A tunnel disease identification and detection method and system based on deep learning

Abstract

This invention belongs to the technical field of tunnel defect detection and discloses a tunnel defect identification and detection method and system based on deep learning. The method involves obtaining historical tunnel defect images, preprocessing them to obtain a training dataset, training a neural network model using the training dataset to obtain a target recognition model, obtaining a tunnel image to be inspected, preprocessing the image, and then inputting the preprocessed image into the target recognition model to obtain a target image containing defects. Finally, the defect areas are extracted from the target image and labeled. This technical solution can improve the training accuracy of the target recognition model and the accuracy of tunnel defect detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel disease detection, and specifically relates to a tunnel disease identification and detection method and system based on deep learning. Background Art

[0002] With the development of society, the scale of tunnel construction is increasing rapidly. At the same time, many projects face harsh construction and maintenance environments, high technical challenges, and numerous safety risks, placing higher demands on maintenance management. Currently, the main methods for detecting tunnel hazards include manual observation, digital image processing and recognition, and radar detection. Digital image processing and recognition methods, especially those based on deep learning, offer advantages such as high real-time performance, high accuracy, and high robustness. They have become the mainstream method for tunnel crack detection and are of great significance to current highway maintenance management.

[0003] A similar prior art includes a Chinese patent application with publication number CN117994710A, which discloses a method for rapid identification of tunnel cracks based on deep learning. The method comprises the following steps: acquiring a tunnel image to obtain a tunnel image dataset; preprocessing the tunnel image dataset to convert the image into an image with consistent format, size, and clarity to obtain an input dataset; dividing the input dataset into a training set and a test set according to a certain ratio, manually annotating the training set and inputting it into a model for training to obtain a trained model; inputting the test set into the trained model and outputting labeled images to obtain a crack dataset; restoring the images in the crack dataset to their original image size and extracting the crack contour; and detecting whether the crack contour size exceeds a threshold. If so, the actual length and width of the crack are calculated and output; otherwise, no data is output. There is also a Chinese patent application with publication number CN108229461A, which discloses a method for rapid identification of tunnel cracks based on deep learning. It creates a deep learning image training set; trains a deep convolutional neural network model; uses the trained convolutional neural network model to detect the image to be detected and outputs a predicted label image; and outputs the detection results based on the predicted label image, including image category, crack coordinate information, and pixel-level width and length values ​​of the crack.

[0004] The aforementioned two methods fail to maintain high image clarity when generating training data or testing images to be inspected, resulting in low accuracy in the recognition model and results. Therefore, providing a deep learning-based tunnel defect recognition and detection method and system to improve the accuracy of target recognition model training and increase the accuracy of tunnel defect detection is an urgent problem to be solved. Summary of the Invention

[0005] In response to the above-mentioned technical problems, the present invention provides a tunnel disease identification and detection method and system based on deep learning.

[0006] In a first aspect, the present invention provides a method for identifying and detecting tunnel defects based on deep learning, the method comprising the following steps:

[0007] Obtain historical tunnel disease images, pre-process the historical disease images, and obtain a training data set;

[0008] The training data set is used to train the neural network model to obtain the target recognition model;

[0009] Acquire an image of the tunnel to be inspected, preprocess the image, and then input the preprocessed image into a target recognition model to obtain a target image with defects.

[0010] Extract the diseased area from the target image and mark it.

[0011] Specifically, the preprocessing includes a first preprocessing, and the first preprocessing includes:

[0012] Step 111: extract any first image to be processed, and divide the first image to be processed into a plurality of sub-images according to a first preset size, wherein the first image to be processed is a historical disease image or a tunnel image to be inspected;

[0013] Step 112: extract any sub-image, determine the image feature value of any sub-image, and determine whether the image feature value is greater than or equal to a first preset value. If so, proceed to step 113; if not, proceed to step 114.

[0014] Step 113: Input any sub-image into the first network model to obtain an interference information image, and generate a first sub-image based on the any sub-image and the interference information image;

[0015] Step 114: Input any sub-image into the second network model to obtain a first sub-image;

[0016] Step 115: After traversing all sub-images, reassemble all first sub-images to generate a first image;

[0017] Step 116: After traversing all first images to be processed, obtain the first pre-processing result.

[0018] Specifically, the image feature value includes a first image feature value and / or a second image feature value, and the first image feature value and the second image feature value are determined by:

[0019] Convert any sub-image into a frequency domain analysis image, extract any frequency domain point in the frequency domain analysis image, and define any frequency domain point as a first-class point when the amplitude of any frequency domain point is greater than a second preset value; otherwise, define any frequency domain point as a second-class point. After traversing all points in the frequency domain analysis image, calculate the average amplitude of all first-class points and define it as the first image eigenvalue;

[0020] Extract any pixel point in any sub-image, perform weighted summation on the intensity value of each color channel at any pixel point, and define it as the first point eigenvalue at any pixel point. After traversing all pixels in any sub-image, calculate the average value of all first point eigenvalues ​​and define it as the second image eigenvalue.

[0021] Specifically, in step 115, when any two adjacent first sub-images are reassembled, the following steps are performed:

[0022] Determine whether there is an intersection area between the two first sub-images. If not, directly stitch the two first sub-images together. If so, cut the two first sub-images into a first stitched image, a second stitched image, a third stitched image, and a fourth stitched image based on the intersection area, where the second stitched image and the third stitched image are stitched images corresponding to the intersection area.

[0023] Extracting a first pixel value at any pixel point on the second stitched image and a second pixel value at any pixel point on the third stitched image, taking an average of the first pixel value and the second pixel value as the pixel value at any pixel point on the fifth stitched image, and generating a fifth stitched image after traversing all pixel points on the second stitched image;

[0024] The first stitched image, the fifth stitched image, and the fourth stitched image are stitched together.

[0025] Specifically, the preprocessing further includes a second preprocessing, and the second preprocessing includes:

[0026] Extracting any second image to be processed, scaling the physical size of any second image to be processed according to a first preset ratio to generate a second image, wherein the second image to be processed is a historical disease image, an image of a tunnel to be inspected, or the first image;

[0027] Inputting the second image into the third network model to obtain a third image, wherein the third image has the same physical size as the second image and a clarity index value of the third image is greater than a clarity index value of the second image;

[0028] After traversing all the second images to be processed, the second preprocessing result is obtained.

[0029] Specifically, the method for generating the third network model is:

[0030] Step 121: Divide the shooting area according to a second preset ratio to obtain multiple sub-areas;

[0031] Step 122: Use an image acquisition device to acquire a first overall image of the shooting area, and use an image acquisition device to acquire a regional image of each sub-area, and integrate all the regional images to acquire a second overall image;

[0032] Step 123: Scale the physical size of the first overall image according to a first preset ratio to generate a third overall image, use the third overall image and the second overall image to train a specific generative adversarial network, and define a generator that changes the third overall image in the direction of the second overall image as a third network model.

[0033] Specifically, extracting the diseased area from the target image and labeling it include:

[0034] Step 41: Divide the target image into a plurality of first partial images according to a second preset size;

[0035] Step 42: extract any first partial image, perform a weighted sum of the intensity values ​​of each color channel at any pixel point in any first partial image, and define the sum as the second point eigenvalue at any pixel point. After traversing all pixels in any first partial image, calculate a first statistic of all second point eigenvalues, and determine whether the first statistic is within a first preset range. If so, mark any first partial image as a first category area.

[0036] Step 43: After traversing all first partial images, define the unlabeled first partial image as the second partial image, extract any second partial image, determine the maximum intensity value and the minimum intensity value of all color channels at any pixel point in any second partial image, calculate the ratio of the difference between the maximum intensity value and the minimum intensity value to the maximum intensity value, and define it as the third point eigenvalue at any pixel point. After traversing all pixels in any second partial image, calculate the second statistical value of all third point eigenvalues, determine whether the second statistical value is within a second preset range, and if so, label any second partial image as a second category area.

[0037] Step 44: After traversing all second partial images, define the unlabeled second partial images as third partial images, extract any third partial image, and determine whether the second statistical value corresponding to any third partial image is within a third preset range. If so, label any third partial image as a third category area.

[0038] Step 45: After traversing all third partial images, define the unlabeled third partial images as fourth partial images, extract any fourth partial image, determine the texture statistical feature value of any fourth partial image, and judge whether the texture statistical feature value is within a fourth preset range. If so, label any fourth partial image as a fourth category area; if not, label any fourth partial image as a fifth category area.

[0039] Specifically, the contours of the sequentially connected fourth category areas in the target image are extracted and marked as diseased areas.

[0040] In a second aspect, the present invention further provides a tunnel disease identification and detection system based on deep learning, the system comprising: a first acquisition module, a preprocessing module, a model training module, a second acquisition module, an image detection module, and a disease extraction module;

[0041] The first acquisition module is used to acquire historical tunnel disease images and input the historical disease images into the preprocessing module;

[0042] The preprocessing module is used to preprocess historical disease images and obtain training data sets;

[0043] A model training module is used to train a neural network model using a training data set to obtain a target recognition model;

[0044] A second acquisition module is used to acquire the image of the tunnel to be detected and input the image of the tunnel to be detected into the preprocessing module;

[0045] The preprocessing module is further used to preprocess the tunnel image to be detected and obtain the preprocessed tunnel image to be detected;

[0046] An image detection module is used to input the pre-processed tunnel image to be inspected into the target recognition model to obtain the target image with defects;

[0047] The disease extraction module is used to extract the diseased area from the target image and mark it.

[0048] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0049] 1. Before model training, preprocessing the historical disease images used for training can improve the quality of training data, reduce noise and interference in the images, make image features more obvious, reduce overfitting, and improve the training efficiency, accuracy, and robustness of the target recognition model. At the same time, preprocessing the tunnel images to be detected can improve the accuracy and efficiency of the recognition results.

[0050] 2. By dividing the target image into multiple local images and classifying and labeling the local images according to the image feature values ​​and preset ranges, the refinement of disease detection is improved and the interpretability and usability of the recognition results are increased. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0052] Figure 1 This is a flow chart of a tunnel disease identification and detection method based on deep learning of the present invention;

[0053] Figure 2 A schematic diagram of the structure of a specific generative adversarial network according to an embodiment of the present invention;

[0054] Figure 3 This is a modular schematic diagram of a tunnel disease identification and detection system based on deep learning of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the specific embodiments described herein are only used to explain the present invention and are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0056] It should be noted that if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0057] Figure 1 FIG2 is a flow chart of an embodiment of a method for identifying and detecting tunnel defects based on deep learning provided by the present invention. The flow chart specifically includes the following steps:

[0058] Step 1: Obtain historical tunnel disease images, preprocess the historical disease images, and obtain a training data set.

[0059] The tunnel historical disease images and the tunnel images to be inspected are images of the tunnel interior taken by mobile inspection equipment, images of the tunnel interior taken by fixed inspection equipment, or images of the tunnel interior taken by drones, etc.

[0060] Before preprocessing the image, the image is converted into an image with consistent format and size.

[0061] Different road sections and different time periods have different lighting conditions and environments. In order to improve the accuracy and efficiency of model training, the images to be trained on the model are preprocessed.

[0062] Specifically, the preprocessing includes a first preprocessing, and the first preprocessing includes:

[0063] Step 111 : extract any first image to be processed, and divide the first image to be processed into a plurality of sub-images according to a first preset size, wherein the first image to be processed is a historical disease image or a tunnel image to be inspected.

[0064] Step 112 , extract any sub-image, determine the image feature value of any sub-image, and judge whether the image feature value is greater than or equal to a first preset value. If so, proceed to step 113 ; otherwise, proceed to step 114 .

[0065] Step 113: Input any sub-image into the first network model, obtain an interference information image, and generate a first sub-image based on the any sub-image and the interference information image.

[0066] Step 114: Input any sub-image into the second network model to obtain the first sub-image.

[0067] Step 115: After traversing all sub-images, all first sub-images are reassembled to generate a first image.

[0068] Step 116: After traversing all first images to be processed, obtain the first pre-processing result.

[0069] The first preset size and the first preset value are set according to the experience of those skilled in the art or according to actual application scenarios, and are not limited in the embodiments of the present application.

[0070] Since the first image to be processed is large in size and the noise interference conditions in various regions thereof are inconsistent, in order to improve the accuracy of the first preprocessing, the first image to be processed is divided into multiple sub-images, and each sub-image is processed separately.

[0071] The first network model is a neural network model for estimating noise components from an image. By inputting a sub-image into the first network model, an interference information image in the sub-image can be extracted. By subtracting the interference information image data from the input sub-image data, the first sub-image data with noise removed can be obtained.

[0072] The second network model is a neural network model that estimates the image data after noise removal from the image. By inputting the sub-image into the second network model, the first sub-image data after noise removal can be obtained.

[0073] If the image feature value of any sub-image is greater than or equal to a first preset value, it indicates that the noise level of the sub-image is high, or the noise has a significant impact on the sub-image. In this case, the first network model is used to pre-process the sub-image. Conversely, if the image feature value of the sub-image is lower, or the noise has a minor impact on the sub-image, the second network model is used to pre-process the sub-image. This effectively improves the denoising effect.

[0074] Specifically, the image feature value includes a first image feature value and / or a second image feature value, and the first image feature value and the second image feature value are determined by:

[0075] Convert any sub-image into a frequency domain analysis image, extract any frequency domain point in the frequency domain analysis image, and define any frequency domain point as a first-class point when the amplitude at any frequency domain point is greater than a second preset value; otherwise, define any frequency domain point as a second-class point. After traversing all points in the frequency domain analysis image, calculate the average amplitude of all first-class points and define it as the first image eigenvalue.

[0076] Extract any pixel point in any sub-image, perform weighted summation on the intensity value of each color channel at any pixel point, and define it as the first point eigenvalue at any pixel point. After traversing all pixels in any sub-image, calculate the average value of all first point eigenvalues ​​and define it as the second image eigenvalue.

[0077] The second preset value is set according to the experience of those skilled in the art or according to actual application scenarios, and is not limited in the embodiments of the present application.

[0078] When performing weighted summation on the intensity value of each color channel at any pixel point, the weight value of each color channel may be the same or different.

[0079] Different image characteristic values ​​correspond to different first preset values.

[0080] Specifically, in step 115, when any two adjacent first sub-images are reassembled, the following steps are performed:

[0081] Determine whether there is an intersection area between the two first sub-images. If not, directly stitch the two first sub-images together. If so, cut the two first sub-images into a first stitched image, a second stitched image, a third stitched image, and a fourth stitched image based on the intersection area, wherein the second stitched image and the third stitched image are stitched images corresponding to the intersection area.

[0082] A first pixel value at any pixel point on the second stitched image and a second pixel value at any pixel point on the third stitched image are extracted, and an average of the first pixel value and the second pixel value is used as the pixel value at any pixel point on the fifth stitched image. After traversing all pixel points on the second stitched image, the fifth stitched image is generated.

[0083] The first stitched image, the fifth stitched image, and the fourth stitched image are stitched together.

[0084] When dividing the first image to be processed, there may or may not be an intersection area between adjacent sub-images. If there is an intersection area, since the first pre-processing is performed on the two adjacent sub-images separately, the processing results of the intersection area may be different. Directly overlapping the two sub-images may cause interference such as artifacts in the intersection area, causing image distortion. Therefore, it is necessary to first perform special processing on the image in the intersection area to ensure that the spliced ​​image is smooth.

[0085] Specifically, the preprocessing further includes a second preprocessing, and the second preprocessing includes:

[0086] Extract any second image to be processed, and scale the physical size of any second image to be processed according to a first preset ratio to generate a second image, wherein the second image to be processed is a historical disease image, an image of a tunnel to be inspected, or the first image.

[0087] The second image is input into a third network model to obtain a third image, wherein the third image has the same physical size as the second image and a clarity index value of the third image is greater than the clarity index value of the second image.

[0088] After traversing all the second images to be processed, the second preprocessing result is obtained.

[0089] The first preset ratio is set according to the experience of those skilled in the art or according to actual application scenarios, and is not limited in the present embodiment. For example, the first preset ratio is 10, and the length and width of the second image to be processed are respectively magnified by 10 times.

[0090] Clarity refers to the pixel density. The greater the pixel density, the finer the image and the higher the realism.

[0091] Specifically, the method for generating the third network model is:

[0092] Step 121: Divide the shooting area according to a second preset ratio to obtain multiple sub-areas.

[0093] Step 122: Use an image acquisition device to acquire a first overall image of the shooting area, and use an image acquisition device to acquire a regional image of each sub-area, and integrate all the regional images to acquire a second overall image.

[0094] Step 123: Scale the physical size of the first overall image according to a first preset ratio to generate a third overall image, use the third overall image and the second overall image to train a specific generative adversarial network, and define a generator that changes the third overall image in the direction of the second overall image as a third network model.

[0095] The second preset ratio is set according to the experience of those skilled in the art or according to actual application scenarios, and is not limited in the embodiments of the present application.

[0096] Preferably, the second preset ratio is the inverse of the first preset ratio. If the first preset ratio is 10, the second preset ratio is 1 / 10. If the area of ​​the shooting area is 100×100, the area of ​​each sub-area is 10×10.

[0097] The second image to be processed, the first overall image and the regional image are acquired using the same image acquisition device. The first overall image and any regional image are identical in digital size (i.e., number of pixels). The shooting area of ​​the second overall image acquired by integrating the regional images is the same as that of the first overall image, but the number of pixels is 100 times that of the first overall image. The first overall image is enlarged according to a first preset ratio to acquire a third overall image so that the number of pixels is consistent with that of the second overall image. However, the information contained in each pixel of the third overall image becomes less. Although the image size (number of pixels) has increased, its clarity index value is lower than that of the second overall image.

[0098] By using the third overall image and the second overall image to train a specific generative adversarial network, a network model that improves the image clarity index value can be obtained.

[0099] by Figure 2The technical solution of the present invention is described by taking as an example, wherein G and F are generators, and D1 and D2 are discriminators. The first step is to use the third whole image P11 and the second whole image P21 to learn the parameters of the discriminator D1 and the discriminator D2 respectively, and at the same time use the third whole image P11 and the second whole image P21 to learn the parameters of the generator G and the generator F respectively, and obtain the output images P12 and P22 respectively; the second step is to use the image P21 and the image P12 to learn the parameters of the discriminator D2, so that the value of the first adversarial loss function between the generator G and the discriminator D2 is maximized, and then use the image P11 to learn the parameters of the generator G, so that the value of the first adversarial loss function between the generator G and the discriminator D2 is minimized. At the same time, use the image P11 and the image P22 to learn the parameters of the discriminator D1, so that the value of the second adversarial loss function between the generator F and the discriminator D1 is maximized, and then use the image P21 to learn the parameters of the generator F, so that the value of the second adversarial loss function between the generator F and the discriminator D1 is minimized. Step 3: Calculate the cycle consistency loss for images P11 and P13, and learn generators G and F to minimize the cycle consistency loss. Then calculate the cycle consistency loss for images P21 and P23, and learn generators G and F to minimize the cycle consistency loss. Step 4: Construct a loss objective function based on the first adversarial loss function, the second adversarial loss function, and the cycle consistency loss. Repeat steps 2 to 3 until the loss objective function converges. The trained generator G is used as the third network model.

[0100] According to the technical solution of the present invention, the third network model can be generated using less training data, which reduces the cost and difficulty of obtaining training data.

[0101] The first pretreatment may be performed first and then the second pretreatment, or the second pretreatment may be performed first and then the first pretreatment, or only one of the first pretreatment and the second pretreatment may be performed, and the technical solution of the present invention is not limited to this.

[0102] Step 2: Use the training data set to train the neural network model to obtain the target recognition model.

[0103] The training data set is manually annotated and then input into the neural network model for training. The above-mentioned target recognition model is used to identify whether there is a disease in the tunnel image and / or identify the diseased area in the tunnel image.

[0104] Step 3: Obtain an image of the tunnel to be inspected, preprocess the image of the tunnel to be inspected, and then input the preprocessed image of the tunnel to be inspected into the target recognition model to obtain a target image with defects.

[0105] Preprocessing the tunnel images to be inspected can identify defects in images under different tunnel environments and conditions, thereby improving the accuracy and efficiency of the identification results.

[0106] Step 4: Extract the diseased area from the target image and mark it.

[0107] There may be a variety of different image features in the target image. Identifying and labeling the target image and accurately extracting the diseased area further verifies the existence and scope of the disease, reduces the possibility of false positives and missed negatives, and ensures the accuracy of the recognition results.

[0108] Specifically, extracting the diseased area from the target image and labeling it include:

[0109] Step 41: Divide the target image into a plurality of first partial images according to a second preset size.

[0110] Step 42: Extract any first partial image, perform a weighted sum of the intensity values ​​of each color channel at any pixel point in any first partial image, and define the sum as the second point eigenvalue at any pixel point. After traversing all pixel points in any first partial image, calculate the first statistical value of all second point eigenvalues, and determine whether the first statistical value is within a first preset range. If so, mark any first partial image as a first category area.

[0111] Step 43: After traversing all first partial images, define the unlabeled first partial image as the second partial image, extract any second partial image, determine the maximum intensity value and the minimum intensity value of all color channels at any pixel point of any second partial image, calculate the ratio of the difference between the maximum intensity value and the minimum intensity value to the maximum intensity value, and define it as the third point eigenvalue at any pixel point, and after traversing all pixels in any second partial image, calculate the second statistical value of all third point eigenvalues, determine whether the second statistical value is within the second preset range, and if so, label any second partial image as the second category area.

[0112] Step 44: After traversing all the second partial images, define the unlabeled second partial images as third partial images, extract any third partial image, and determine whether the second statistical value corresponding to any third partial image is within a third preset range. If so, label any third partial image as a third category area.

[0113] Step 45: After traversing all third partial images, define the unlabeled third partial images as fourth partial images, extract any fourth partial image, determine the texture statistical feature value of any fourth partial image, and judge whether the texture statistical feature value is within a fourth preset range. If so, label any fourth partial image as a fourth category area; if not, label any fourth partial image as a fifth category area.

[0114] The second preset size, the first preset range, the second preset range, the third preset range, and the fourth preset range are set based on the experience of those skilled in the art or based on actual application scenarios, and are not limited in the embodiments of the present application. For example, the second preset size is 25×25 pixels, the first preset range is [160, ∞), the second preset range is [15, ∞), the third preset range is [6, 11], and the fourth preset range is [0.35, ∞).

[0115] When weighted summing the intensity values ​​of each color channel at any pixel, the weight values ​​of each color channel can be the same or different. The first statistical value is the average of all second-point eigenvalues. The first category area is the illumination area. The second statistical value is the dispersion value of all third-point eigenvalues. The second category area is the cable area. The third category area is free calcium oxide. The texture statistical eigenvalue is a statistical value reflecting the uniformity of the local image grayscale distribution and the coarseness of the texture. The fourth category area is the crack area. The fifth category area is the normal area or the area that does not need to be identified.

[0116] According to the technical solution of the present invention, by dividing the target image into multiple local images and identifying, excluding and labeling each local image in turn, the refinement of disease detection can be improved, and the diseased area can be extracted and labeled more accurately.

[0117] The image feature statistics corresponding to the various categories of regions and the preset ranges of the image feature statistics are obtained by statistically analyzing the color features and texture features of the various categories of regions in a preset number of tunnel images.

[0118] Specifically, the contours of the sequentially connected fourth category areas in the target image are extracted and marked as diseased areas.

[0119] Tunnel diseases include cracks, water seepage, and spalling. The present invention extracts and labels crack diseases. Based on this technical solution, the image feature statistical value category and the preset range corresponding to the image feature statistical value can be adjusted to adapt to different types of disease detection needs.

[0120] Figure 3 The figure shows a schematic diagram of the structure of an embodiment of a tunnel disease identification and detection system based on deep learning provided by the present invention. Figure 3 As shown, the system includes: a first acquisition module 10, a preprocessing module 20, a model training module 30, a second acquisition module 40, an image detection module 50 and a disease extraction module 60.

[0121] The first acquisition module 10 is used to acquire historical tunnel disease images and input the historical disease images into the pre-processing module 20 .

[0122] The preprocessing module 20 is used to preprocess the historical disease images to obtain a training data set.

[0123] The model training module 30 is used to train the neural network model using the training data set to obtain a target recognition model.

[0124] The second acquisition module 40 is used to acquire the image of the tunnel to be detected and input the image of the tunnel to be detected into the pre-processing module.

[0125] The preprocessing module 20 is further configured to preprocess the tunnel image to be detected and obtain the preprocessed tunnel image to be detected.

[0126] The image detection module 50 is used to input the pre-processed tunnel image to be detected into the target recognition model to obtain the target image with defects.

[0127] The disease extraction module 60 is used to extract the diseased area from the target image and mark it.

[0128] It should be understood that, although the various steps in the flow chart of each embodiment of the present invention are shown in sequence according to the indication of the arrows, these steps are not necessarily performed in sequence according to the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0129] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. The above-mentioned program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0130] The above embodiments merely represent preferred implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention.

Claims

1. A tunnel disease identification and detection method based on deep learning, characterized in that: The steps include: Acquire historical tunnel disease images, preprocess the historical disease images, and obtain a training data set; Using the training data set to train a neural network model to obtain a target recognition model; Acquiring an image of the tunnel to be inspected, performing the preprocessing on the image of the tunnel to be inspected, and then inputting the preprocessed image of the tunnel to be inspected into the target recognition model to obtain a target image with a disease; Extracting the diseased area from the target image and marking it; The preprocessing includes a first preprocessing, and the first preprocessing includes: Step 111: extract any first image to be processed, and divide any first image to be processed into a plurality of sub-images according to a first preset size, wherein the first image to be processed is the historical disease image or the tunnel image to be inspected; Step 112: extract any sub-image, determine the image feature value of any sub-image, and determine whether the image feature value is greater than or equal to a first preset value. If so, proceed to step 113; if not, proceed to step 114. Step 113: input any of the sub-images into a first network model to obtain an interference information image, and generate a first sub-image based on any of the sub-images and the interference information image; Step 114: Input any of the sub-images into the second network model to obtain the first sub-image; Step 115: After traversing all sub-images, reassemble all first sub-images to generate a first image; Step 116: After traversing all first images to be processed, obtain the first preprocessing result; In step 112, the image feature value includes a first image feature value and / or a second image feature value, and the first image feature value and the second image feature value are determined by: Convert any of the sub-images into a frequency domain analysis image, extract any frequency domain point in the frequency domain analysis image, and define any of the frequency domain points as a first-category point when the amplitude of any of the frequency domain points is greater than a second preset value; otherwise, define any of the frequency domain points as a second-category point. After traversing all points in the frequency domain analysis image, calculate the average of the amplitudes of all first-category points and define the average as the first image feature value. Extract any pixel point in any of the sub-images, perform a weighted sum of the intensity values ​​of each color channel at any of the pixel points, and define the sum as the first point eigenvalue at any of the pixel points; after traversing all of the pixels in any of the sub-images, calculate the average value of all of the first point eigenvalues, and define the average value as the second image eigenvalue; Extracting the diseased area from the target image and marking it include: Step 41: Divide the target image into a plurality of first partial images according to a second preset size; Step 42: Extract any first partial image, perform a weighted sum of the intensity values ​​of each color channel at any pixel point in any first partial image, and define the sum as a second point eigenvalue at any pixel point. After traversing all pixels in any first partial image, calculate a first statistic of all second point eigenvalues, determine whether the first statistic is within a first preset range, and if so, mark any first partial image as a first category area. Step 43: After traversing all first partial images, define the unlabeled first partial images as second partial images, extract any second partial image, determine the maximum intensity value and the minimum intensity value of all color channels at any pixel point in any second partial image, calculate the ratio of the difference between the maximum intensity value and the minimum intensity value to the maximum intensity value, and define this as the third point eigenvalue at any pixel point, and after traversing all pixels in any second partial image, calculate a second statistical value of all the third point eigenvalues, determine whether the second statistical value is within a second preset range, and if so, label any second partial image as a second category area; Step 44: After traversing all second partial images, define the unlabeled second partial images as third partial images, extract any third partial image, and determine whether the second statistical value corresponding to any third partial image is within a third preset range. If so, label any third partial image as a third category area. Step 45: After traversing all third partial images, define the unlabeled third partial images as fourth partial images, extract any fourth partial image, determine the texture statistical feature value of any fourth partial image, and judge whether the texture statistical feature value is within a fourth preset range. If so, label any fourth partial image as a fourth category area; if not, label any fourth partial image as a fifth category area.

2. The method according to claim 1, characterized in that In step 115, when any two adjacent first sub-images are reassembled, the following steps are performed: determining whether there is an intersection region between the two first sub-images; if not, directly stitching the two first sub-images together; and if so, dividing the two first sub-images into a first stitched image, a second stitched image, a third stitched image, and a fourth stitched image based on the intersection region, wherein the second stitched image and the third stitched image are stitched images corresponding to the intersection region; extracting a first pixel value at any pixel point on the second stitched image and a second pixel value at any pixel point on the third stitched image, taking an average of the first pixel value and the second pixel value as the pixel value at any pixel point on the fifth stitched image, and generating the fifth stitched image after traversing all pixel points on the second stitched image; The first stitched image, the fifth stitched image, and the fourth stitched image are stitched together.

3. The method according to claim 1, characterized in that The preprocessing further includes a second preprocessing, wherein the second preprocessing includes: Extracting any second image to be processed, scaling the physical size of any second image to be processed according to a first preset ratio to generate a second image, wherein the second image to be processed is the historical disease image, the image of the tunnel to be inspected, or the first image; Inputting the second image into a third network model to obtain a third image, wherein the third image has the same physical size as the second image and a clarity index value of the third image is greater than a clarity index value of the second image; After traversing all the second images to be processed, the result of the second preprocessing is obtained.

4. The method according to claim 3, characterized in that The method for generating the third network model is: Step 121: Divide the shooting area according to a second preset ratio to obtain multiple sub-areas; Step 122: using an image acquisition device to acquire a first overall image of the shooting area, and also using the image acquisition device to acquire a regional image of each of the sub-areas, and integrating all the regional images to acquire a second overall image; Step 123: Scale the physical size of the first overall image according to the first preset ratio to generate a third overall image, use the third overall image and the second overall image to train a specific generative adversarial network, and define the generator that changes the third overall image in the direction of the second overall image as the third network model.

5. The method according to claim 1, wherein The contours of the sequentially connected fourth category areas in the target image are extracted and marked as the diseased areas.

6. A tunnel disease identification and detection system based on deep learning, used to implement the method according to any one of claims 1 to 5, characterized in that: include: a first acquisition module, a preprocessing module, a model training module, a second acquisition module, an image detection module, and a disease extraction module; The first acquisition module is used to acquire historical tunnel disease images and input the historical disease images into the pre-processing module; The preprocessing module is used to preprocess the historical disease images to obtain a training data set; The model training module is used to train a neural network model using the training data set to obtain a target recognition model; The second acquisition module is used to acquire the image of the tunnel to be detected and input the image of the tunnel to be detected into the preprocessing module; The preprocessing module is further configured to preprocess the tunnel image to be detected to obtain the preprocessed tunnel image to be detected; The image detection module is used to input the pre-processed tunnel image to be detected into the target recognition model to obtain the target image with the disease; The disease extraction module is used to extract the diseased area from the target image and mark it.

Images