Tunnel lining defect detection method based on visual identification

Abstract

The invention discloses a tunnel lining defect detection method based on visual identification, and the method comprises the steps: obtaining a tunnel lining surface image through a camera, and generating a first image data set; extracting features according to the second image data set, and generating a first feature set by adopting a pre-trained lightweight convolutional neural network model; re-extracting features from the third image data set, and updating the lightweight convolutional neural network model by using a transfer learning method to obtain a second feature set; aiming at the second feature set, carrying out weighting processing on the disease area by adopting an attention mechanism to generate a third feature set; judging the disease type and severity from the third feature set by adopting a preset classifier to obtain a first detection result; and from the fourth feature set, the optimized classifier is adopted to re-judge the disease type and severity, and a final detection result is obtained. According to the invention, the accuracy and real-time performance of detection are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel lining defect detection, and in particular relates to a tunnel lining defect detection method based on visual recognition. Background Art

[0002] Tunnel engineering is a vital component of infrastructure construction, and its long-term safe operation is crucial for ensuring smooth traffic and public safety. However, tunnel linings, exposed to complex geological stresses and harsh environmental factors, are prone to cracks, spalling, leakage, and other defects, threatening structural stability and traffic safety. Traditional manual inspection methods are inefficient, labor-intensive, and limited by subjective judgment, making them difficult to guarantee accurate inspections. Existing automated inspection technologies, such as laser scanning and infrared imaging, rely on expensive equipment and have high deployment costs, making them difficult to widely apply in small and medium-sized projects. With the advancement of computer vision technology, intelligent inspection based on image recognition has gradually become a research hotspot, but its application in practical engineering still faces significant technical bottlenecks. Current solutions rely on traditional image processing methods that rely on manually designed feature extraction rules. These methods are poorly adaptable to complex environments such as changing lighting, occlusion, and noise, resulting in unstable inspection results. While general deep learning models perform well under ideal conditions, they require large amounts of labeled data for training. They are prone to overfitting in data-scarce tunnel inspection scenarios. Furthermore, model complexity makes inference speed insufficient to meet the requirements of real-time detection. While multi-sensor fusion systems can improve detection accuracy, they suffer from high hardware costs, complex data processing, and significant maintenance challenges. These limitations collectively point to core technical challenges. Tunnel defect detection requires high accuracy and real-time performance in complex environments, but insufficient data is a primary constraint. Due to the diverse morphologies of tunnel defects and the high cost of acquiring labeled data, models struggle to fully learn diverse defect characteristics, resulting in insufficient generalization in real-world scenarios. This data insufficiency further complicates the model's robustness to environmental variations. For example, in low light or rainy or foggy conditions, existing methods are prone to false or missed detections. This lack of robustness, in turn, creates real-time challenges, as complex models often sacrifice inference speed for accuracy, making them difficult to adapt to resource-constrained edge computing devices. Therefore, how to optimize model structure and data processing strategies to improve the detection system's adaptability to complex environments while maintaining high accuracy and real-time performance, given limited data, has become a key issue in the field of intelligent tunnel defect detection. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention proposes a tunnel lining defect detection method based on visual recognition, which improves the accuracy and real-time performance of detection.

[0004] To achieve the above objectives, the present invention provides a tunnel lining defect detection method based on visual recognition, comprising:

[0005] Acquiring a tunnel lining surface image to generate a first image data set;

[0006] performing equalization processing on the first image data set to enhance image contrast under low light conditions, thereby obtaining a second image data set;

[0007] Extracting features from the second image dataset and using a pre-trained lightweight convolutional neural network model to generate a first feature set;

[0008] If the confidence level of the disease feature in the first feature set is lower than a first preset threshold, processing the second image data set to generate a third image data set;

[0009] Re-extracting features based on the third image dataset, updating the lightweight convolutional neural network model, and obtaining a second feature set;

[0010] Based on the second feature set, weighting the diseased area to generate a third feature set;

[0011] Determine the type and severity of the disease using a preset classifier according to the third feature set to obtain a first detection result;

[0012] If the false positive rate of the first detection result is higher than a second preset threshold, updating the third feature set through images collected in real time by the edge device to generate a fourth feature set;

[0013] The optimized classifier is used according to the fourth feature set to re-judge the type and severity of the disease and obtain the tunnel lining defect detection result.

[0014] Technical effect of the present invention: The present invention discloses a method for detecting tunnel lining defects based on visual recognition, which obtains tunnel lining surface images through a camera, adopts adaptive histogram equalization processing to enhance image contrast under low-light conditions, and uses a lightweight convolutional neural network to extract features. When the feature confidence is insufficient, the model is optimized through data enhancement and transfer learning. Subsequently, the attention mechanism is used to weight the diseased area, and a classifier is used to determine the type and severity of the disease. If the false detection rate is too high, the feature set is updated in real time through an online learning method, and finally an accurate detection result is obtained. The present invention can effectively solve the problems of low light, insufficient feature extraction, high false detection rate, etc. in tunnel lining disease detection, improves the accuracy and real-time performance of detection, and provides reliable technical support for tunnel safety maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0016] Figure 1 The figure is a flow chart of a tunnel lining defect detection method based on visual recognition according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0018] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0019] like Figure 1 As shown, this embodiment provides a tunnel lining defect detection method based on visual recognition, including: acquiring a tunnel lining surface image to generate a first image data set;

[0020] performing equalization processing on the first image data set to enhance image contrast under low light conditions, thereby obtaining a second image data set;

[0021] Extracting features from the second image dataset and using a pre-trained lightweight convolutional neural network model to generate a first feature set;

[0022] If the confidence level of the disease feature in the first feature set is lower than a first preset threshold, processing the second image data set to generate a third image data set;

[0023] Re-extracting features based on the third image dataset, updating the lightweight convolutional neural network model, and obtaining a second feature set;

[0024] Based on the second feature set, weighting the diseased area to generate a third feature set;

[0025] Determine the type and severity of the disease using a preset classifier according to the third feature set to obtain a first detection result;

[0026] If the false positive rate of the first detection result is higher than a second preset threshold, updating the third feature set through images collected in real time by the edge device to generate a fourth feature set;

[0027] The optimized classifier is used according to the fourth feature set to re-judge the type and severity of the disease and obtain the tunnel lining defect detection result.

[0028] Furthermore, acquiring the tunnel lining surface image and generating the first image data set includes:

[0029] Acquire tunnel lining surface images using a camera at a preset acquisition frequency, and generate a first image sequence of an original image set including a timestamp when the ambient illumination meets a preset threshold;

[0030] If the original image set has uneven illumination, the first image sequence is processed using histogram equalization to generate a second image set with uniform image quality;

[0031] unifying the image size and color space of the second image set using an automated tool according to the image format of the second image set to generate the third image set suitable for data annotation;

[0032] The third image set is automatically annotated using preset annotation rules, and a timestamp and a location tag are added in combination with time synchronization information to generate a first image data set.

[0033] Specifically, in an embodiment of the present invention, tunnel lining surface images are acquired via a camera at a preset frequency. For example, five frames per second are captured to ensure that details of the tunnel surface are captured. Ambient lighting must meet a preset threshold, such as when the light intensity reaches 500 lux. A first image sequence is generated, including a set of original images with a timestamp. Timestamps are accurate to the millisecond, such as 2025-05-13 10:00:00.123, to facilitate subsequent data traceability. This approach ensures the temporal continuity of the image sequence, providing a reliable foundation for defect detection. For example, the original image sequence may have uneven illumination due to uneven distribution of light sources within the tunnel. Histogram equalization is used to process the first image sequence to enhance image contrast and generate the second image set. Specifically, the pixel grayscale value distribution is adjusted to enhance the clarity of dark details, such as by expanding the grayscale value range from 0-100 to 0-255. This method effectively eliminates lighting variations, improves image quality consistency, and provides high-quality input for subsequent processing. In one embodiment, the second image set must be formatted uniformly to generate the third image set suitable for data annotation. For example, automated tools are used to standardize image sizes to 1920x1080 pixels and convert color space to RGB for different camera output formats, such as JPEG or PNG. Format consistency checks ensure that all image parameters are consistent, preventing annotation tools from reporting errors due to format differences. This standardization process improves data processing compatibility and efficiency.

[0034] Furthermore, performing equalization processing on the first image dataset to enhance image contrast under low light conditions to obtain a second image dataset includes:

[0035] Redistributing pixel grayscale values ​​of the first image data set to generate a second image sequence with enhanced contrast, thereby obtaining the second image set;

[0036] Standardizing the image size and color space of the second image set to generate a third image set in a unified format;

[0037] According to a preset annotation rule and in combination with the time synchronization information, the third image set is automatically annotated, and a timestamp and a location tag are added to generate a fourth image set;

[0038] The contrast and clarity of the fourth image set are detected, and if the detection result meets a preset threshold, a second image data set is generated.

[0039] Specifically, in the embodiments of the present invention, when processing the first image dataset using an adaptive histogram equalization algorithm, the core focus is on optimizing for local illumination variations within the tunnel lining surface images. In principle, the algorithm divides the image into multiple small regions and adjusts the grayscale distribution of each region to enhance local contrast. For example, within a tunnel, due to varying light source angles, some areas may be too dark, with grayscale values ​​concentrated between 0 and 50. The algorithm remaps the grayscale values ​​in these areas to a range of 0 to 200, preserving detail while avoiding the over-enhancement that could result from global equalization. To address the texture characteristics of the tunnel lining surface, the region size can be set to 64x64 pixels, ensuring that details such as cracks are more easily discernible after enhancement. This method is particularly well-suited for complex lighting scenarios in tunnel environments. In one possible implementation, when normalizing the second image dataset, it is necessary to ensure that the image size and color space are consistent to accommodate subsequent annotation tools. For example, tunnel lining images may come from different camera models, with resolutions ranging from 1280x720 to 2560x1440. The automated tool resizes all images to a uniform 1920x1080 pixel resolution and converts the color space to RGB. Specifically, the conversion process checks for loss of detail due to compression. For example, JPEG formats can introduce compression artifacts, and the tool prioritizes lossless formats such as PNG. This standardized process ensures compatibility among annotation tools.

[0040] Furthermore, extracting features based on the second image dataset and using a pre-trained lightweight convolutional neural network model to generate a first feature set includes:

[0041] acquiring image data according to the second image data set, dividing each image into regions using an image segmentation algorithm, and generating a fifth image set including a target region;

[0042] Extracting features of the target region using a pre-trained convolutional neural network based on the fifth image set to generate a sixth feature set including regional features;

[0043] If the feature dimension of the sixth feature set is greater than a preset threshold, a principal component analysis algorithm is used to reduce the dimension of the feature to generate a seventh feature set with reduced dimension;

[0044] According to the seventh feature set, a clustering algorithm is used to group features to generate a first feature set.

[0045] Furthermore, generating the third image dataset includes:

[0046] If the confidence level of the disease feature in the first feature set is lower than a first preset threshold, performing a rotation transformation on the second image data set using an image processing algorithm to generate a fourth image set including a rotation angle change;

[0047] According to the fourth image set, each image is horizontally and vertically flipped using a flip transformation algorithm to generate a fifth image set containing flip features;

[0048] According to the fifth image set, a noise addition algorithm is used to inject random Gaussian noise into each image to generate a third image data set.

[0049] Furthermore, re-extracting features based on the third image dataset and updating the lightweight convolutional neural network model to obtain a second feature set includes:

[0050] Acquiring image data of the third image data set, dividing each image into regions using a preset image segmentation algorithm to generate a sixth image set comprising a plurality of sub-regions, and determining a segmented image data set;

[0051] According to the sixth image set, grayscale values ​​of each sub-region image are adjusted using a grayscale transformation algorithm to generate a seventh image set, thereby obtaining an image data set after grayscale adjustment;

[0052] If the feature significance of the sub-region images in the seventh image set is lower than a second preset threshold, performing contrast enhancement on each image using a histogram equalization algorithm to generate an eighth image set, and determining the enhanced image data set;

[0053] According to the eighth image set, a transfer learning method is used to fine-tune the parameters of a pre-established lightweight convolutional neural network to generate a second feature set.

[0054] Furthermore, based on the second feature set, weighted processing is performed on the diseased area to generate a third feature set including:

[0055] acquiring image feature data based on the second feature set, dividing each image into regions using a preset image segmentation algorithm to generate a ninth image set comprising a plurality of sub-regions, and determining a segmented image data set;

[0056] Calculating the texture features of each sub-region image using a gray level co-occurrence matrix algorithm based on the ninth image set to generate a tenth image set, thereby obtaining an image dataset with enhanced texture features;

[0057] If the significance of the texture features of the sub-region images in the tenth image set is lower than a third preset threshold, adjusting the pixel grayscale distribution of each image by a histogram equalization algorithm to generate an eleventh image set, and determining the adjusted image data set;

[0058] Based on the eleventh image set, an attention mechanism is used to perform weighted processing on the features of the diseased area to generate a third feature set.

[0059] Specifically, in an embodiment of the present invention, for images of water seepage areas taken in low light, the texture features may be blurred due to insufficient lighting. Histogram equalization can redistribute grayscale values ​​so that the grayscale range of water seepage traces is expanded from 90-120 to 70-190, thereby enhancing the visibility of the texture. This adjustment can improve the analytical adaptability of images under different lighting conditions. Based on the eleventh image set, the attention mechanism is used to weight the features of the diseased area to generate a third feature set. The attention mechanism calculates the weight W through a normalization function to highlight the features of cracks or water seepage areas. For example, for crack images, the attention weight can be calculated based on texture and edge features. The weight of the crack area may be 0.8, while the weight of the background area may be 0.2. In one embodiment, for water seepage areas, the attention mechanism can prioritize enhancing sub-regions with complex textures, making their features more prominent in the third feature set. This weighted processing can optimize the model's attention to key defects. It can be understood that the above process gradually optimizes image features through segmentation, texture analysis, grayscale adjustment, and attention weighting. For example, for a 0.5 mm wide crack, region growing segmentation accurately locates the crack, the gray-level co-occurrence matrix extracts linear textures, histogram equalization enhances contrast in low-light images, and the attention mechanism highlights the crack's features. This progressive processing provides higher-quality feature input for defect detection, improving the reliability of subsequent analysis.

[0060] Furthermore, the first detection result is obtained by using a preset classifier to determine the type and severity of the disease based on the third feature set:

[0061] Obtaining image feature data from the third feature set, finely dividing the diseased area using a preset regional clustering algorithm, generating a fourth image set containing multiple sub-areas, and determining a segmented image data set;

[0062] Calculating the texture features of each sub-region image using a local binary pattern algorithm based on the fourth image set to generate a fifth image set, thereby obtaining an image dataset with enhanced texture features;

[0063] If the significance of the texture features of the sub-region images in the fifth image set is lower than a fourth preset threshold, adjusting the grayscale distribution of each image by an adaptive histogram equalization algorithm to generate a sixth image set, and determining the adjusted image data set;

[0064] Based on the sixth image set, a support vector machine classifier is used to classify the features of the diseased area to generate a first detection result.

[0065] Furthermore, if the false detection rate of the first detection result is higher than a second preset threshold, the third feature set is updated using the image collected in real time by the edge device, and the fourth feature set is generated, including:

[0066] According to the image data collected in real time by the edge device, a preset image preprocessing algorithm is used to perform denoising and grayscale normalization on the image data to generate a first image set and obtain a standardized image data set;

[0067] For the first image set, an optical flow algorithm is used to calculate the motion trajectory of pixels in the image sequence to generate a second image set, and an image data set with enhanced motion features is determined;

[0068] If the motion feature significance of the second image set is lower than a third preset threshold, dividing the second image set into regions using an adaptive threshold segmentation algorithm to generate a third image set, and determining the image data set after the region segmentation;

[0069] According to the third image set, an online learning method is adopted to update the weight parameters of the third feature set through a gradient descent algorithm to generate a fourth feature set.

[0070] Specifically, in this embodiment of the present invention, a Gaussian filter algorithm is used to denoise the collected tunnel images, with a filter kernel size of 3x3 pixels. This smooths out noise while preserving crack edge features. Grayscale normalization normalizes the image grayscale values ​​to a range of 0-255 to ensure consistency in subsequent processing. The resulting first image set provides a clear, standardized data foundation for subsequent analysis. For this first image set, an optical flow algorithm is used to calculate the motion trajectories of pixels in the image sequence, generating a second image set. The optical flow algorithm analyzes the brightness changes of pixels in consecutive frames to calculate the displacement vector V(x, y) of the pixels along the time axis. Specifically, during tunnel inspection, when a moving camera captures the lining surface, the optical flow algorithm can capture the subtle displacements of crack regions caused by camera movement. For example, crack pixels may exhibit linear motion along the crack direction, with a displacement of approximately 2-5 pixels per frame, while background regions exhibit more random displacements. This motion signature enhances the ability to distinguish crack regions. If the significance of the motion signature in the second image set falls below a preset threshold T3, for example, if the modulus of the displacement vector in the crack region is less than 1 pixel per frame, the image is segmented using an adaptive threshold segmentation algorithm to generate a third image set. Adaptive threshold segmentation dynamically determines the segmentation threshold based on local pixel grayscale and motion characteristics.

[0071] Furthermore, the optimized classifier is used to re-judge the type and severity of the defect based on the fourth feature set, and the tunnel lining defect detection results obtained include:

[0072] Acquire feature vector data from the fourth feature set, and normalize the feature vector using a preset feature normalization algorithm to obtain a fifth feature set;

[0073] Based on the fifth feature set, a support vector machine algorithm is used to perform classification training to generate a first classification result and determine a preliminary disease type;

[0074] If the confidence level of the first classification result is lower than a fourth preset threshold, a second analysis is performed on the fifth feature set using a logical reasoning algorithm, and feature patterns are matched based on a preset disease feature library to generate a second classification result to determine a more accurate disease type.

[0075] According to the second classification result, a weighted scoring algorithm is used to calculate the severity of the disease and obtain the tunnel lining defect detection result.

[0076] Specifically, in tunnel inspection, the fifth feature set includes crack width features and grayscale features of water seepage areas. A support vector machine constructs a classification hyperplane to distinguish between cracks, water seepage, and other defects. In one embodiment, when the crack width feature value is greater than 0.5 and the water seepage grayscale feature value is less than 0.3, the classification result is determined to be a crack defect with a confidence level of 0.85. If the confidence level falls below a preset threshold T4, such as 0.7, further analysis is required. It should be noted that support vector machines improve classification robustness by maximizing the classification margin, making them particularly suitable for tunnel environments with complex feature distributions. When the confidence level of the first classification result is insufficient, a logical inference algorithm performs a secondary analysis based on a pre-set defect feature library. This defect feature library contains patterns such as the geometry of cracks and the color distribution of water seepage. The logical inference algorithm matches the crack width and water seepage grayscale features in the fifth feature set to more accurately determine the defect type. For example, if the crack width feature value is 0.6 and exhibits a linear distribution, the inference result may be determined to be a structural crack rather than a surface crack. This method overcomes the limitations of support vector machines in low-confidence scenarios through rule matching. Based on the second classification results, a weighted scoring algorithm was used to calculate the severity of the disease.

[0077] The present invention discloses a method for detecting tunnel lining defects based on visual recognition. The tunnel lining surface image is acquired by a camera, and adaptive histogram equalization is used to enhance the image contrast under low-light conditions. A lightweight convolutional neural network is used to extract features. When the feature confidence is insufficient, the model is optimized through data enhancement and transfer learning. Subsequently, the attention mechanism is used to weight the diseased area, and a classifier is used to determine the type and severity of the disease. If the false detection rate is too high, the feature set is updated in real time through an online learning method, and an accurate detection result is finally obtained. The present invention can effectively solve the problems of low light, insufficient feature extraction, high false detection rate, etc. in tunnel lining disease detection, improve the accuracy and real-time performance of detection, and provide reliable technical support for tunnel safety maintenance.

[0078] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A tunnel lining defect detection method based on visual recognition, characterized in that: include: Acquiring a tunnel lining surface image to generate a first image data set; performing equalization processing on the first image data set to enhance image contrast under low light conditions, thereby obtaining a second image data set; Extracting features from the second image dataset and using a pre-trained lightweight convolutional neural network model to generate a first feature set; If the confidence level of the disease feature in the first feature set is lower than a first preset threshold, processing the second image data set to generate a third image data set; Re-extracting features based on the third image dataset, updating the lightweight convolutional neural network model, and obtaining a second feature set; Based on the second feature set, weighting the diseased area to generate a third feature set; Determine the type and severity of the disease using a preset classifier according to the third feature set to obtain a first detection result; If the false positive rate of the first detection result is higher than a second preset threshold, updating the third feature set through images collected in real time by the edge device to generate a fourth feature set; The optimized classifier is used according to the fourth feature set to re-judge the type and severity of the disease and obtain the tunnel lining defect detection result.

2. The tunnel lining defect detection method based on visual recognition according to claim 1, characterized in that: Acquiring a tunnel lining surface image and generating a first image data set includes: Acquire tunnel lining surface images using a camera at a preset acquisition frequency, and generate a first image sequence of an original image set including a timestamp when the ambient illumination meets a preset threshold; If the original image set has uneven illumination, the first image sequence is processed using histogram equalization to generate a second image set with uniform image quality; unifying the image size and color space of the second image set using an automated tool according to the image format of the second image set to generate the third image set suitable for data annotation; The third image set is automatically annotated using preset annotation rules, and a timestamp and a location tag are added in combination with time synchronization information to generate a first image data set.

3. The tunnel lining defect detection method based on visual recognition according to claim 1, characterized in that: Performing equalization processing on the first image dataset to enhance image contrast under low light conditions to obtain a second image dataset includes: Redistributing pixel grayscale values ​​of the first image data set to generate a second image sequence with enhanced contrast, thereby obtaining the second image set; Standardizing the image size and color space of the second image set to generate a third image set in a unified format; According to a preset annotation rule and in combination with the time synchronization information, the third image set is automatically annotated, and a timestamp and a location tag are added to generate a fourth image set; The contrast and clarity of the fourth image set are detected, and if the detection result meets a preset threshold, a second image data set is generated.

4. The tunnel lining defect detection method based on visual recognition according to claim 1, characterized in that: Extracting features from the second image dataset and using a pre-trained lightweight convolutional neural network model to generate a first feature set includes: acquiring image data according to the second image data set, dividing each image into regions using an image segmentation algorithm, and generating a fifth image set including a target region; Extracting features of the target region using a pre-trained convolutional neural network based on the fifth image set to generate a sixth feature set including regional features; If the feature dimension of the sixth feature set is greater than a preset threshold, a principal component analysis algorithm is used to reduce the dimension of the feature to generate a seventh feature set with reduced dimension; According to the seventh feature set, a clustering algorithm is used to group features to generate a first feature set.

5. The tunnel lining defect detection method based on visual recognition according to claim 1, characterized in that: Generating the third image dataset includes: If the confidence level of the disease feature in the first feature set is lower than a first preset threshold, performing a rotation transformation on the second image data set using an image processing algorithm to generate a fourth image set including a rotation angle change; According to the fourth image set, each image is horizontally and vertically flipped using a flip transformation algorithm to generate a fifth image set containing flip features; According to the fifth image set, a noise addition algorithm is used to inject random Gaussian noise into each image to generate a third image data set.

6. The tunnel lining defect detection method based on visual recognition according to claim 1, characterized in that: Re-extracting features based on the third image dataset and updating the lightweight convolutional neural network model to obtain a second feature set includes: Acquiring image data of the third image data set, dividing each image into regions using a preset image segmentation algorithm to generate a sixth image set comprising a plurality of sub-regions, and determining a segmented image data set; According to the sixth image set, grayscale values ​​of each sub-region image are adjusted using a grayscale transformation algorithm to generate a seventh image set, thereby obtaining an image data set after grayscale adjustment; If the feature significance of the sub-region images in the seventh image set is lower than a second preset threshold, performing contrast enhancement on each image using a histogram equalization algorithm to generate an eighth image set, and determining the enhanced image data set; According to the eighth image set, a transfer learning method is used to fine-tune the parameters of a pre-established lightweight convolutional neural network to generate a second feature set.

7. The method for detecting tunnel lining defects based on visual recognition according to claim 1, characterized in that: Based on the second feature set, weighted processing is performed on the diseased area to generate a third feature set including: acquiring image feature data based on the second feature set, dividing each image into regions using a preset image segmentation algorithm to generate a ninth image set comprising a plurality of sub-regions, and determining a segmented image data set; Calculating the texture features of each sub-region image using a gray level co-occurrence matrix algorithm based on the ninth image set to generate a tenth image set, thereby obtaining an image dataset with enhanced texture features; If the significance of the texture features of the subregion images in the tenth image set is lower than a third preset threshold, adjusting the pixel grayscale distribution of each image by a histogram equalization algorithm to generate an eleventh image set, and determining the adjusted image data set; Based on the eleventh image set, an attention mechanism is used to perform weighted processing on the features of the diseased area to generate a third feature set.

8. The method for detecting tunnel lining defects based on visual recognition according to claim 1, characterized in that: Determining the disease type and severity using a preset classifier based on the third feature set to obtain a first detection result includes: Obtaining image feature data from the third feature set, finely dividing the diseased area using a preset regional clustering algorithm, generating a fourth image set containing multiple sub-areas, and determining a segmented image data set; Calculating the texture features of each sub-region image using a local binary pattern algorithm based on the fourth image set to generate a fifth image set, thereby obtaining an image dataset with enhanced texture features; If the significance of the texture features of the sub-region images in the fifth image set is lower than a fourth preset threshold, adjusting the grayscale distribution of each image by an adaptive histogram equalization algorithm to generate a sixth image set, and determining the adjusted image data set; Based on the sixth image set, a support vector machine classifier is used to classify the features of the diseased area to generate a first detection result.

9. The tunnel lining defect detection method based on visual recognition according to claim 1, characterized in that: If the false positive rate of the first detection result is higher than the second preset threshold, the third feature set is updated through the image collected in real time by the edge device to generate the fourth feature set, including: According to the image data collected in real time by the edge device, a preset image preprocessing algorithm is used to perform denoising and grayscale normalization on the image data to generate a first image set and obtain a standardized image data set; For the first image set, an optical flow algorithm is used to calculate the motion trajectory of pixels in the image sequence to generate a second image set, and an image data set with enhanced motion features is determined; If the motion feature significance of the second image set is lower than a third preset threshold, dividing the second image set into regions using an adaptive threshold segmentation algorithm to generate a third image set, and determining the image data set after the region segmentation; According to the third image set, an online learning method is adopted to update the weight parameters of the third feature set through a gradient descent algorithm to generate a fourth feature set.

10. The tunnel lining defect detection method based on visual recognition according to claim 1, characterized in that: The optimized classifier is used based on the fourth feature set to re-judge the type and severity of the defect, and the tunnel lining defect detection results obtained include: Acquire feature vector data from the fourth feature set, and normalize the feature vector using a preset feature normalization algorithm to obtain a fifth feature set; Based on the fifth feature set, a support vector machine algorithm is used to perform classification training to generate a first classification result and determine a preliminary disease type; If the confidence level of the first classification result is lower than a fourth preset threshold, a second analysis is performed on the fifth feature set using a logical reasoning algorithm, and feature patterns are matched based on a preset disease feature library to generate a second classification result to determine a more accurate disease type. According to the second classification result, a weighted scoring algorithm is used to calculate the severity of the disease and obtain the tunnel lining defect detection result.

Images