Tunnel face crack identification method and system based on parallel cavity convolution

Abstract

The invention provides a tunnel face crack identification method and system based on parallel cavity convolution, and relates to the technical field of deep learning tunnel identification, and the method comprises the steps: obtaining a construction site face image; inputting the tunnel face image into a tunnel face crack recognition model, and outputting a crack recognition result; wherein in the tunnel face crack recognition model, a high-level semantic feature map is extracted from a face image through a MobileNetv3 backbone network, then multi-scale features are extracted through a plurality of parallel cavity convolution layers, the features of different scales are spliced, and unified feature representation is obtained; inputting the uniform feature representation into an ECA attention module, mining a relationship between channels to improve the feature representation capability, obtaining global information of the channels, generating a channel description vector, generating a channel weight vector from the channel description vector through one-dimensional convolution and a Sigmoid activation function, and performing channel weighting by using the channel weight vector to obtain a channel weight vector; and realizing re-calibration of the original input feature map, and completing identification of the tunnel face crack.

Description

Technical Field

[0001] The present disclosure relates to the technical field of deep learning tunnel identification, and in particular to a method and system for identifying cracks in a tunnel face based on parallel void convolution. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] With the acceleration of urbanization and the continuous advancement of infrastructure construction, tunnel engineering plays an increasingly important role in the fields of transportation, water conservancy, mining, etc. However, during the tunnel construction process, the crack problem of the tunnel face (tunnel excavation working face) has always been one of the key factors affecting construction safety and project quality. Timely identification and treatment of tunnel face cracks are of great significance for preventing tunnel collapse, improving construction efficiency and ensuring project safety.

[0004] Traditional crack identification methods mainly rely on manual inspections and empirical judgments, which are not only inefficient but also easily affected by human factors, making it difficult to ensure the accuracy and reliability of the identification results. In recent years, with the rapid development of computer vision and deep learning technologies, automated crack identification technology based on image analysis has gradually become a research hotspot.

[0005] As a powerful machine learning method, deep learning has achieved remarkable results in the fields of image classification, target detection and semantic segmentation. As an important task in image processing, semantic segmentation aims to classify each pixel in the image into a specific semantic category, so as to achieve a fine-grained understanding of the image. In the identification of cracks in the tunnel face, semantic segmentation technology can effectively distinguish cracks from background areas, thereby providing more accurate crack location information. For example, fully convolutional networks (FCN), U-Net, DeepLab, etc., through end-to-end training methods, can automatically learn feature representations in images and achieve high-precision crack identification in complex tunnel face images. These models can not only process high-resolution image data, but also have strong generalization capabilities, and are suitable for crack identification tasks in different environments and working conditions.

[0006] However, due to the complexity and diversity of the tunnel construction environment, higher requirements are placed on crack recognition technology. For example, the ResNet model is used for tunnel disease detection, and the recognition accuracy is improved through transfer learning. The performance of Faster R-CNN and Mask R-CNN in crack detection is compared. The tunnel crack detection method combining improved Retinex and deep learning significantly improves the detection accuracy and speed. However, the current crack detection is mainly aimed at the disease detection of tunnel lining during the operation period, and there are relatively few studies on the detection of cracks on the face during the construction period. In addition, the scale of cracks on the face during the construction period varies greatly, the background is complex, and it is difficult to perform fine extraction. Factors such as lighting conditions, background noise, and material changes at the construction site may affect the recognition effect. In addition, the current crack detection is mainly aimed at the disease detection of tunnel lining during the operation period, and there are relatively few studies on the detection of cracks on the face during the construction period. In addition, the scale of cracks on the face during the construction period varies greatly, the background is complex, and it is difficult to perform fine extraction. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a tunnel face crack recognition method and system based on parallel void convolution, by introducing a semantic segmentation model DeepLabv3+, adding an ECA attention module to the semantic segmentation model DeepLabv3+, and introducing a weighted cross entropy loss function, a new model Crack-DeepLabv3+ suitable for tunnel face crack recognition is formed, the lightweight design of the model is maintained, making it more suitable for deployment on engineering sites, thereby improving the efficiency and reliability of crack detection during construction.

[0008] According to some embodiments, the present disclosure adopts the following technical solutions:

[0009] The tunnel face crack identification method based on parallel void convolution includes:

[0010] Acquire the tunnel face image at the construction site and pre-process it;

[0011] The preprocessed tunnel face image is input into the tunnel tunnel face crack recognition model, and the crack recognition result is output;

[0012] In the tunnel face crack recognition model, the tunnel face image first passes through the MobileNetv3 backbone network to extract high-level semantic feature maps, and then uses multiple parallel hole convolution layers to extract multi-scale features, and then splices features of different scales to obtain a unified feature representation.

[0013] The unified feature representation is input into the ECA attention module, the relationship between channels is mined to improve the feature representation ability, the global information of the channel is obtained, and the channel description vector is generated. The channel description vector is converted into a channel weight vector through one-dimensional convolution and Sigmoid activation function. The channel weight vector is used to weight the channels to achieve recalibration of the original input feature map and complete the identification of the cracks in the tunnel face.

[0014] According to some embodiments, the present disclosure adopts the following technical solutions:

[0015] The tunnel face crack identification system based on parallel void convolution includes:

[0016] Data acquisition module, used to acquire and pre-process the images of the tunnel face at the construction site;

[0017] A crack recognition module is used to input the preprocessed tunnel face image into the tunnel face crack recognition model and output the crack recognition result;

[0018] In the tunnel face crack recognition model, the tunnel face image first passes through the MobileNetv3 backbone network to extract high-level semantic feature maps, and then uses multiple parallel hole convolution layers to extract multi-scale features, and then splices features of different scales to obtain a unified feature representation.

[0019] The unified feature representation is input into the ECA attention module, the relationship between channels is mined to improve the feature representation ability, the global information of the channel is obtained, and the channel description vector is generated. The channel description vector is converted into a channel weight vector through one-dimensional convolution and Sigmoid activation function. The channel weight vector is used to weight the channels to achieve recalibration of the original input feature map and complete the identification of the cracks in the tunnel face.

[0020] According to some embodiments, the present disclosure adopts the following technical solutions:

[0021] A computer program product comprises a computer program, wherein when the computer program is executed by a processor, the method for identifying cracks in a tunnel face based on parallel void convolution is implemented.

[0022] According to some embodiments, the present disclosure adopts the following technical solutions:

[0023] A non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the tunnel face crack identification method based on parallel void convolution is implemented.

[0024] According to some embodiments, the present disclosure adopts the following technical solutions:

[0025] An electronic device comprises: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the tunnel face crack identification method based on parallel void convolution.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The tunnel face crack recognition method based on parallel hole convolution disclosed in this paper designs an improved DeepLabv3+ model. By integrating the MobileNetv3 backbone network in the model, the model's lightweight and effective features are improved, and the ECA attention module is added to effectively capture important features and enhance the model's feature expression ability. At the same time, multi-scale feature extraction is achieved through the ASPP module, and the fusion of low-level features and high-level features is combined to improve the accuracy of identifying subtle cracks. In addition, the weighted cross entropy loss is used to compensate for the class imbalance problem, so that the model can handle complex crack detection tasks more accurately.

[0028] The tunnel face crack identification method based on parallel void convolution disclosed in this paper solves the problem that the model size in the current tunnel crack identification task is large and it is difficult to identify tiny features with high precision. It introduces the semantic segmentation model DeepLabv3+ and makes improvements on this basis, so that it not only achieves good results in detection accuracy, but also maintains the lightweight design of the model, making it more suitable for deployment on the construction site, and improving the efficiency and reliability of crack detection during the construction period. It solves the detection difficulty caused by large scale changes and complex background in the crack detection of the face during the construction period. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0030] Figure 1 Schematic diagram of the original DeepLabv3+ network structure of the embodiment of the present disclosure;

[0031] Figure 2 A schematic diagram of the MobileNetv3 backbone network of an embodiment of the present disclosure;

[0032] Figure 3 Schematic diagram of an ECA attention module according to an embodiment of the present disclosure;

[0033] Figure 4 is a flow chart of the application method of the embodiment of the present disclosure;

[0034] Figure 5 It is a schematic diagram of the verification result of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0035] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0036] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0037] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0038] Example 1

[0039] In one embodiment of the present disclosure, a tunnel face crack identification method based on parallel hole convolution is provided, comprising the following steps:

[0040] Step 1: Obtain the tunnel face image at the construction site and pre-process it;

[0041] Step 2: input the preprocessed tunnel face image into the tunnel face crack recognition model, and output the crack recognition result;

[0042] In the tunnel face crack recognition model, the tunnel face image first passes through the MobileNetv3 backbone network to extract high-level semantic feature maps, and then uses multiple parallel hole convolution layers to extract multi-scale features, and then splices features of different scales to obtain a unified feature representation.

[0043] The unified feature representation is input into the ECA attention module, the relationship between channels is mined to improve the feature representation ability, the global information of the channel is obtained, and the channel description vector is generated. The channel description vector is converted into a channel weight vector through one-dimensional convolution and Sigmoid activation function. The channel weight vector is used to weight the channels to achieve recalibration of the original input feature map and complete the identification of the cracks in the tunnel face.

[0044] As an embodiment, the specific implementation process of the tunnel face crack identification method based on parallel void convolution disclosed in the present invention is as follows:

[0045] Step 1: For crack identification in tunnel faces during construction, DeepLabv3+ was built as the baseline network, MobileNetv3 was used as the backbone network, the ECA attention module was added to the network, and the weighted cross entropy loss function was introduced to form a new model Crack-DeepLabv3+ suitable for crack identification in tunnel faces, and training was carried out, including:

[0046] S1: Construct a dataset of cracks on the tunnel face during the construction period. This embodiment uses pictures collected on site to construct a dataset of cracks on the tunnel face during the construction period, which has the characteristics of large scale changes and complex features. The resolution of each image is 512*512 pixels, the number of target categories is set to 2, and the category name is set to cracks and background. A total of 4,000 training sample libraries and 800 test sample libraries are established. The Labelme annotation tool is used to manually annotate the contour shape of the cracks on the tunnel face, and a png file is generated as a label.

[0047] S2: Build the DeepLabv3+ model.

[0048] First of all, the basis of DeepLabv3+ is a backbone network (Backbone), and this disclosure uses MobileNetv3 as the backbone network. The convolution layer of the backbone network is responsible for capturing the complex features of the input image. The output of the backbone network is a high-level feature map, which contains rich semantic information and usually takes the output of the last layer of convolution of the backbone network. In addition, in order to enhance the segmentation accuracy of the model, low-level features need to be extracted from the backbone network. This can be achieved by accessing earlier convolution layers. These extracted low-level features include texture and edge information, which is particularly important for identifying the details and edges of cracks. These detailed information can help the feature decoder better restore the spatial resolution of the image.

[0049] Next, DeepLabv3+ introduces the Atrous Spatial Pyramid Pooling (ASPP) module, which is its core part. ASPP extracts multi-scale features through multiple parallel dilated convolutions. By setting different dilation rates (such as 1, 6, 12, and 18), ASPP can capture objects of different sizes in the image under different receptive fields. In addition, ASPP also contains a global average pooling operation to extract global contextual information. After all these feature maps of different scales are spliced, a 1x1 convolution is used to reduce the number of output channels to obtain a unified feature representation. The cracks in the face may vary greatly in size and shape. The multi-scale feature extraction of the ASPP module captures these diverse crack information through different dilation rates, which helps to identify cracks that are both wide and narrow.

[0050] After obtaining the high-level features and the multi-scale features extracted by ASPP, the next step is the decoder. The goal of the decoder is to restore the spatial resolution of the image and enhance the segmentation accuracy of the model. Specifically, the decoder first upsamples the output of the ASPP module, usually using bilinear interpolation, to restore its spatial resolution to the size of the input image to ensure the accurate location of the crack is identified. Then, the upsampled feature map is spliced ​​with the low-level features extracted from the backbone network, so that the spatial details in the low-level features can be utilized. In order to further process the spliced ​​feature map, the decoder will also include several convolutional layers, and finally output a feature map with the corresponding number of category channels to improve the recognition accuracy of the crack edge.

[0051] Finally, the output layer maps the decoder's output feature map to a specific category number through a 1x1 convolution, which maps the feature map to a crack or non-crack category to generate an accurate binary segmentation map. These segmentation maps can be used to calculate the loss function to measure the difference between the predicted segmentation map and the true label.

[0052] Through these steps, the improved DeepLabv3+ can provide an accurate and efficient solution for tunnel face crack identification, and its powerful segmentation performance can be applied to engineering practice to assist in intelligent detection and assessment of crack risks.

[0053] In the improved DeepLabv3+ model, the ECA attention module is added after the last deep convolution of the backbone network. Because the feature map size at this position is moderate, it can maintain computational efficiency and effectively improve the feature expression ability of the model. ECA (Efficient Channel Attention) is an efficient channel attention mechanism that aims to improve the accuracy of crack identification by strengthening the focus on important features. The basic idea of ​​the ECA module is to improve the feature representation ability by effectively mining the relationship between channels without increasing excessive computational overhead.

[0054] As an embodiment, the improved DeepLabv3+ model model as a whole includes four main parts: the backbone network (Backbone), the ASPP module, the decoder and the output layer. The processing order is as follows: First, MobileNetv3 is used as the backbone network to capture the high-level features of the input image and extract low-level features to enhance the detail information. Next, multi-scale feature extraction is performed through the ASPP module, and multi-scale semantic information is obtained by using hole convolutions with different hole rates and global average pooling operations. Then, the decoder part upsamples the output of the ASPP module and splices it with the low-level features extracted from the backbone network, and restores the spatial resolution of the image through further convolution processing to enhance the segmentation accuracy. Finally, the output feature map of the decoder is mapped to a specific number of categories through 1x1 convolution to generate an accurate binary segmentation map for calculating the loss function and measuring the difference between the prediction and the true label.

[0055] Its detailed structure can be divided into the following steps:

[0056] First, the ECA module receives the input feature map, which is usually a three-dimensional tensor with a feature map shape of ×H×W, where C represents the number of channels, and H and W represent the height and width of the feature map, respectively. Next, the module performs global average pooling on the input features. The main purpose of this step is to obtain the global information of the channel from the feature map and generate a one-dimensional channel description vector with a shape of C×1×1.

[0057] After obtaining the channel description vector, the ECA module then calculates the relationship between channels through a one-dimensional convolution (1D Convolution) operation. The convolution kernel size of this one-dimensional convolution is usually set to a small value, such as 3 or 5. Unlike the traditional fully connected layer, the ECA module captures the relationship between different channels through one-dimensional convolution. This step is very critical because different channels may correspond to different texture features, edge features or other important information for identifying cracks, and one-dimensional convolution can effectively reduce the number of parameters and calculations when extracting channel relationships. In this way, the model can more effectively identify and associate important but not obvious crack information. The feature map calculated by the one-dimensional convolution is then activated by the Sigmoid function to generate a channel weight vector. Each value of this weight vector represents the importance information of the corresponding channel.

[0058] Finally, the ECA module completes channel weighting by element-by-element multiplication of the generated weight vector with the channels of the original input feature map, realizing the recalibration of the feature map, so that the features that are critical to crack identification are enhanced and irrelevant features are suppressed. This precise weighting helps to improve the model's ability to distinguish crack details in complex rock backgrounds. In general, the ECA module implements an effective channel attention mechanism with less computational and parameter overhead, increasing the possibility of small cracks being correctly identified in complex backgrounds. Considering the simple design of the ECA module, the small amount of computation and parameters, this allows it to be embedded in a real-time processing system without significantly increasing the hardware burden. It can better handle complex and changing environments and provide more stable, accurate and efficient crack detection capabilities.

[0059] In the improved DeepLabv3+ model, since the existence of cracks is usually small and complex, the improved loss function can effectively help the model better learn and recognize these subtle features. Therefore, after the decoder output of the improved DeepLabv3+, the weighted cross entropy loss is calculated at the position of the final generated segmentation map. The formula of the weighted cross entropy loss function is as follows:

[0060]

[0061] Among them, C represents the total number of categories, y i is the true label, p i is the probability predicted by the model that a pixel belongs to category i, w i is the weight of the corresponding category. In code implementation, it can usually be directly called through WeightedCrossEntropyLoss in the loss function library.

[0062] In tunnel face images, the number of crack pixels is usually far less than that of background pixels, resulting in serious class imbalance. Weighted cross entropy loss can alleviate this problem by assigning higher weights to crack categories. Moreover, by increasing the weight of crack categories, the model can pay more attention to these small and important features, thereby improving the recognition accuracy of cracks.

[0063] This paper improves the DeepLabv3+ model and generates a new tunnel face crack recognition model Crack-DeepLabv3+, the specific structure of which is shown in Table 1 below:

[0064] Table 1: Crack-DeepLabv3+ specific structure

[0065]

[0066]

[0067] The constructed network model is used to train and optimize the sample library to obtain the optimal model. The optimal model obtained by S6 is used to verify the recognition of the validation set to obtain the recognition result.

[0068] Step 2: Obtain the tunnel face image at the construction site and preprocess it; input the preprocessed tunnel face image into the tunnel tunnel face crack recognition model trained in step 1, and output the crack recognition result;

[0069] Specifically, the model processing process is as follows:

[0070] Step 1. Input an image in the initial convolutional layer with a size of 512×512×3 (color image). After one convolutional layer, the output size becomes 256×256×16. This step extracts the basic features of the image, such as edge, texture, and color information. These features lay the foundation for subsequent feature extraction, enabling the model to recognize some basic structures and shapes in the image.

[0071] Step 2.: The 256×256×16 feature map is input to the middle layer convolution block 1. After the depthwise separable convolution (DW convolution), the output size becomes 128×128×24. This step further extracts the intermediate features of the image, such as local shape and structure. These features begin to contain the preliminary shape and direction information of the cracks, which is helpful for crack recognition in the middle and late stages.

[0072] Step 3. In the middle layer convolution block 2, the input feature map of 128×128×24 is passed through another depth-wise separable convolution (DW convolution), and the output size becomes 64×64×40. This step continues to deepen the feature extraction and form more complex intermediate features such as contours and textures.

[0073] Step 4. Input the 64×64×40 feature map in the middle layer to extract low-level features, which are mainly processed by convolution blocks and the output size becomes 128×128×24. This step extracts low-level features of the image, which retain more spatial detail information, such as the edges and local structures of the cracks, and provides an important basis for subsequent feature fusion.

[0074] Step 5. The feature map of 64×64×40 is input to the final convolution of the Bottleneck layer. After the final convolution of the Bottleneck layer, the output size becomes 32×32×96. This step extracts high-level features that represent more abstract concepts, such as the overall shape of the cracks, their distribution locations and other related structures, and their relationship with the overall background.

[0075] Step 6. Input the 32×32×96 feature map into the ECA attention module, and after global average pooling and 1D convolution, obtain the weighted feature map, and the output size is still 32×32×96. This step enhances the importance of key features and ensures that the model can pay more attention to the key parts in recognition, especially the characteristics of cracks.

[0076] Step 7. Input the 32×32×96 feature map into the ASPP module, and after multi-scale dilated convolution and splicing, the output size becomes 32×32×128. This step enhances the feature expression capability through multi-scale information fusion, so that crack features of different scales can be comprehensively considered.

[0077] Step 8. Upsampling: Input the feature map of 32×32×128, and after bilinear interpolation upsampling, the output size becomes 128×128×128. This step restores the feature map to half the size of the original image, ensuring that subsequent processing can maintain the spatial information of the image.

[0078] Step 9. Concatenate the upsampled feature map (128 × 128 × 128) with the low-level features (128 × 128 × 24) for feature fusion, and the output size becomes 128 × 128 × 152. This step integrates features at different levels, combining the spatial details of low-level features and the semantic information of high-level features.

[0079] Step 10. Input the 128×128×152 feature map to the convolution layer. After convolution processing, the output size becomes 512×512×64. This step further enhances the feature expression and helps the model describe complex crack information more accurately.

[0080] Step 11. Input the 512 × 512 × 64 feature map to the output layer, and after 1x1 convolution, the output size becomes 512 × 512 × 2. This step generates the final crack recognition result, providing classification information for each pixel, indicating whether a crack exists or not.

[0081] Example 2

[0082] In one embodiment of the present disclosure, a tunnel face crack identification system based on parallel void convolution is provided, comprising:

[0083] Data acquisition module, used to acquire and pre-process the images of the tunnel face at the construction site;

[0084] A crack recognition module is used to input the preprocessed tunnel face image into the tunnel face crack recognition model and output the crack recognition result;

[0085] In the tunnel face crack recognition model, the tunnel face image first passes through the MobileNetv3 backbone network to extract high-level semantic feature maps, and then uses multiple parallel hole convolution layers to extract multi-scale features, and then splices features of different scales to obtain a unified feature representation.

[0086] The unified feature representation is input into the ECA attention module, the relationship between channels is mined to improve the feature representation ability, the global information of the channel is obtained, and the channel description vector is generated. The channel description vector is converted into a channel weight vector through one-dimensional convolution and Sigmoid activation function. The channel weight vector is used to weight the channels to achieve recalibration of the original input feature map and complete the identification of the cracks in the tunnel face.

[0087] Example 3

[0088] In one embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the method for identifying cracks in a tunnel face based on parallel void convolution is implemented.

[0089] Example 4

[0090] In one embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the tunnel face crack identification method based on parallel void convolution is implemented.

[0091] Example 5

[0092] In one embodiment of the present disclosure, an electronic device is provided, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the tunnel face crack identification method based on parallel void convolution.

[0093] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0094] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0095] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A tunnel face crack identification method based on parallel void convolution, characterized in that: include: Acquire the tunnel face image at the construction site and pre-process it; The preprocessed tunnel face image is input into the tunnel tunnel face crack recognition model, and the crack recognition result is output; In the tunnel face crack recognition model, the tunnel face image first passes through the MobileNetv3 backbone network to extract high-level semantic feature maps, and then uses multiple parallel hole convolution layers to extract multi-scale features, and then splices features of different scales to obtain a unified feature representation. The unified feature representation is input into the ECA attention module, the relationship between channels is mined to improve the feature representation ability, the global information of the channel is obtained, and the channel description vector is generated. The channel description vector is converted into a channel weight vector through one-dimensional convolution and Sigmoid activation function. The channel weight vector is used to weight the channels to achieve recalibration of the original input feature map and complete the identification of the cracks in the tunnel face.

2. The tunnel face crack identification method based on parallel void convolution according to claim 1, characterized in that: MobileNetv3 is used as the backbone network of the road face crack recognition model. The convolution layer of the backbone network is responsible for extracting the complex features of the input image. The output of the backbone network is a high-level semantic feature map containing rich semantic information. The output of the last convolution layer of the backbone network is taken.

3. The tunnel face crack identification method based on parallel void convolution according to claim 1, characterized in that: The tunnel face crack recognition model introduces parallel hole convolution, extracts multi-scale features through multiple parallel hole convolutions, captures objects of different sizes in the image under different receptive fields by setting different hole rates, and extracts global context information through global average pooling. After splicing feature maps of different scales, a unified feature representation is output by reducing the number of channels.

4. The tunnel face crack identification method based on parallel void convolution according to claim 1, characterized in that: The ECA attention module is added after the last deep convolution of the backbone network. The ECA module receives the input feature map, which is a three-dimensional tensor with a feature map of shape C×H×W. It performs global average pooling on the input feature map, obtains the global information of the channel from the feature map, and generates a one-dimensional channel description vector.

5. The tunnel face crack identification method based on parallel void convolution according to claim 4, characterized in that: After obtaining the channel description vector, the relationship between channels is calculated through a one-dimensional convolution operation. The convolution kernel size of the one-dimensional convolution is set to a small value. The feature map calculated by the one-dimensional convolution is then activated by the Sigmoid function to generate a channel weight vector. Each value of the channel weight vector represents the importance information of the corresponding channel.

6. The tunnel face crack identification method based on parallel void convolution according to claim 1, characterized in that: Channel weighting is completed by element-by-element multiplication of the generated weight vector with the channel of the original input feature map, and the feature map is recalibrated so that the features that are critical to crack identification are enhanced and irrelevant features are suppressed.

7. The tunnel face crack identification system based on parallel void convolution is characterized by: include: Data acquisition module, used to acquire and pre-process the images of the tunnel face at the construction site; A crack recognition module is used to input the preprocessed tunnel face image into the tunnel face crack recognition model and output the crack recognition result; In the tunnel face crack recognition model, the tunnel face image first passes through the MobileNetv3 backbone network to extract high-level semantic feature maps, and then uses multiple parallel hole convolution layers to extract multi-scale features, and then splices features of different scales to obtain a unified feature representation. The unified feature representation is input into the ECA attention module, the relationship between channels is mined to improve the feature representation ability, the global information of the channel is obtained, and the channel description vector is generated. The channel description vector is converted into a channel weight vector through one-dimensional convolution and Sigmoid activation function. The channel weight vector is used to weight the channels to achieve recalibration of the original input feature map and complete the identification of the cracks in the tunnel face.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for identifying cracks in a tunnel face based on parallel void convolution according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the tunnel face crack identification method based on parallel void convolution according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the tunnel face crack identification method based on parallel void convolution as described in any one of claims 1 to 7.

Images