Tunnel apparent disease identification method and system

Abstract

The invention provides a tunnel apparent disease identification method, which comprises the steps of preprocessing a collected tunnel apparent image set to obtain a preprocessed image set, and marking a disease area label on the preprocessed image set as a first training set; inputting the first training set into a disease area detection model for training to obtain a trained disease area detection model, and obtaining an output result of the disease area detection model; a classification label is added to the output result to obtain a classification label image set, the classification label image set serves as a second training set, the second training set is input into the improved PSPNet semantic segmentation model for training, and the trained improved PSPNet semantic segmentation model is obtained; sequentially inputting a to-be-detected tunnel appearance image into the trained disease area detection model and the trained improved PSPNet semantic segmentation model to obtain a semantic segmentation image; and determining a tunnel apparent disease identification result according to the obtained semantic segmentation image. According to the invention, a block detection method is adopted to amplify disease detail features, and the accuracy and recall rate of small target detection are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of highway tunnel civil structure detection, relates to tunnel apparent disease detection, and specifically relates to a method and system for identifying tunnel apparent diseases. Background Art

[0002] With the acceleration of the urbanization process and the increasing complexity of the transportation network, tunnels have become an indispensable and important part of modern transportation infrastructure. However, with the increase in service life and the change of external environment, various diseases often appear in tunnel structures and surfaces. These diseases not only affect the safety of tunnels, but may also lead to traffic accidents or equipment damage, bringing great risks to public safety and economic activities. Therefore, how to detect tunnel apparent diseases in a timely and accurate manner and take effective repair measures has become an important issue in tunnel operation and management.

[0003] Traditional tunnel disease detection methods mainly rely on manual inspections or traditional mechanized detection means. These methods not only have a high labor intensity, but are also limited by the experience and visual ability of detection personnel, easily leading to problems such as missed detections and misdetections. In addition, manual inspections cannot cover large areas of tunnels in a short time, and it is difficult to carry out efficiently in complex or dangerous environments, especially in the case of insufficient light and narrow space inside the tunnel, the detection of diseases is more difficult.

[0004] With the development of deep learning and computer vision technologies, it has become possible to use automated means to quickly and accurately detect tunnel surface diseases. However, the existing deep learning-based detection methods have low detection accuracy and recall rates for tunnel lining cracks, and at the same time have problems such as poor generalization, and urgently need to be improved. Summary of the Invention

[0005] Aiming at the deficiencies and defects of the above-mentioned existing technologies, the purpose of the present invention is to provide a method and system for identifying tunnel apparent diseases to solve the technical problem of high false detection rate and missed detection rate in tunnel apparent disease detection in the existing technology.

[0006] To solve the above technical problems, the present invention is implemented by adopting the following technical solutions:

[0007] A method for identifying tunnel apparent diseases includes the following steps:

[0008] Step 1: Preprocess the collected tunnel apparent image set to obtain a preprocessed image set, and label the disease areas thereon as the first training set;

[0009] Step 2: Input the first training set into the disease area detection model for training to obtain a trained disease area detection model, and obtain the output result of the disease area detection model;

[0010] Step 3: Add classification labels to the output results to obtain a classified label image set, which is used as the second training set. Input the second training set into the improved PSPNet semantic segmentation model for training to obtain a trained improved PSPNet semantic segmentation model;

[0011] Step 4: Input the tunnel apparent images to be detected into the trained disease area detection model and the trained improved PSPNet semantic segmentation model in sequence to obtain semantic segmentation images;

[0012] Step 5: Determine the tunnel apparent disease recognition results based on the obtained semantic segmentation images.

[0013] The present invention also has the following technical features:

[0014] Specifically, the disease area detection model includes a Backbone module, a Neck module, and a Head module connected in sequence;

[0015] The Backbone module is used to extract multi-scale features of the input image, the Neck module is used to fuse the multi-scale features extracted by the Backbone module, and the Head module is used to classify and locate the multi-scale features fused by the Neck module;

[0016] The Backbone module includes a Silence unit, a first Conv unit, a second Conv unit, a first GELAN unit, a first ADOWN unit, a second GELAN unit, a second ADOWN unit, a third GELAN unit, a third ADOWN unit, a fourth GELAN unit, and an SPFFLAN unit connected in sequence;

[0017] Among them, the first Conv unit and the second Conv unit are both used to downsample the input image, the first GELAN unit, the second GELAN unit, the third GELAN unit, and the fourth GELAN unit are used to extract features from the input image to obtain multi-scale feature maps, and the SPFFLAN unit is used to extract multi-scale features from the multi-scale feature maps.

[0018] Furthermore, the Neck module includes a third Conv unit, a first ASFF2 unit, a fifth GELAN unit, a first ASFF3 unit, and a sixth GELAN unit connected in sequence; a fourth Conv unit, a second ASFF2 unit, a seventh GELAN unit, a second ASFF3 unit, and an eighth GELAN unit connected in sequence; and a fifth Conv unit, a third ASFF3 unit, and a ninth GELAN unit connected in sequence;

[0019] Among them, the output of the SPFFLAN unit is connected to the input of the fifth Conv unit; the output of the second GELAN unit is connected to the input of the third Conv unit; the output of the third GELAN unit is connected to the input of the fourth Conv unit; the third Conv unit is also respectively connected to the first ASFF2 unit and the second ASFF2 unit; the fourth Conv unit is also respectively connected to the first ASFF2 unit and the second ASFF2 unit; the sixth GELAN unit and the ninth GELAN unit are both connected to the eighth GELAN unit;

[0020] Both the first ASFF2 unit and the second ASFF2 unit are used for dynamically fusing features of two levels, and both the first ASFF3 unit, the second ASFF3 unit and the third ASFF3 unit are used for dynamically fusing features of three levels.

[0021] Furthermore, the Head module includes a first detector, a second detector, a third detector and 3 Concat modules. The first detector, the second detector and the third detector all include a Conv-reg module and a Conv-cls module arranged in parallel. The input ends of the Conv-reg module and the Conv-cls module are both connected to the Neck module, and the output ends of the Conv-reg module and the Conv-cls module are both connected to a corresponding Concat module;

[0022] The Conv-cls module is used to output the category of the disease, the Conv-reg module is used to output the localization information of the disease, and the Concat module is used to splice the output results of the Conv-cls module and the Conv-reg module to obtain the output result of the disease area detection model.

[0023] Furthermore, the preprocessing in step 1 specifically includes the following sub-steps:

[0024] Step 1.1: Use the multi-scale restricted contrast adaptive histogram equalization algorithm to enhance the contrast of the tunnel apparent image to obtain a tunnel apparent image with enhanced contrast;

[0025] Step 1.2: Perform smoothing denoising, gamma correction, random cropping, scaling, flipping and Mosaic processing on the obtained tunnel apparent image with enhanced contrast in sequence to obtain the preprocessed image.

[0026] Further, the improved PSPNet semantic segmentation model replaces the ResNet50 module of the original PSPNet semantic segmentation model with a RepVGG Backbone module, and the RepVGG Backbone module is composed of multiple stacked RepVGG modules. Each RepVGG module includes a 3x3 convolutional layer, a 1x1 convolutional branch layer, and an identity mapping branch arranged in parallel.

[0027] Further, the classification label in step 3 is a number, where crack is 0 and spalling is 1.

[0028] Further, determining the tunnel apparent disease recognition result according to the obtained semantic segmentation image in step 4 includes determining the crack length, or the crack width, or the spalling area according to the semantic segmentation image.

[0029] Further, the crack length is determined by the skeleton line extraction method, the crack width is determined by the maximum inscribed circle method, and the spalling area is determined by calculating the connected domain area using Green's formula.

[0030] The present invention also protects a tunnel apparent disease recognition system, including a data acquisition and preprocessing module, a disease area detection model training module, an improved PSPNet semantic segmentation model training module, a detection module, and a disease recognition module;

[0031] The data acquisition and preprocessing module is used to preprocess the collected tunnel apparent image set to obtain a preprocessed image set, and label it with disease area labels as the first training set;

[0032] The disease area detection model training module is used to input the first training set into the disease area detection model for training to obtain a trained disease area detection model, and obtain the output result of the disease area detection model;

[0033] The improved PSPNet semantic segmentation model training module is used to add classification labels to the output result to obtain a classification label image set, which is used as the second training set, and input the second training set into the improved PSPNet semantic segmentation model for training to obtain a trained improved PSPNet semantic segmentation model;

[0034] The detection module is used to sequentially input the to-be-detected tunnel apparent image into the trained disease area detection model and the trained improved PSPNet semantic segmentation model to obtain a semantic segmentation image;

[0035] The disease recognition module is used to determine the tunnel apparent disease recognition result according to the obtained semantic segmentation image.

[0036] Compared with the prior art, the present invention has the following technical effects:

[0037] (1) Starting from the input preprocessed images and the network structure of the improved deep learning model, the method of the present invention improves the accuracy of the algorithm. By adopting the block detection method, the detailed features of the diseases are magnified, and the accuracy and recall rate of small target detection are improved.

[0038] (2) The present invention first adopts multi-scale adaptive histogram equalization to improve the image quality. Aiming at the problem of information degradation existing in the existing model, with the help of the constructed CARAFE-AFPN network, the upsampling kernel is determined according to semantic features and multi-scale feature fusion between different layers is carried out, improving the accuracy of target detection. Finally, the RepVGG module and the EMA attention network are adopted to improve the accuracy of pixel classification, obtaining accurate disease contour information, which is beneficial to improving the calculation accuracy of the subsequent crack length, width or spalling area. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flowchart of the method of the present invention;

[0040] Figure 2 is a schematic structural diagram of the disease area detection model in Embodiment 1;

[0041] Figure 3 is a schematic structural diagram of the improved PSPNet semantic segmentation model;

[0042] Figure 4 is a schematic structural diagram of the RepVGG module.

[0043] The following further elaborates on the specific content of the present invention in conjunction with embodiments. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] 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 the present application falls within the protection scope of the present invention.

[0045] Embodiment 1

[0046] Following the above technical solution, this embodiment provides a method for identifying tunnel apparent diseases, as Figure 1 shown, including the following steps:

[0047] Step 1. Preprocess the collected tunnel apparent image set to obtain a preprocessed image set, and label the disease areas thereon as the first training set;

[0048] Specifically, high-quality apparent images inside the tunnel are collected by using a high-definition camera or a laser scanner; the preprocessing specifically includes the following sub-steps:

[0049] Step 1.1: Use the multi-scale limited contrast adaptive histogram equalization algorithm to enhance the contrast of the acquired apparent image, obtaining a tunnel apparent image with enhanced contrast.

[0050] Specifically, in this embodiment, the acquired tunnel apparent image is subjected to limited contrast adaptive histogram equalization processing at three scales of (16, 16), (32, 32), and (64, 64). The limited contrast adaptive histogram equalization processing is mainly used to enhance the local contrast of the image while avoiding the noise amplification and over-enhancement problems that may be introduced by traditional histogram equalization. Through processing, the gray values of the overly dark image can be evenly distributed, the contrast of the image is enhanced, and the detail information of the image can be retained.

[0051] Step 1.2: Use bilateral filtering to perform smoothing denoising, gamma correction, random cropping, scaling, flipping, and Mosaic processing on the tunnel apparent image with enhanced contrast obtained in Step 1.1, adjust the image size in the preprocessed image set to 640×640, and label the diseases areas in the preprocessed image set as the first training set.

[0052] Step 2: Input the first training set into the disease area detection model for training to obtain a trained disease area detection model and the output result of the disease area detection model.

[0053] As Figure 2 shown, in this embodiment, the disease area detection model is based on the single-stage yolov9 model, including a Backbone module, a Neck module, and a Head module connected in sequence.

[0054] Among them, the Backbone module is used to extract multi-scale features of the input image, the Neck module is used to fuse the multi-scale features extracted by the Backbone module, and the Head module is used to classify and locate the multi-scale features fused by the Neck module.

[0055] The Backbone module includes a Silence unit, a first Conv unit, a second Conv unit, a first GELAN unit, a first ADOWN unit, a second GELAN unit, a second ADOWN unit, a third GELAN unit, a third ADOWN unit, a fourth GELAN unit, and a SPFFLAN unit connected in sequence.

[0056] Among them, the Silence unit is used to assist the branch call to input the image into the network; both the first Conv unit and the second Conv unit use a 3×3 convolution kernel to downsample the input image; the first GELAN unit, the second GELAN unit, the third GELAN unit, and the fourth GELAN unit are used for feature fusion, and the SPFFLAN unit is used for multi-scale feature extraction of multi-scale feature maps.

[0057] In this embodiment, the first GELAN unit, the second GELAN unit, the third GELAN unit, and the fourth GELAN unit all adopt the RepNCSPELAN structure composed of RepCSPNet and ELAN network. Among them, RepCSPNet combines the existing RepConv and CSPNet to achieve feature fusion, increase rich gradient flow information, and improve the model performance. During inference, RepNCSPELAN can perform structural reparameterization, changing the multi-path structure into a single-path structure, reducing the number of parameters, reducing the video memory occupancy, and improving the inference speed.

[0058] As a preferred solution of this embodiment, the Neck module includes a third Conv unit, a first ASFF2 unit, a fifth GELAN unit, a first ASFF3 unit, and a sixth GELAN unit connected in sequence; a fourth Conv unit, a second ASFF2 unit, a seventh GELAN unit, a second ASFF3 unit, and an eighth GELAN unit connected in sequence; and a fifth Conv unit, a third ASFF3 unit, and a ninth GELAN unit connected in sequence;

[0059] Among them, the output of the SPFFLAN unit is connected to the input of the fifth Conv unit; the output of the second GELAN unit is connected to the input of the third Conv unit; the output of the third GELAN unit is connected to the input of the fourth Conv unit; the third Conv unit is also respectively connected to the first ASFF2 unit and the second ASFF2 unit; the fourth Conv unit is also respectively connected to the first ASFF2 unit and the second ASFF2 unit; both the sixth GELAN unit and the ninth GELAN unit are connected to the eighth GELAN unit;

[0060] Both the first ASFF2 unit and the second ASFF2 unit are used for dynamically fusing the features of two levels, and both the first ASFF3 unit, the second ASFF3 unit, and the third ASFF3 unit are used for dynamically fusing the features of three levels, and the first ASFF3 unit, the second ASFF3 unit, and the third ASFF3 unit all adopt CARAFE to perform upsampling.

[0061] Specifically, in the Neck module, the existing PANet is replaced with CARAFE-AFPN composed of the CARAFE network and the AFPN network. CARAFE is used to implement upsampling. CARAFE can generate an upsampling kernel according to the input feature semantic information, and has a large receptive field, which can make good use of the surrounding information. ASFF, that is, the Spatial Adaptive Network, where N in ASFFX is an Arabic numeral, indicating that N layers of features are fused together.

[0062] Specifically, CARAFE includes an upsampling module and a feature recombination module with connection settings. Among them, the upsampling kernel prediction module is used to predict the upsampling kernel. For example, if the upsampling ratio is σ, and the size of the input feature map is H×W×C, then the upsampling kernel prediction module first compresses the channels of the input image to obtain a feature map of H×W×C1, then performs feature encoding to obtain a feature map of H×W×σ 2 ×k 2 of the feature map, then performs convolution to obtain the upsampling kernel, and finally normalizes the upsampling kernel; the feature recombination module convolves the input data, convolves the obtained feature map with the upsampling kernel to obtain a feature map of σH×σW×σC. AFPN is a progressive feature pyramid network that can solve the problem of feature information degradation in PANet.

[0063] In this embodiment, the Neck module uses the features C3 (corresponding to a feature dimension of 80×80), C4 (corresponding to a feature dimension of 40×40), and C5 (corresponding to a feature dimension of 20×20) of different scales generated by the BackBone module as inputs. The low-level features C3 and C4 are fused in the feature pyramid network, and then further fused with the feature C5 to obtain the fused features, ultimately improving the model accuracy and generalization.

[0064] The Head module includes a first detector, a second detector, a third detector, and 3 Concat modules. The first detector, the second detector, and the third detector all include a Conv-reg module and a Conv-cls module arranged in parallel. The input ends of the Conv-reg module and the Conv-cls module are both connected to a corresponding Neck module, and the output ends of the Conv-reg module and the Conv-cls module are both connected to the Concat module to obtain the final prediction result;

[0065] The Concat module is used to splice the output results of the Conv-cls module and the Conv-reg module to obtain the output result of the disease area detection model.

[0066] Step 3: Add classification labels to the output results to obtain a classification label image set, which is used as the second training set. Input the second training set into the improved PSPNet semantic segmentation model for training to obtain a trained improved PSPNet semantic segmentation model;

[0067] In this embodiment, to improve the accuracy, misdetected disease pictures can be manually deleted first, and then Segment-Anything can be used for automatic annotation to generate semantic segmentation labels in Labelme format. The role of semantic segmentation is to classify pixels, and the semantic segmentation labels are numbers. Among them, cracks are 0 and spalling is 1.

[0068] Such as Figure 3 shown, in the improved PSPNet semantic segmentation model, the ResNet50 module of the original PSPNet semantic segmentation model is replaced by the RepVGG Backbone module. The RepVGG Backbone module is stacked by multiple RepVGG modules. Each of the RepVGG modules includes a 3x3 convolutional layer, a 1x1 convolutional branch layer, and an identity mapping branch arranged in parallel.

[0069] Specifically, the improved PSPNet semantic segmentation model includes a RepVGG Backbone module, an attention mechanism module EMA, a pooling layer, a convolutional layer I, an upsampling module, and a convolutional layer II connected in sequence.

[0070] The improved PSPNet semantic segmentation model adopts the idea of reparameterization to convert parallel multi-branches into a single-branch model, improving the inference speed of the model; input the features extracted by the RepVGG Backbone module into the EMA attention module to learn efficient multi-scale attention features through grouped cross-space learning; input into the PSPModule, downsample to different sizes through a pooling layer, reduce the number of channels of the features through the convolutional layer conv, then use the upsampling operation to restore the feature maps of different sizes to the original image size, then splice the feature maps obtained by the EMA attention module and the feature maps output by the upsampling in the channel dimension, and finally conv reduces the number of channels to the number of categories, performs softmax processing to obtain a probability map, and obtains a predicted mask map. Different-scale feature fusion is used to obtain context information with different details, improving the performance of the model.

[0071] Step 4: Determine the tunnel apparent disease recognition result according to the obtained semantic segmentation image.

[0072] Determining the tunnel apparent disease recognition result according to the obtained semantic segmentation image includes determining the crack length, or the crack width, or the spalling area according to the semantic segmentation image.

[0073] Furthermore, the crack length is determined by the existing skeleton line extraction method, the crack width is determined by the existing maximum inscribed circle method, and the spalling area is determined by calculating the connected domain area through the existing Green's formula.

[0074] Embodiment 2

[0075] This embodiment provides a tunnel apparent disease identification system for implementing the tunnel apparent disease identification method disclosed in Embodiment 1, including a data acquisition and preprocessing module, a disease area detection model training module, an improved PSPNet semantic segmentation model training module, a detection module, and a disease identification module.

[0076] The data acquisition and preprocessing module is used to preprocess the acquired tunnel apparent image set to obtain a preprocessed image set, and label it with disease area labels as the first training set.

[0077] The disease area detection model training module is used to input the first training set into the disease area detection model for training to obtain a trained disease area detection model and the output result of the disease area detection model.

[0078] The improved PSPNet semantic segmentation model training module is used to add classification labels to the output result to obtain a classification label image set, which is used as the second training set, and input the second training set into the improved PSPNet semantic segmentation model for training to obtain a trained improved PSPNet semantic segmentation model.

[0079] The detection module is used to sequentially input the tunnel apparent image to be detected into the trained disease area detection model and the trained improved PSPNet semantic segmentation model to obtain a semantic segmentation image.

[0080] The disease identification module is used to determine the tunnel apparent disease identification result according to the obtained semantic segmentation image.

[0081] The above implementation process is only an example clearly described for this application, rather than a limitation on the implementation manner. For those of ordinary skill in the art, other different forms of changes or variations can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. And the obvious changes or variations derived therefrom are still within the protection scope of this application type.

Claims

1. A method for identifying tunnel surface defects, characterized in that: The following steps are involved: Step 1: preprocess the collected tunnel surface image set to obtain a preprocessed image set, and label the image set with a diseased area as the first training set; Step 2: input the first training set into the diseased area detection model for training, obtain a trained diseased area detection model, and obtain an output result of the diseased area detection model; Step 3, adding classification labels to the output results to obtain a classification label image set, using it as a second training set, and inputting the second training set into the improved PSPNet semantic segmentation model for training to obtain a trained improved PSPNet semantic segmentation model; Step 4, inputting the tunnel surface image to be detected into the trained diseased area detection model and the trained improved PSPNet semantic segmentation model in sequence to obtain a semantic segmentation image; Step 5: Determine the tunnel surface defect recognition result based on the obtained semantic segmentation image.

2. The tunnel apparent disease identification method according to claim 1, characterized in that: The diseased area detection model includes a Backbone module, a Neck module and a Head module connected in sequence; The Backbone module is used to extract multi-scale features of the input image, the Neck module is used to fuse the multi-scale features extracted by the Backbone module, and the Head module is used to classify and locate the multi-scale features fused by the Neck module; The Backbone module includes a Silence unit, a first Conv unit, a second Conv unit, a first GELAN unit, a first ADOWN unit, a second GELAN unit, a second ADOWN unit, a third GELAN unit, a third ADOWN unit, a fourth GELAN unit and a SPFFLAN unit connected in sequence; Among them, the Silence unit is used to assist the branch in calling the image input into the network, the first Conv unit and the second Conv unit are both used to downsample the input image, the first GELAN unit, the second GELAN unit, the third GELAN unit and the fourth GELAN unit are used to extract features of the input image to obtain a multi-scale feature map, and the SPFFLAN unit is used to perform multi-scale feature extraction on the multi-scale feature map.

3. The tunnel apparent disease identification method according to claim 2, characterized in that: The Neck module includes a third Conv unit, a first ASFF2 unit, a fifth GELAN unit, a first ASFF3 unit, and a sixth GELAN unit connected in sequence; a fourth Conv unit, a second ASFF2 unit, a seventh GELAN unit, a second ASFF3 unit, and an eighth GELAN unit connected in sequence; and a fifth Conv unit, a third ASFF3 unit, and a ninth GELAN unit which are sequentially connected; The output of the SPFFLAN unit is connected to the input of the fifth Conv unit; the output of the second GELAN unit is connected to the input of the third Conv unit; the output of the third GELAN unit is connected to the input of the fourth Conv unit; the third Conv unit is also connected to the first ASFF2 unit and the second ASFF2 unit respectively; the fourth Conv unit is also connected to the first ASFF2 unit and the second ASFF2 unit respectively; the sixth GELAN unit and the ninth GELAN unit are both connected to the eighth GELAN unit; The first ASFF2 unit and the second ASFF2 unit are both used for dynamically fusing features of two levels, and the first ASFF3 unit, the second ASFF3 unit and the third ASFF3 unit are both used for dynamically fusing features of three levels.

4. The tunnel apparent disease identification method according to claim 2, characterized in that: The Head module includes a first detector, a second detector, a third detector and three Concat modules, wherein the first detector, the second detector and the third detector each include a Conv-reg module and a Conv-cls module arranged in parallel, the input ends of the Conv-reg module and the Conv-cls module are both connected to the Neck module, and the output ends of the Conv-reg module and the Conv-cls module are both connected to a corresponding Concat module; The Conv-cls module is used to output the category of the disease, the Conv-reg module is used to output the location information of the disease, and the Concat module is used to splice the output results of the Conv-cls module and the Conv-reg module to obtain the output result of the disease area detection model.

5. The tunnel apparent disease identification method according to claim 1, characterized in that: The pretreatment described in step 1 specifically includes the following sub-steps: Step 1.1, using a multi-scale limited contrast adaptive histogram equalization algorithm to perform contrast enhancement on the tunnel surface image to obtain a contrast enhanced tunnel surface image; Step 1.2: The obtained contrast-enhanced tunnel surface image is subjected to smoothing denoising, gamma correction, random cropping, scaling, flipping and Mosaic processing in sequence to obtain a preprocessed image.

6. The tunnel apparent disease identification method according to claim 1, characterized in that: The improved PSPNet semantic segmentation model replaces the ResNet50 module of the original PSPNet semantic segmentation model with a RepVGG Backbone module. The RepVGG Backbone module is composed of a plurality of stacked RepVGG modules. Each of the RepVGG modules includes a 3x3 convolution layer, a 1x1 convolution branch layer and an identity mapping branch arranged in parallel.

7. The tunnel apparent disease identification method according to claim 1, characterized in that: The classification labels described in step 3 are numbers, where cracks are 0 and peeling is 1.

8. The tunnel apparent disease identification method according to claim 1, characterized in that: Determining the tunnel surface disease recognition result according to the obtained semantic segmentation image in step 4 includes determining the crack length and crack width, or the spalling area according to the semantic segmentation image.

9. The tunnel apparent disease identification method according to claim 8, characterized in that: The crack length is determined by a skeleton line extraction method, the crack width is determined by a maximum inscribed circle method, and the spalling area is determined by a connected domain area calculation method using Green's formula.

10. A tunnel apparent disease identification system, characterized in that: It includes data collection and preprocessing module, disease area detection model training module, improved PSPNet semantic segmentation model training module, detection module and disease identification module; The data acquisition and preprocessing module is used to preprocess the acquired tunnel surface image set to obtain a preprocessed image set, and label the preprocessed image set with a diseased area label as a first training set; The diseased area detection model training module is used to input the first training set into the diseased area detection model for training, obtain a trained diseased area detection model, and obtain an output result of the diseased area detection model; The improved PSPNet semantic segmentation model training module is used to add classification labels to the output results to obtain a classification label image set, which is used as a second training set, and the second training set is input into the improved PSPNet semantic segmentation model for training to obtain a trained improved PSPNet semantic segmentation model; The detection module is used to sequentially input the tunnel surface image to be detected into the trained diseased area detection model and the trained improved PSPNet semantic segmentation model to obtain a semantic segmentation image; The disease recognition module is used to determine the tunnel apparent disease recognition result according to the obtained semantic segmentation image.

Images