Bridge apparent crack identification method based on adaptive subset search and deep learning

Abstract

The application provides a bridge apparent crack identification method based on adaptive subset search and deep learning, comprising the following steps: obtaining a high-speed railway bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; using a random sampling method to select multiple samples from the candidate dataset as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the prediction results of the picture dataset by the deep neural network model in the adjacent two iteration training processes; if the relative loss of the prediction results of the picture dataset in the continuous multiple iterations is less than the allowable value, it is determined that the deep neural network model converges; saving the deep neural network model, and using the deep neural network model to identify the high-speed railway bridge apparent crack. The application can solve the technical problem of sampling local optimization in the active learning of the high-speed railway bridge apparent crack identification in the prior art, and improve the efficiency and accuracy of crack identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a bridge apparent crack identification method based on adaptive subset search and deep learning. BACKGROUND

[0002] Large-span high-speed railway (high-speed rail) bridges will undergo performance degradation during long-term service, resulting in a decrease in the safety performance of the bridge structure, which seriously threatens people's lives and property safety. Performing state evaluation of high-speed rail bridges can effectively identify the degree of performance degradation of the bridge and prevent disasters from occurring in advance, which is an important means to ensure the safe operation of high-speed rail bridges. Surface cracking is an important early indicator of bridge structure damage, and identifying apparent cracks is an important task of state evaluation of high-speed rail bridges.

[0003] Traditional manual detection methods are costly, and manual high-altitude work has safety hazards, making it difficult to efficiently detect a large number of high-speed rail bridges. In recent years, computer vision technology based on deep learning has developed rapidly and has been successfully applied to the field of structural apparent crack identification, which can greatly promote the automation of high-speed rail bridge apparent crack identification. However, deep learning models used for apparent crack computer vision identification require a large number of labeled crack datasets for training to achieve sufficient recognition accuracy. In actual engineering, it is usually difficult to establish a large training dataset, and the cost of labeling the dataset is relatively high. Compared with traditional training methods, active learning methods can train deep learning models with relatively few training data, but can achieve high recognition accuracy. However, existing expression-based and uncertainty-based active learning methods fail to balance exploration and utilization during the active learning process, which can easily lead to local optimal sampling problems, resulting in low efficiency of active learning and waste of training datasets. SUMMARY

[0004] The present application aims to overcome the above technical deficiencies and provide a bridge apparent crack identification method based on adaptive subset search and deep learning, which can solve the technical problem of local optimal sampling in active learning of high-speed railway bridge apparent crack identification, and improve the efficiency and accuracy of high-speed railway bridge apparent crack identification.

[0005] To achieve the above technical purpose, in a first aspect, the technical solution of the present application provides a bridge apparent crack identification method based on adaptive subset search and deep learning, comprising the following steps:

[0006] Obtaining a high-speed rail bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset;

[0007] a plurality of samples are selected from the candidate dataset as new training samples by random sampling, the training samples are labeled, and are added to the training set;

[0008] The deep neural network model for crack identification is updated using the training set, and the relative loss of the prediction results of the picture dataset in the adjacent two iteration training processes is calculated;

[0009] If the relative loss of the prediction results of the picture dataset in the continuous multiple iterations is less than a first allowable value, the deep neural network model converges;

[0010] The deep neural network model is saved, and is used for high-speed railway bridge apparent crack identification.

[0011] Compared with the prior art, the beneficial effects of the present application include:

[0012] The bridge apparent crack identification method based on adaptive subset search and deep learning provided by the present application adopts a global-to-local strategy to select a subset and a training sample from a high-speed railway bridge apparent crack dataset according to the uncertainty of the sample, and trains a deep neural network model for crack identification; the relative loss of the prediction results of the picture dataset in the adjacent two iterations is calculated to determine the model convergence and subset convergence. The present application measures the global prediction accuracy of the deep neural network model according to the subset convergence, adaptively reduces the sampling range, and balances the exploration and utilization in the active learning process; the model convergence is used as the termination condition of the active learning, which can effectively improve the efficiency and accuracy of the high-speed railway bridge apparent crack identification.

[0013] According to some embodiments of the present application, if the deep neural network model does not meet the convergence condition, the prediction label of the current subset in the last iteration is used as the true label, and the prediction label of the current subset in the current iteration is predicted by the deep neural network model;

[0014] The relative loss of the prediction results of the current subset in the two iterations is calculated according to the true label, the prediction label and the loss function of the network, and whether the current subset meets the convergence condition is determined;

[0015] When the relative loss of the prediction results of the current subset in the continuous multiple iterations is less than a second allowable value, the current subset meets the convergence condition, the current subset is deleted from the candidate dataset to reduce the search range of the active learning;

[0016] The samples of the candidate dataset are predicted by the deep neural network model, the uncertainty of the samples of the candidate dataset is calculated based on the information entropy of the prediction results, and the uncertainty of the samples of the candidate dataset is updated.

[0017] According to some embodiments of the present application, after judging whether the current subset meets the convergence condition, the method comprises the steps of:

[0018] If the current subset does not meet the convergence condition, the uncertainty of the candidate dataset sample is calculated based on the information entropy of the prediction result to update the uncertainty of the sample of the candidate dataset.

[0019] According to some embodiments of the present application, the relative loss of the prediction result of the picture dataset of the deep neural network model in the adjacent two iteration training processes is calculated, comprising the steps of:

[0020] The prediction label of the picture dataset in the last iteration by the deep neural network model is used as the real label, the prediction label of the picture dataset is predicted by the deep neural network model in the current iteration, and then the relative loss of the dataset in the two iterations is calculated according to the real label, the prediction label and the loss function of the deep neural network model.

[0021] According to some embodiments of the present application, the loss function of the deep neural network model comprises two parts of Dice Loss loss function and cross-entropy loss function.

[0022] Wherein the Dice Loss loss function is defined as:

[0023]

[0024] In the formula, S i p and S i g are the predicted and real pixel classifications respectively, and N is the total number of pixels.

[0025] The definition of the cross-entropy loss function is:

[0026]

[0027] x is the input and also the output of the last layer of the deep neural network model, and class is the real label index value of the pixel.

[0028] The loss function of the deep neural network model is:

[0029] loss = loss CE + loss Dice .

[0030] According to some embodiments of the present application,

[0031] According to some embodiments of the present application, the uncertainty of the candidate dataset samples is calculated based on the information entropy of the prediction results, including:

[0032] The uncertainty calculation formula based on information entropy is:

[0033]

[0034] p(S i ) is the probability of the pixel being predicted as class i, n is the total number of classes, and N is the total number of pixels in the sample picture.

[0035] According to some embodiments of the present application, after updating the uncertainty of the samples of the candidate dataset, including the step of:

[0036] According to the uncertainty of the samples, the first N s samples with the minimum uncertainty are selected from the candidate dataset as the new current subset, where N s is defined as:

[0037] N s = p s N c

[0038] p s is the proportionality coefficient, and N c is the number of remaining candidate dataset samples.

[0039] According to some embodiments of the present application, if the relative loss of the prediction results of the picture dataset in consecutive iterations is less than a first allowable value, the deep neural network model converges, including the step of:

[0040] If the relative loss of the prediction results of the picture dataset in consecutive iterations is less than the first allowable value, the first allowable value is 0.15, and the deep neural network model converges.

[0041] According to some embodiments of the present application, when the relative loss of the prediction results of the current subset in consecutive iterations is less than a second allowable value, the current subset satisfies the convergence condition, including the step of:

[0042] When the relative loss of the prediction results of the current subset in consecutive iterations is less than the second allowable value, the second allowable value is 0.3, and the current subset satisfies the convergence condition.

[0043] In a second aspect, the present application provides a bridge apparent crack identification system based on adaptive subset search and deep learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the bridge apparent crack identification method based on adaptive subset search and deep learning as described in any one of the first aspect.

[0044] Additional aspects and advantages of the present application will be in part apparent and in part pointed out below. BRIEF DESCRIPTION OF DRAWINGS

[0045] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the Abstract, of the embodiments, with reference to the following drawings:

[0046] Figure 1 A flow chart of the bridge apparent crack identification method based on adaptive subset search and deep learning provided for an embodiment of the present application;

[0047] Figure 2 A flow chart of the bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application;

[0048] Figure 3 A flow chart of the bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application;

[0049] Figure 4 A computer vision model schematic diagram of the bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application;

[0050] Figure 5 An original apparent crack picture of the bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application;

[0051] Figure 6 A predicted apparent crack picture of the bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application;

[0052] Figure 7 A crack prediction Dice coefficient convergence process of the bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application. DETAILED DESCRIPTION

[0053] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0054] It should be noted that although the functional modules are divided in the system schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described can be performed in a manner different from the module division in the system or the order in the flowchart. The terms "first", "second", etc. in the specification and claims and the above drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0055] The present application provides a bridge apparent crack identification method based on adaptive subset search and deep learning, which adopts a global-to-local sampling strategy to select a subset and training samples from a high-speed railway bridge apparent crack data set, and trains a deep neural network model for crack identification. According to the model convergence as the termination condition of active learning, the efficiency and accuracy of high-speed railway bridge apparent crack identification can be effectively improved. The present application can effectively balance the exploration and utilization in the active learning process, avoid the local optimal problem of sampling, and has higher sampling efficiency and crack identification accuracy.

[0056] The embodiments of the present application will be further described below with reference to the accompanying drawings.

[0057] Referring to Figure 1 and Figure 2 , Figure 1 The flowchart of the bridge apparent crack identification method based on adaptive subset search and deep learning provided by an embodiment of the present application is shown in the figure. Figure 2 The flowchart of the bridge apparent crack identification method based on adaptive subset search and deep learning provided by another embodiment of the present application is shown in the figure. The bridge apparent crack identification method based on adaptive subset search and deep learning includes but is not limited to the following steps:

[0058] Step S110, acquiring a high-speed railway bridge apparent crack picture data set, selecting a candidate data set from the picture data set, and selecting a current subset from the candidate data set;

[0059] Step S120, selecting a plurality of samples as new training samples from the candidate data set in a random sampling manner, marking the training samples, and adding them to the training set;

[0060] Step S130, updating the deep neural network model for crack identification using the training set, and calculating the relative loss of the prediction results of the picture data set by the deep neural network model in the adjacent two iteration training processes;

[0061] Step S140, if the relative loss of the prediction result of the picture data set in continuous multiple iterations is less than the first allowable value, the deep neural network model converges.

[0062] Step S150, save the deep neural network model, and use the deep neural network model for high-speed railway bridge apparent crack identification.

[0063] In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning comprises the steps of: obtaining a high-speed railway bridge apparent crack picture data set, selecting a candidate data set from the picture data set, and selecting a current subset from the candidate data set; using a random sampling method to select multiple samples from the candidate data set as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the prediction result of the deep neural network model for the picture data set in the adjacent two iteration training processes; if the relative loss of the prediction result of the picture data set in continuous multiple iterations is less than the first allowable value, the deep neural network model converges; save the deep neural network model, and use the deep neural network model for high-speed railway bridge apparent crack identification.

[0064] The embodiment provides a bridge apparent crack identification method based on adaptive subset search and deep learning, which adopts a whole-to-local sampling strategy to select a subset and training samples from a high-speed railway bridge apparent crack data set, and trains a deep neural network model for crack identification; the application adopts a whole-to-local sampling strategy to select a subset and training samples from a high-speed railway bridge apparent crack data set according to the uncertainty of the samples, and trains a deep neural network model for crack identification; the relative loss of the prediction result of the data in the adjacent two iterations is calculated, and the model convergence and subset convergence are distinguished; the global prediction accuracy of the deep neural network model is measured according to the subset convergence, the sampling range is adaptively reduced, and the exploration and utilization in the active learning process are balanced; the model convergence is used as the termination condition of the active learning, and the efficiency and precision of the high-speed railway bridge apparent crack identification are effectively improved. Compared with the existing method, the application can effectively balance the exploration and utilization in the active learning process, avoid the local optimal problem of sampling, and has higher sampling efficiency and crack identification precision.

[0065] Reference Figure 3 , Figure 3 The embodiment provides a bridge apparent crack identification method based on adaptive subset search and deep learning; the bridge apparent crack identification method based on adaptive subset search and deep learning comprises but is not limited to the following steps:

[0066] Step S210, if the deep neural network model does not satisfy the convergence condition, the predicted label of the current subset in the last iteration of the deep neural network model is used as the real label, and the predicted label of the current subset in the current iteration of the deep neural network model is predicted.

[0067] Step S220, the relative loss of the prediction result of the current subset in the two iterations is calculated according to the real label, the predicted label and the loss function of the network, and it is judged whether the current subset satisfies the convergence condition.

[0068] Step S230, when the relative loss of the prediction result of the current subset in the continuous several iterations is less than the second allowable value, the current subset satisfies the convergence condition, the current subset is deleted from the candidate dataset to reduce the search range of active learning.

[0069] Step S240, the samples of the candidate dataset are predicted by using the deep neural network model, and the uncertainty of the samples of the candidate dataset is calculated based on the information entropy of the prediction result, so as to update the uncertainty of the samples of the candidate dataset.

[0070] In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning comprises the steps of: acquiring a high-speed rail bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; randomly selecting a plurality of samples from the candidate dataset as new training samples in a sampling manner, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the prediction result of the picture dataset in the adjacent two iteration training processes by using the deep neural network model; if the relative loss of the prediction result of the picture dataset in the continuous multiple iterations is less than the first allowable value, the deep neural network model converges; saving the deep neural network model, and using the deep neural network model to identify the high-speed rail bridge apparent crack.

[0071] If the deep neural network model does not satisfy the convergence condition, the predicted label of the current subset in the last iteration of the deep neural network model is used as the real label, and the predicted label of the current subset in the current iteration of the deep neural network model is predicted; the relative loss of the prediction result of the current subset in the two iterations is calculated according to the real label, the predicted label and the loss function of the network, and it is judged whether the current subset satisfies the convergence condition; when the relative loss of the prediction result of the current subset in the continuous several iterations is less than the second allowable value, the current subset satisfies the convergence condition, the current subset is deleted from the candidate dataset to reduce the search range of active learning; the samples of the candidate dataset are predicted by using the deep neural network model, and the uncertainty of the samples of the candidate dataset is calculated based on the information entropy of the prediction result, so as to update the uncertainty of the samples of the candidate dataset.

[0072] Reference Figure 4 to Figure 7 ,Figure 4 A computer vision model schematic diagram of a bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application; Figure 5 A raw apparent crack picture of a bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application; Figure 6 A predicted apparent crack picture of a bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application; Figure 7 A crack prediction Dice coefficient convergence process of a bridge apparent crack identification method based on adaptive subset search and deep learning provided for another embodiment of the present application.

[0073] (1) In an embodiment, a high-speed railway bridge apparent crack identification method based on adaptive subset search and deep learning is used for high-speed railway bridge apparent crack identification.

[0074] (2) First, prepare a high-speed railway bridge apparent crack dataset. The original dataset can be obtained by field shooting. Then, perform preprocessing such as cropping and conversion on the data, so that each picture sample is a 3-channel RGB picture, and adjust all samples to a resolution of 3x256x256. It is worth noting that the samples in the dataset do not need to be labeled in advance, so it is easier to obtain enough samples. Finally, number the samples in the dataset to facilitate network data calling.

[0075] (3) Select a candidate dataset from the dataset. Generally, all datasets can be selected as candidate datasets;

[0076] (4) Select a subset from the candidate dataset. In the first iteration, an empty set can be selected as the current subset, and in other iterations, a certain proportion of samples with smaller uncertainty can be selected from the candidate dataset as the current subset. Generally, 10% of the candidate samples can be taken as the subset each time;

[0077] (5) Select new training samples from the candidate dataset. Random sampling is used to select a small number of samples from the candidate dataset as new training samples each time. In the first iteration, 50 samples can be selected, and 20 samples can be selected each time. Label the new samples and add them to the training set;

[0078] (6) Update the deep neural network model for crack identification using the training set. The present application is suitable for many kinds of deep neural network models currently in use, such as the U-Net deep neural network model. In the first iteration, a U-Net deep neural network model needs to be built as follows: Figure 2The U-Net deep neural network model shown only needs to import the model of the last iteration for retraining in other iterations. The loss function in the training process includes two parts of Dice Loss and cross-entropy, wherein the Dice Loss is defined as

[0079]

[0080] In the formula, S i p and S i g are the predicted and real pixel classifications respectively, and N is the total number of pixels. The definition of the cross-entropy loss function is

[0081]

[0082] In the formula, x is the input and also the output of the last layer of the deep neural network model, and class is the real label index value of the pixel. After defining the network and the loss function, the deep neural network model is trained in a small batch stochastic gradient descent manner;

[0083] (7) The convergence of the deep neural network model is judged. The predicted label of the data set in the last iteration of the deep neural network model is taken as the real label The label of the data set is predicted by the deep neural network model in the current iteration Then, the relative loss loss of the data set in the two iterations is calculated according to the real label, the predicted label and the loss function of the network d When the loss loss data = 0.15 for n d = 2 consecutive iterations is less than the allowable value loss dm = 0.15, it is considered that the model convergence condition is met, otherwise the model convergence condition is not met;

[0084] (8) When the model convergence condition is not met, the next iteration calculation needs to be performed again. First, the subset convergence judgment is performed. The predicted label of the subset in the last iteration of the deep neural network model is taken as the real label The label of the subset is predicted by the deep neural network model in the current iteration Then, the relative loss loss of the predicted results of the subset in the two iterations is calculated according to the real label, the predicted label and the loss function of the network s When the loss loss sub = 0.3 for n s = 1 consecutive iteration is less than the allowable value loss sm = 0.3, it is considered that the subset convergence condition is met, otherwise the subset convergence condition is not met;

[0085] (9) If the subset convergence condition is met, the subset is deleted from the candidate dataset, thereby narrowing the search range of active learning;

[0086] (10) Updating the uncertainty of the candidate dataset samples. The candidate dataset samples are predicted by using the deep neural network model, and the uncertainty of the candidate dataset samples is calculated based on the information entropy of the prediction result. The uncertainty based on the information entropy is defined as

[0087]

[0088] In the formula, p(S i ) is the probability of the pixel being predicted as class i, n is the total number of classes, and N is the total number of pixels of the sample picture;

[0089] (11) According to the uncertainty of the samples, the first N s samples with the minimum uncertainty are selected from the candidate dataset as the subset. Wherein N s

[0090] N s = p s N c (4)

[0091] In the formula, p s is a proportionality coefficient, which can be 0.1; and N c is the number of remaining candidate dataset samples.

[0092] (12) When the model convergence condition is not met, the active learning iteration process is terminated, and the trained deep neural network model is saved.

[0093] (13) The trained deep neural network model is used for high-speed railway bridge apparent crack identification.

[0094] Figure 5 and Figure 6 are the original apparent crack picture and the predicted apparent crack picture, respectively. As can be seen from the figure, the present application has high prediction accuracy. Figure 7 is the crack prediction Dice coefficient convergence process, and as can be seen from the figure, the Dice coefficient of the present application converges quickly, indicating that the present application has high calculation efficiency.

[0095] ​In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning comprises the steps of: obtaining a high-speed railway bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; using random sampling to select multiple samples from the candidate dataset as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the deep neural network model for the picture dataset prediction results in the adjacent two iteration training processes; if the relative loss of the picture dataset prediction results in the continuous iterations is less than a first allowable value, the deep neural network model converges; saving the deep neural network model, and using the deep neural network model to identify the high-speed railway bridge apparent crack.

[0096] If the deep neural network model does not meet the convergence condition, the prediction label of the deep neural network model in the last iteration for the current subset is used as the true label, and the prediction label of the deep neural network model in the current iteration for the current subset is used as the prediction label; the relative loss of the current subset prediction results in the two iterations is calculated according to the true label, the prediction label and the loss function of the network, and whether the current subset meets the convergence condition is judged; if the current subset does not meet the convergence condition, the uncertainty of the candidate dataset samples is calculated based on the information entropy of the prediction results, so as to update the uncertainty of the samples in the candidate dataset.

[0097] In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning comprises the steps of: obtaining a high-speed railway bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; using random sampling to select multiple samples from the candidate dataset as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the deep neural network model for the picture dataset prediction results in the adjacent two iteration training processes; if the relative loss of the picture dataset prediction results in the continuous iterations is less than a first allowable value, the deep neural network model converges; saving the deep neural network model, and using the deep neural network model to identify the high-speed railway bridge apparent crack.

[0098] The relative loss of the deep neural network model for the picture dataset prediction results in the adjacent two iteration training processes comprises the steps of: using the prediction label of the deep neural network model in the last iteration for the picture dataset as the true label, using the prediction label of the deep neural network model in the current iteration for the picture dataset as the prediction label, and then calculating the relative loss of the dataset in the two iterations according to the true label, the prediction label and the loss function of the deep neural network model.

[0099] In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning includes the steps of: obtaining a high-speed rail bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; using random sampling to select multiple samples from the candidate dataset as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the deep neural network model for picture dataset prediction results in the adjacent two iteration training processes; if the relative loss of the picture dataset prediction results in the continuous iterations is less than a first allowable value, the deep neural network model converges; saving the deep neural network model, and using the deep neural network model for high-speed rail bridge apparent crack identification.

[0100] The relative loss of the deep neural network model for picture dataset prediction results in the adjacent two iteration training processes is calculated, including the steps of: using the predicted label of the picture dataset in the last iteration of the deep neural network model as the true label, using the predicted label of the picture dataset in the current iteration of the deep neural network model, and then calculating the relative loss of the dataset in the two iterations according to the true label, the predicted label and the loss function of the deep neural network model. The loss function of the deep neural network model includes two parts of Dice Loss loss function and cross-entropy loss function;

[0101] The Dice Loss loss function is defined as:

[0102]

[0103] In the formula, and are the predicted and true pixel classifications, respectively, and N is the total number of pixels;

[0104] The definition of the cross-entropy loss function is:

[0105]

[0106] x is the input and also the output of the last layer of the deep neural network model, and class is the true label index value of the pixel;

[0107] The loss function of the deep neural network model is:

[0108] loss=loss CE +loss Dice .

[0109] In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning comprises the steps of: acquiring a high-speed railway bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; using random sampling to select multiple samples from the candidate dataset as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the deep neural network model for picture dataset prediction results in adjacent two iteration training processes; if the relative loss of the picture dataset prediction results in continuous multiple iterations is less than a first allowable value, the deep neural network model converges; saving the deep neural network model, and using the deep neural network model to identify high-speed railway bridge apparent cracks.

[0110] If the deep neural network model does not meet the convergence condition, the prediction label of the current subset in the last iteration is used as the true label, and the prediction label of the current subset is predicted using the deep neural network model in the current iteration; the relative loss of the prediction results of the current subset in the two iterations is calculated according to the true label, the prediction label and the loss function of the network, and it is judged whether the current subset meets the convergence condition; when the relative loss of the prediction results of the current subset in continuous iterations is less than a second allowable value, the current subset meets the convergence condition, and the current subset is deleted from the candidate dataset to reduce the search range of active learning; the samples of the candidate dataset are predicted using the deep neural network model, and the uncertainty of the samples of the candidate dataset is calculated based on the information entropy of the prediction results to update the uncertainty of the samples of the candidate dataset.

[0111] The uncertainty of the samples of the candidate dataset is calculated based on the information entropy of the prediction results, comprising:

[0112] The uncertainty calculation formula based on information entropy is:

[0113]

[0114] p(S i ) is the probability of pixels being predicted as class i, n is the total number of classes, and N is the total number of pixels in the sample picture.

[0115] In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning includes the steps of: acquiring a high-speed railway bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; using random sampling to select multiple samples from the candidate dataset as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the deep neural network model for picture dataset prediction results in the adjacent two iteration training processes; if the relative loss of the picture dataset prediction results in the continuous multiple iterations is less than a first allowable value, the deep neural network model converges; saving the deep neural network model, and using the deep neural network model for high-speed railway bridge apparent crack identification.

[0116] If the deep neural network model does not meet the convergence condition, the prediction label of the deep neural network model in the last iteration for the current subset is used as the true label, and the prediction label of the deep neural network model in the current iteration for the current subset is used as the prediction label; the relative loss of the current subset prediction results in the two iterations is calculated according to the true label, the prediction label and the network loss function, and it is judged whether the current subset meets the convergence condition; when the relative loss of the current subset prediction results in the continuous multiple iterations is less than a second allowable value, the current subset meets the convergence condition, and the current subset is deleted from the candidate dataset to reduce the search range of active learning; the deep neural network model is used to predict the samples of the candidate dataset, and the uncertainty of the samples of the candidate dataset is calculated based on the information entropy of the prediction results to update the uncertainty of the samples of the candidate dataset.

[0117] According to the uncertainty of the samples, the first N s samples with the minimum uncertainty are selected from the candidate dataset as a new current subset, where N s is defined as:

[0118] N s = p s N c

[0119] p s is a proportionality coefficient, and N c is the number of remaining candidate dataset samples.

[0120] In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning comprises the steps of: obtaining a high-speed railway bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; using random sampling to select multiple samples from the candidate dataset as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the deep neural network model for the picture dataset prediction results in the adjacent two iteration training processes; if the relative loss of the picture dataset prediction results is less than a first allowable value in two consecutive iterations, the first allowable value is 0.15, and the deep neural network model converges; saving the deep neural network model, and using the deep neural network model to identify high-speed railway bridge apparent cracks.

[0121] In an embodiment, the bridge apparent crack identification method based on adaptive subset search and deep learning comprises the steps of: obtaining a high-speed railway bridge apparent crack picture dataset, selecting a candidate dataset from the picture dataset, and selecting a current subset from the candidate dataset; using random sampling to select multiple samples from the candidate dataset as new training samples, marking the training samples, and adding them to the training set; updating the deep neural network model for crack identification using the training set, calculating the relative loss of the deep neural network model for the picture dataset prediction results in the adjacent two iteration training processes; if the relative loss of the picture dataset prediction results is less than a first allowable value in two consecutive iterations, the first allowable value is 0.15, and the deep neural network model converges; saving the deep neural network model, and using the deep neural network model to identify high-speed railway bridge apparent cracks.

[0122] If the deep neural network model does not meet the convergence condition, the prediction label of the current subset in the last iteration of the deep neural network model is used as the true label, and the prediction label of the current subset is predicted using the deep neural network model in the current iteration; the relative loss of the current subset prediction results in the two iterations is calculated according to the true label, the prediction label and the network loss function, and it is judged whether the current subset meets the convergence condition; when the relative loss of the current subset prediction results in one iteration is less than a second allowable value, the second allowable value is 0.3, the current subset meets the convergence condition, the current subset is deleted from the candidate dataset to reduce the search range of active learning; the samples of the candidate dataset are predicted using the deep neural network model, and the uncertainty of the samples of the candidate dataset is calculated based on the information entropy of the prediction results to update the uncertainty of the samples of the candidate dataset.

[0123] The application further provides a bridge apparent crack identification system based on adaptive subset search and deep learning, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the bridge apparent crack identification method based on adaptive subset search and deep learning as described above when executing the computer program.

[0124] The processor and the memory can be connected through a bus or other means.

[0125] The memory, as a non-transitory computer readable storage medium, can be used to store non-transitory software programs and non-transitory computer executable programs. In addition, the memory can include a high-speed random access memory and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory remotely arranged relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0126] The device embodiments described above are only schematic, and units described as separate components can or can not be physically separate, i.e., can be located in one place or distributed over multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment.

[0127] In addition, one embodiment of the application further provides a computer readable storage medium storing computer executable instructions, and the computer executable instructions are executed by a processor or a controller, for example, a processor in the terminal embodiment described above, so that the processor executes the bridge apparent crack identification method based on adaptive subset search and deep learning in the above embodiment.

[0128] Those of ordinary skill in the art will appreciate that all or certain steps, systems in the methods disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Certain physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software can be distributed on computer readable media, which can include computer storage media (or non-transitory media) and communication media (or transitory media). As is well known to those of ordinary skill in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by a computer. Further, it is common knowledge to those of ordinary skill in the art that communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media.

[0129] The above description is specifically explained with reference to the preferred embodiments of the present application, but the present application is not limited to the above-described embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application, and these equivalent modifications or replacements are included in the scope of the present application defined by the claims.

[0130] The above description of the specific embodiments of the present application does not constitute a limitation on the scope of protection of the present application. Any various other corresponding changes and modifications made in accordance with the technical concept of the present application should be included in the scope of protection of the claims of the present application.

Claims

1. A method for identifying apparent cracks in bridges based on adaptive subset search and deep learning, characterized in that, Includes the following steps: Obtain a dataset of images showing apparent cracks in high-speed railway bridges, select candidate datasets from the image dataset, and then select the current subset from the candidate datasets; Multiple samples are selected from the candidate dataset using random sampling as new training samples, and the training samples are labeled and added to the training set. The deep neural network model for crack identification is updated using the training set, and the relative loss of the deep neural network model on the prediction results of the image dataset is calculated in two adjacent training iterations. If the relative loss of the image dataset prediction results is less than the first allowable value in multiple consecutive iterations, the deep neural network model converges. Save the deep neural network model and use the deep neural network model to identify apparent cracks in high-speed railway bridges; If the deep neural network model does not meet the convergence condition, the predicted label of the current subset by the deep neural network model in the previous iteration is used as the true label, and the predicted label of the current subset is predicted by the deep neural network model in the current iteration. Calculate the relative loss of the prediction results of the current subset in two iterations based on the real label, the predicted label, and the network's loss function, and determine whether the current subset satisfies the convergence condition; If the relative loss of the prediction results of the current subset is less than the second tolerance value in several consecutive iterations, the current subset satisfies the convergence condition, and the current subset is deleted from the candidate dataset to narrow the search scope of active learning. A deep neural network model is used to predict the samples of the candidate dataset, and the uncertainty of the candidate dataset samples is calculated based on the information entropy of the prediction results in order to update the uncertainty of the samples of the candidate dataset. After updating the uncertainty of the samples in the candidate dataset, the steps include: Based on the uncertainty of the samples, the top candidates with the least uncertainty are selected from the candidate dataset. 1 sample is used as the new current subset, where Defined as: This is the proportionality coefficient. This represents the number of remaining candidate dataset samples.

2. The bridge apparent crack identification method based on adaptive subset search and deep learning according to claim 1, characterized in that, After determining whether the current subset satisfies the convergence condition, the following steps are included: If the current subset does not meet the convergence condition, the uncertainty of the candidate dataset samples is calculated based on the information entropy of the prediction results, so as to update the uncertainty of the candidate dataset samples.

3. The bridge apparent crack identification method based on adaptive subset search and deep learning according to claim 1, characterized in that, Calculating the relative loss of the deep neural network model's prediction results for the image dataset during two adjacent training iterations includes the following steps: The predicted labels of the image dataset by the deep neural network model in the previous iteration are used as the real labels. The predicted labels of the image dataset are predicted by the deep neural network model in the current iteration. Then, the relative loss of the dataset in the two iterations is calculated based on the real labels, the predicted labels, and the loss function of the deep neural network model.

4. The bridge apparent crack identification method based on adaptive subset search and deep learning according to claim 3, characterized in that, The loss function of the deep neural network model includes two parts: the Dice Loss function and the cross-entropy loss function. The loss function of the deep neural network model is: 。 5. The bridge apparent crack identification method based on adaptive subset search and deep learning according to claim 1, characterized in that, If the relative loss of the image dataset prediction results is less than a first tolerance value in multiple consecutive iterations, the deep neural network model converges, including the following steps: If the relative loss of the image dataset prediction results is less than the first tolerance value (0.15) in two consecutive iterations, the deep neural network model converges.

6. The bridge apparent crack identification method based on adaptive subset search and deep learning according to claim 1, characterized in that, When the relative loss of the prediction results of the current subset is less than the second allowable value in several consecutive iterations, the current subset satisfies the convergence condition, including the following steps: The current subset satisfies the convergence condition when the relative loss of the prediction result of the current subset in a consecutive iteration is less than the second tolerance value, which is 0.

3.

7. A bridge apparent crack identification system based on adaptive subset search and deep learning, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the bridge apparent crack identification method based on adaptive subset search and deep learning as described in any one of claims 1 to 6.

Images