network, and the encoded channel information and multi-scale feature information in two spatial directions are learned respectively through two 1*1 global average pooling branches and a 3*3 convolutional branch. Finally, the cross-space information learned in different spatial dimension directions is aggregated, and these outputs are aggregated through matrix dot product operations to generate the first spatial attention map. Then, the output feature maps within each group are aggregated through the Sigmoid function of the spatial attention weight values generated by the parallel sub-network to enhance the learning of attention features; the final output of the EMA attention module is the same size as the input.

[0073] Step 403, the output of the EMA attention module is used as the input of the Neck layer. The Neck layer is composed of a main branch, an auxiliary reversible branch, and multi-level auxiliary information.

[0074] Specifically, in this embodiment, the main branch structure adopts the network structure of PANet (Path Aggregation Network) to participate in the training and inference of the network; the auxiliary reversible branch and multi-level auxiliary information only perform parameter fine-tuning during training to avoid semantic loss caused by deep networks and help improve the performance of the model.

[0075] Specifically, in this embodiment, replacing GELAN in the Neck layer with GELAN-EMA can further learn high-dimensional attention features. The Head layer uses a decoupled head to separately judge class and location information, uses the Focaler-SIOU loss function for bounding box regression and the Focal Loss loss function for class judgment, and at the same time adds the Center Loss loss function to improve the accuracy of classification.

[0076] Specifically, in this embodiment, Center Loss is optimized by defining the center vector of each class and minimizing the distance between each sample feature and the center of its corresponding class. In the object detection task, the Euclidean distance between the feature vector of each candidate box and the center of the corresponding class is calculated, and the classification ability of the model is further optimized by minimizing this distance.

[0077] Specifically, in this embodiment, the Focaler-SIOU loss function takes into account the impact of the imbalance in the distribution of easy and hard samples on bounding box regression, and can focus on different detection samples in different detection tasks to improve the performance of the detector. At the same time, SIOU adds the angular deviation between the predicted box and the ground truth box to the distance deviation on the basis of calculating the IOU loss and the shape loss, which is specifically as follows:

[0078] First, the angular loss:

[0079] Calculate Figure 4 the angular deviation θ between the center point coordinates (c x , c y ) of the ground truth box B and the center point coordinates (c ′ x , c ′ y ) of the predicted box P as:

[0080]

[0081]

[0082] In the formula:

[0083] x is Figure 4 the sine value of the angle α in

[0084] d is the distance between the center point of the ground truth box and the center point of the predicted box.

[0085] Second, the distance loss:

[0086] Calculate Figure 5 the height h, width w of the minimum bounding rectangle in x , c y ) and the center point coordinates (c ′x , c ′ y The deviation between them, combined with the above angular distance, redefines the distance loss δ as:

[0087]

[0088] γ = 2 - θ

[0089] In the formula:

[0090] ρ t is the weight of the distance loss, and the square of the distance is used as the weight here;

[0091] γ is the weight of the angular loss.

[0092] Third, the rectangle loss:

[0093] Calculate the loss σ between the width and height of the predicted box and the ground truth box as;

[0094]

[0095] In the formula:

[0096] w and h represent the width and height of the predicted box respectively;

[0097] w gt , h gt represent the width and height of the ground truth box respectively;

[0098] ω t represents the shape difference between the predicted box and the ground truth box;

[0099] τ is the exponent of the shape loss, used to control the attention of the shape loss.

[0100] From the above, the SIOU loss function L siou is:

[0101]

[0102] In the formula:

[0103] IOU represents the ratio of the intersection and union of the predicted box and the ground truth box.

[0104] Focaler - SIOU considers the difficulty of different sample regressions and redesigned the piecewise loss function IOU folcaer as:

[0105]

[0106] In the formula:

[0107] u is the minimum value of the IOU ratio;

[0108] d is the maximum value of the IOU ratio.

[0109] u, d ∈ [0, 1]. By adjusting the values of u and d, the IOU folcaer can focus on different samples.

[0110] Therefore, the loss function L of Focaler-SIOU focaler-siou is as follows:

[0111] L focaler-siou = L siou + IOU - IOU focaler

[0112] Specifically, in this embodiment, after the ground penetrating radar waveform analysis model is trained, it can automatically analyze the ground penetrating radar waveform diagram and find out the problematic areas. This method can improve the accuracy of tunnel internal disease detection and at the same time greatly reduce the labor cost and misjudgment.

Claims

1. An improved yolov9 highway tunnel lining internal disease detection method, the method comprising the following steps: Step 1: Data collection: Ground penetrating radar is used for data collection to obtain GPR waveforms; Features: Step 2, improvement of yolov9 model: On the basis of yolov9 model, the yolov9 model is improved to obtain the improved yolov9 model; The process of improving the yolov9 model includes: Step 201, adding an EMA attention module in Backbone; Step 202, replace GELAN in the Neck layer with GELAN-EMA, and add the EMA module after the first convolutional layer of GELAN; Step 203, using the Focaler-SIOU loss function to perform bounding box regression; Step 204, adding a Center Loss function and performing weighted summation with the Focal Loss to achieve classification judgment; Step 3: Training of the improved yolov9 model: The improved yolov9 model in step 2 is trained to obtain an improved yolov9 model after training, which is a disease detection model; Step 4: Disease detection: The GPR waveform obtained in step one is input into the disease detection model trained in step three. The features are extracted through BackBone, and then the features are fused through the Neck network. Finally, the internal disease detection results of the highway tunnel lining are obtained through the head layer.

2. The improved yolov9 road tunnel lining internal disease detection method as claimed in claim 1 is characterized in that: In step three, the GPR waveform collected in step one is labeled with disease data using labelme, and label conversion is performed after the labeling is completed, and the json file annotated by labelme is converted into a yolo label format file; the disease data in the yolo label format file is divided into a training set and a validation set, the training data is enhanced, and then the improved yolov9 model in step two is trained to obtain the improved yolov9 model after training, which is the disease detection model.

3. The improved yolov9 road tunnel lining internal disease detection method as claimed in claim 1, characterized in that: In step 4, the disease detection model includes the following steps: Step 401, the Backbone is composed of Conv, GELAN, ADOWN and EMA attention modules. First, the input data is convolved twice, and then the convolution result is input into the first GELAN module, and then ADOWN is used for pooling. After pooling, GELAN and ADOWN operations are performed twice, and finally the third down-sampling feature is passed into the EMA attention module to learn efficient multi-scale attention features across space; Step 402, the EMA attention module first groups features, each group learns different semantic features, and then inputs the grouped features into the parallel sub-network, respectively, and learns the encoding channel information and multi-scale feature information in two spatial directions through two 1*1 global average pooling branches and one 3*3 convolution branch, and finally aggregates the cross-spatial information learned in different spatial dimensions, aggregates these outputs through matrix dot product operations, generates the first spatial attention map, and then aggregates the output feature map in each group through the Sigmoid function of the spatial attention weight value generated by the parallel sub-network to enhance the learning of attention features; The final output of the EMA attention module has the same size as the input; Step 403, the output of the EMA attention module is used as the input of the Neck layer, which consists of a main branch, an auxiliary reversible branch, and multi-level auxiliary information.

Images