Method and system for realizing tunnel crack detection based on Deeplab

Abstract

The invention provides a method and a system for realizing tunnel crack detection based on Deeplab, belongs to the field of tunnel crack detection, and is used for solving the problem of poor effect and accuracy of tunnel crack detection in related technologies. And the prediction effect of the improved Deeplab network is evaluated by using the segmentation performance parameters, so that a segmentation model with a relatively good prediction effect can be output, and better detection of tunnel cracks by four lines is facilitated.

Description

Technical Field

[0001] The present application relates to the field of tunnel crack detection, and in particular to a method and system for implementing tunnel crack detection based on DeepLab. Background Art

[0002] In the field of tunnel crack detection, semantic segmentation models based on deep learning face challenges such as high data acquisition and annotation costs and strong diversity of crack morphology. The traditional method of dividing the training set, validation set, and test set in a fixed ratio (such as 7:2:1) is difficult to adapt to the special needs of tunnel scenarios. Especially when crack samples are scarce, the static division strategy can easily lead to insufficient learning of small target features by the model, resulting in missed detection or misjudgment. In addition, complex background interference such as uneven lighting and structural occlusion in the tunnel environment makes it impossible for fixed-ratio data division to effectively cover multimodal features, reducing the generalization ability of the model.

[0003] Existing tunnel crack detection models (such as the improved DeepLab series) rely on a fixed ratio of training data for feature learning, and it is difficult to dynamically adapt to the differences in crack distribution in different tunnel structures. When there is a deviation between the crack type in the training data set and the actual detection scenario, the model cannot optimize the feature representation by adjusting the data partitioning strategy in real time, resulting in problems such as blurred segmentation boundaries and missed detection of small cracks under complex geological conditions. In addition, traditional methods lack quantitative analysis of multi-scale characteristics of tunnel structures (such as crack width and direction), and it is difficult to enhance the model's perception of cracks of different scales through data partitioning strategies.

[0004] In terms of model training termination conditions, existing technologies are mainly based on the convergence judgment of the validation set loss value or a single indicator (such as IoU), which cannot fully reflect the multi-dimensional requirements of tunnel crack detection tasks. For example, in engineering applications, it is necessary to take into account the accuracy of crack identification, boundary positioning accuracy and real-time requirements at the same time, and the traditional termination strategy lacks a comprehensive evaluation of these indicators, resulting in the model stopping training prematurely or falling into invalid iterations before meeting engineering requirements, affecting detection efficiency and reliability.

[0005] The current tunnel crack detection technology has not yet established a mathematical correlation model between data partition ratio, training parameters and detection performance. When the evaluation results of the validation set do not meet expectations, the manual adjustment of parameters such as learning rate and batch size lacks systematic guidance, and is not combined with the data enhancement strategy of the tunnel scene (such as simulating lighting changes and adding noise interference) for coordinated optimization. This isolated parameter adjustment mode is difficult to break through the constraints of data limitation on model performance, resulting in a long-term contradiction between detection accuracy and engineering practicality. Summary of the invention

[0006] The present application provides a method and system for realizing tunnel crack detection based on improved DeepLab segmentation, which can improve the accuracy and efficiency of tunnel crack detection by adjusting the segmentation ratio of labeled data sets, and provide strong support for tunnel structure health monitoring.

[0007] In a first aspect, the present application provides a method for tunnel crack detection based on DeepLab. The method comprises: Obtain annotated dataset of tunnel cracks; The labeled data set is randomly divided into training set, test set and validation set according to a:b:c; Use the training set and test set to train and improve the Deeplab network to obtain the segmentation model; Use the validation set to verify whether the segmentation performance parameter of the segmentation model is greater than the performance parameter threshold; If so, output the segmentation model; Otherwise, adjust the training strategy; The method for obtaining the segmentation performance parameter includes: Suppose there are n types of recognition results corresponding to image features, and the i-th type recognition result is predicted as the i-th type result , predicting the i-th category recognition result as the j-th category result is , predicting the j-th category recognition result as the i-th category result is , , the segmentation performance parameter is z, then .

[0008] By adopting the above technical solution, the segmentation model is obtained by training the improved Deeplab network, and the segmentation performance parameters are used to evaluate the prediction effect of the improved Deeplab network, so as to output a segmentation model with better prediction effect, which is conducive to better detection of cracks in four-line pairs of tunnels.

[0009] Further, .

[0010] Further, ,The two types of recognition results are tunnel cracks and background.

[0011] Furthermore, the adjusting the training strategy includes changing training parameters, data sets, and learning rates.

[0012] Furthermore, the improved Deeplab network includes an encoding model and a decoding model; The coding model includes: Use the MobileNetV2 backbone network to extract the initial features of the input image; The initial features are processed through three branches to obtain branch features. The three branches are 1*1 convolution branch, dilation rate rata 2, 6 hollow convolution branch, and pooling branch. The branch features are processed by 1*1 convolution to obtain the encoded output features; The decoding model includes: Collect the intermediate down-sampling features of the MobileNetV2 backbone network; Perform 1*1 convolution on the intermediate downsampled features to obtain downsampled fusion features; The encoding output features are concatenated and fused with the down-sampling fusion features to obtain the up- and down-sampling fusion features; The up- and down-sampled fusion features are sequentially processed by 3*3 convolution and up-sampled to obtain the output segmented image.

[0013] In a second aspect, the present application provides a system for tunnel crack detection based on DeepLab. The system comprises: A data acquisition module, used to obtain a labeled data set of tunnel cracks; The data segmentation module is used to randomly split the labeled data set into training set, test set and validation set according to a:b:c; A model training module is used to train and improve the Deeplab network using the training set and the test set to obtain a segmentation model; and A training management module is used to use the validation set to verify whether the segmentation performance parameter of the segmentation model is greater than the performance parameter threshold; if so, the segmentation model is output; otherwise, the training strategy is adjusted; The training management module is further configured such that the method for obtaining the segmentation performance parameter includes: Suppose there are n types of recognition results corresponding to image features, and the i-th type recognition result is predicted as the i-th type result , predicting the i-th category recognition result as the j-th category result is , predicting the j-th category recognition result as the i-th category result is , , the segmentation performance parameter is z, then .

[0014] Furthermore, the data segmentation module is further configured to: .

[0015] Furthermore, the training management module is further configured to: ,The two types of recognition results are tunnel cracks and background.

[0016] Furthermore, the training management module is further configured such that the adjustment of the training strategy includes changing training parameters, data sets, and learning rates.

[0017] Furthermore, the model training module is further configured such that the improved Deeplab network includes an encoding model and a decoding model; The coding model includes: Use the MobileNetV2 backbone network to extract the initial features of the input image; The initial features are processed through three branches to obtain branch features. The three branches are 1*1 convolution branch, dilation rate rata 2, 6 hollow convolution branch, and pooling branch. The branch features are processed by 1*1 convolution to obtain the encoded output features; The decoding model includes: Collect the intermediate down-sampling features of the MobileNetV2 backbone network; Perform 1*1 convolution on the intermediate downsampled features to obtain downsampled fusion features; The encoding output features are concatenated and fused with the down-sampling fusion features to obtain the up- and down-sampling fusion features; The up- and down-sampled fusion features are sequentially processed by 3*3 convolution and up-sampled to obtain the output segmented image.

[0018] In summary, this application at least has the following beneficial effects: 1. A method and system for tunnel crack detection based on Deeplab is provided, which obtains a better segmentation model by improving the Deeplab network and the designed segmentation performance parameter calculation model, which is conducive to better detection of tunnel cracks; 2. The Deeplab network is independently designed and improved. The model is lightweight and fully considers the influence of computational complexity and detail information. The model can expand the receptive field, retain the internal data structure and avoid the loss of feature information caused by downsampling, so as to realize tunnel crack detection more efficiently and accurately.

[0019] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present application, nor are they intended to limit the scope of the present application. Other features of the present application will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other features, advantages and aspects of the embodiments of the present application will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein: Figure 1 A schematic diagram of a method for realizing tunnel crack detection based on DeepLab in an embodiment of the present application is shown; Figure 2The schematic diagram of the improved Deeplab network in the embodiment of the present application is shown; Figure 3 A schematic diagram of a system for tunnel crack detection based on DeepLab in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0022] In addition, the term "and / or" in this article is only a description of the association relationship between the associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0023] The present application provides a method and system for realizing tunnel crack detection based on Deeplab, which obtains a better segmentation model through an independently designed improved Deeplab network and a segmentation performance parameter calculation model, which is conducive to better realizing tunnel crack detection.

[0024] In a first aspect, an embodiment of the present application discloses a method for tunnel crack detection based on DeepLab. The method can be executed by a server.

[0025] refer to Figure 1 The method specifically comprises the following steps: S110: Obtain a labeled dataset of tunnel cracks.

[0026] The annotated data set contains annotated image data, which is an image containing various tunnel cracks or no tunnel cracks in the background of tunnel crack detection, and the tunnel cracks in the image have been annotated. The annotated image data can be obtained by processing the original tunnel detection video data, wherein the tunnel detection video data is collected under different conditions of different dimensions such as weather, lighting, visibility, and environment to determine the diversity of tunnel detection video content, and the tunnel detection video is processed into an image sequence to obtain a frame image to be processed, and the frame image to be processed is pre-processed by image enhancement, denoising, geometric distortion correction, motion blur removal, etc., and then annotated to obtain the annotated image data, and all the annotated image data form an annotated data set.

[0027] S120: Randomly split the labeled data set into a training set, a test set, and a validation set according to a:b:c.

[0028] The segmentation ratio here refers to the quantity ratio, specifically the quantity ratio of the labeled image data contained in the training set, test set, and validation set. In a specific example, Of course, other split ratios, such as 7:3:1, can also be considered.

[0029] S130: Use the training set and test set to train and improve the Deeplab network to obtain a segmentation model.

[0030] In this step, the DeepLab network is improved to be a self-designed model. Its structure is adjusted based on DeepLab v3+ and has good effects in the field of tunnel detection. Figure 2 A detailed introduction to the improved Deeplab network is given.

[0031] Reference Figure 2 , the improved Deeplab network includes an encoding model and a decoding model; The coding model includes: Use the MobileNetV2 backbone network to extract the initial features of the input image; The initial features are processed through three branches to obtain branch features. The three branches are 1*1 convolution branch, dilation rate rata 2, 6 hollow convolution branch, and pooling branch. The branch features are processed by 1*1 convolution to obtain the encoded output features; The decoding model includes: Collect the intermediate down-sampling features of the MobileNetV2 backbone network; Perform 1*1 convolution on the intermediate downsampled features to obtain downsampled fusion features; The encoding output features are concatenated and fused with the down-sampling fusion features to obtain the up- and down-sampling fusion features; The up- and down-sampled fusion features are sequentially processed by 3*3 convolution and up-sampled to obtain the output segmented image.

[0032] In the case where the Deeplab network has been improved, the specific means of using the training set and the test set to train and improve the Deeplab network are common knowledge to those skilled in the art, and the specific training means do not belong to the improvement of this application and will not be described here.

[0033] S140: Using the validation set to verify whether the segmentation performance parameter of the segmentation model is greater than a performance parameter threshold.

[0034] In the embodiment of the present application, the segmentation performance parameters designed independently are used to evaluate the effect of the segmentation model on tunnel detection. The specific calculation method of the segmentation performance parameters is introduced below.

[0035] In one embodiment, the method for obtaining the segmentation performance parameter includes: assuming that there are n types of recognition results corresponding to image features, predicting the i-th type recognition result as the i-th type result , predicting the i-th category recognition result as the j-th category result is , predicting the j-th category recognition result as the i-th category result is , , the segmentation performance parameter is z, then .

[0036] In a specific example, , the two recognition results are tunnel cracks and background, and the two recognition results are 0 and 1. Specifically, the segmentation performance parameters ,in, To correctly predict the number of classes, , are the number of false positives and false negatives, respectively.

[0037] The performance parameter thresholds are preset in the server.

[0038] S150: If the judgment in step S140 is yes, the segmentation model is output.

[0039] When the segmentation performance parameter satisfies and is greater than the performance parameter threshold, it indicates that the segmentation model has met the expected effect. At this time, the optimal training result is output as the segmentation model, and the segmentation model is used to detect tunnel cracks with good effect.

[0040] S160: If the determination in step S140 is no, adjust the training strategy.

[0041] When the segmentation performance parameter does not meet the performance parameter threshold, it means that the segmentation model does not meet the expected effect. At this time, it is necessary to adjust the training strategy to continue training the segmentation model. In the method of this step, the adjustment of the training strategy includes changing the training parameters, data sets, learning rates, etc. Adjusting the training strategy is a conventional technical means for those skilled in the art and does not belong to the improvement of this application, so it is not described here.

[0042] In summary, this method can realize the detection of tunnel cracks, the detection process is more efficient, and the detection results are more accurate.

[0043] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the embodiments of the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0044] In a second aspect, the present application discloses a system for tunnel crack detection based on DeepLab. The system may be included in a server or implemented as a server.

[0045] Figure 3 A schematic diagram of a system for tunnel crack detection based on DeepLab in an embodiment of the present application is shown.

[0046] Reference Figure 3 The system specifically includes: A data acquisition module, used to obtain a labeled data set of tunnel cracks; The data segmentation module is used to randomly split the labeled data set into training set, test set and validation set according to a:b:c; A model training module is used to train and improve the Deeplab network using the training set and the test set to obtain a segmentation model; and A training management module is used to use the validation set to verify whether the segmentation performance parameter of the segmentation model is greater than the performance parameter threshold; if so, the segmentation model is output; otherwise, the training strategy is adjusted; The training management module is further configured such that the method for obtaining the segmentation performance parameter includes: Suppose there are n types of recognition results corresponding to image features, and the i-th type recognition result is predicted as the i-th type result , predicting the i-th category recognition result as the j-th category result is , predicting the j-th category recognition result as the i-th category result is , , the segmentation performance parameter is z, then .

[0047] Furthermore, the data segmentation module is further configured to: .

[0048] Furthermore, the training management module is further configured to: ,The two types of recognition results are tunnel cracks and background.

[0049] Furthermore, the training management module is further configured such that the adjustment of the training strategy includes changing training parameters, data sets, and learning rates.

[0050] Furthermore, the model training module is further configured such that the improved Deeplab network includes an encoding model and a decoding model; The coding model includes: Use the MobileNetV2 backbone network to extract the initial features of the input image; The initial features are processed through three branches to obtain branch features. The three branches are 1*1 convolution branch, dilation rate rata 2, 6 hollow convolution branch, and pooling branch. The branch features are processed by 1*1 convolution to obtain the encoded output features; The decoding model includes: Collect the intermediate down-sampling features of the MobileNetV2 backbone network; Perform 1*1 convolution on the intermediate downsampled features to obtain downsampled fusion features; The encoding output features are concatenated and fused with the down-sampling fusion features to obtain the up- and down-sampling fusion features; The up- and down-sampled fusion features are sequentially processed by 3*3 convolution and up-sampled to obtain the output segmented image.

[0051] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0052] In summary, this application at least has the following beneficial effects: 1. Based on the use of backbone network: By analyzing the data set samples and considering the impact of computational complexity on real-time performance, the lightweight network mobilenetv2 is used as the backbone network for training the segmentation model; DeepLab v3+ is a semantic segmentation algorithm that uses a spatial pyramid module and an encoder-decoder structure, and obtains richer contextual information through pooling operations at different resolutions. The encoding module uses mobilenetv2 as the backbone network for feature extraction, thereby linking the ASSP module to extract multi-scale information of the fused image. The feature map of the decoding module is fused with the intermediate information of mobilenetv2. The feature map of the encoding module provides semantic information, and the intermediate downsampling of mobilenetv2 provides detail information, and then upsampling is used to obtain the segmentation result.

[0053] 2. Using dilated convolution with secondary rates equal to 2 and 6 respectively, one layer of dilated convolution calculation is reduced compared with the DeepLab v3+ network, which can expand the receptive field to a certain extent, preserve the internal data structure and avoid the loss of feature information caused by downsampling, thereby obtaining a feature map with more information. The number of parameters and calculations can be reduced through deep separable convolution.

[0054] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the aforementioned disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other to form a technical solution.

Claims

1. A method for tunnel crack detection based on DeepLab, characterized in that: include: Obtain annotated dataset of tunnel cracks; The labeled data set is randomly divided into training set, test set and validation set according to a:b:c; Use the training set and test set to train and improve the Deeplab network to obtain the segmentation model; Use the validation set to verify whether the segmentation performance parameter of the segmentation model is greater than the performance parameter threshold; If so, output the segmentation model; Otherwise, adjust the training strategy; The method for obtaining the segmentation performance parameter includes: Suppose there are n types of recognition results corresponding to image features, and the i-th type recognition result is predicted as the i-th type result , predicting the i-th category recognition result as the j-th category result is , predicting the j-th category recognition result as the i-th category result is , , the segmentation performance parameter is z, then .

2. The method according to claim 1, characterized in that 。 3. The method according to claim 1, characterized in that ,The two types of recognition results are tunnel cracks and background.

4. The method according to claim 1, characterized in that: The adjustment of the training strategy includes changing the training parameters, data set, and learning rate.

5. The method according to any one of claims 1 to 4, characterized in that: The improved Deeplab network includes an encoding model and a decoding model; The coding model includes: Use the MobileNetV2 backbone network to extract the initial features of the input image; The initial features are processed through three branches to obtain branch features. The three branches are 1*1 convolution branch, dilation rate rata 2, 6 hollow convolution branch, and pooling branch. The branch features are processed by 1*1 convolution to obtain the encoded output features; The decoding model includes: Collect the intermediate down-sampling features of the MobileNetV2 backbone network; Perform 1*1 convolution on the intermediate downsampled features to obtain downsampled fusion features; The encoding output features are concatenated and fused with the down-sampling fusion features to obtain the up- and down-sampling fusion features; The up- and down-sampled fusion features are sequentially processed by 3*3 convolution and up-sampled to obtain the output segmented image.

6. A system for tunnel crack detection based on DeepLab, characterized in that: include: A data acquisition module, used to obtain a labeled data set of tunnel cracks; The data segmentation module is used to randomly split the labeled data set into training set, test set and validation set according to a:b:c; Model training module, used to train and improve the Deeplab network using training sets and test sets to obtain a segmentation model; as well as A training management module, used to verify whether the segmentation performance parameter of the segmentation model is greater than the performance parameter threshold using the verification set; If so, output the segmentation model; Otherwise, adjust the training strategy; The training management module is configured such that the method for obtaining the segmentation performance parameter includes: Suppose there are n types of recognition results corresponding to image features, and the i-th type recognition result is predicted as the i-th type result , predicting the i-th category recognition result as the j-th category result is , predicting the j-th category recognition result as the i-th category result is , , the segmentation performance parameter is z, then .

7. The system according to claim 6, characterized in that The data segmentation module is configured to: .

8. The system according to claim 6, characterized in that The training management module is configured to: ,The two types of recognition results are tunnel cracks and background.

9. The system according to claim 6, characterized in that The training management module is configured such that the adjustment of the training strategy includes changing training parameters, data sets, and learning rates.

10. The system according to any one of claims 6 to 9, characterized in that: The model training module is configured such that the improved Deeplab network includes an encoding model and a decoding model; The coding model includes: Use the MobileNetV2 backbone network to extract the initial features of the input image; The initial features are processed through three branches to obtain branch features. The three branches are 1*1 convolution branch, dilation rate rata 2, 6 hollow convolution branch, and pooling branch. The branch features are processed by 1*1 convolution to obtain the encoded output features; The decoding model includes: Collect the intermediate down-sampling features of the MobileNetV2 backbone network; Perform 1*1 convolution on the intermediate downsampled features to obtain downsampled fusion features; The encoding output features are concatenated and fused with the down-sampling fusion features to obtain the up- and down-sampling fusion features; The up- and down-sampled fusion features are sequentially processed by 3*3 convolution and up-sampled to obtain the output segmented image.

Images