Intelligent Identification Method for Road Subsurface Diseases Based on AlexNet Model

Abstract

The present invention discloses an intelligent recognition method for underground road diseases based on the AlexNet model, comprising: S1, collecting three-dimensional ground penetrating radar image data and performing preprocessing to obtain radar maps in two-dimensional directions, namely C-scan maps and B-scan maps; S2, classifying the radar maps corresponding to different target bodies to obtain radar map data sets corresponding to different target bodies; S3, adding class labels to the C-scan maps and B-scan maps of each radar map data set; S4, constructing a cascaded model based on the AlexNet deep network; S5, inputting the data to be predicted into the model and outputting a prediction result. The present invention constructs a cascaded model based on the AlexNet deep network. Inputting the data to be predicted into the deep learning model can obtain accurate prediction results, improve the detection accuracy and detection efficiency, and can intelligently recognize underground road disease bodies with high precision and high accuracy.

Description

Technical Field

[0001] The present invention relates to an intelligent recognition method for underground road diseases based on the AlexNet model. Background Art

[0002] The interpretation of ground penetrating radar data is a process of identifying meaningful medium characteristics and medium distribution laws in radar data, which is an embodiment of the detection target of ground penetrating radar. At present, the interpretation of radar images mainly relies on manual interpretation methods, that is, technicians observe and analyze radar images one by one and obtain the detection results of the target body based on manual experience. However, in the process of comprehensively detecting roads, tunnels, etc. by radar, the detection data faced by interpreters is often extremely large, and the following problems exist when manually interpreting radar images: (1) The professional requirements for the interpretation process are relatively high, and there is a lack of sufficient data interpreters; (2) The interpretation process is subjective. For the same radar image, the interpretation results obtained by different interpreters often vary, making the detection accuracy overly dependent on the technical level of the interpreters; (3) The manual interpretation process is time-consuming, with a huge workload and low interpretation efficiency. These problems have limited the application and popularization of ground penetrating radar technology to a certain extent.

[0003] In recent years, with the rapid development of computer vision algorithms, more and more large-scale information processing tasks in various fields (such as public security, biology, transportation, medical treatment, etc.) have gradually shifted from manual processing to computer automatic processing. Under this background, it is necessary to introduce various image recognition algorithms and technologies in the field of computer vision into the automatic interpretation method of radar images, so as to realize the automatic interpretation of radar data or images. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent recognition method for underground road diseases based on the AlexNet model, which can improve the detection accuracy and detection efficiency.

[0005] The purpose of the present invention is achieved by the following technical measures: An intelligent recognition method for underground road diseases based on the AlexNet model, which is characterized by including the following steps:

[0006] S1. Collect three-dimensional ground penetrating radar image data and perform preprocessing to obtain radar maps in two-dimensional directions, namely C-scan maps and B-scan maps;

[0007] S2. Classify the radar maps corresponding to different target bodies to obtain radar map data sets corresponding to different target bodies;

[0008] S3. Add category labels to the C-scan maps and B-scan maps of each radar map data set;

[0009] S4. Construct a cascaded model based on the AlexNet deep network:

[0010] Input the C-scan atlas with class labels as training samples into the first-level neural network, and train according to the parameters of each layer of this neural network to output prediction values. Update the parameters of each layer of the first-level neural network by continuously reducing the error between the prediction value and the true value, and finally obtain the trained first-level neural network.

[0011] Input the C-scan atlas and B-scan atlas with class labels as training samples into the second-level neural network, and train according to the parameters of each layer of this neural network to output prediction values. Update the parameters of each layer of the second-level neural network by continuously reducing the error between the prediction value and the true value, and finally obtain the trained second-level neural network.

[0012] S5. Input the C-scan atlas of the data to be predicted into the trained first-level neural network for preliminary screening of diseases. If there is no disease, output a healthy prediction result.

[0013] If there is a disease, splice the C-scan atlas and B-scan atlas of the data to be predicted along their length directions and input them into the trained second-level neural network to determine the disease type and output the prediction result.

[0014] The present invention constructs a cascaded model based on the AlexNet deep network, and obtains the deep learning models of the first-level neural network and the second-level neural network through neural network training. After inputting the data to be predicted into this deep learning model, accurate prediction results can be obtained, which not only improves the detection accuracy, but also improves the detection efficiency, and can realize intelligent identification of road underground disease bodies with high precision and high accuracy.

[0015] In the step S4 of the present invention, both the first-level neural network and the second-level neural network include a convolutional layer, a pooling layer and a fully connected layer. Among them, the convolutional operation formula:

[0016] In the formula: Convolve the training sample S ij to obtain the output S i'j' at the corresponding position. i represents the i-th echo, and j represents the j-th point of the echo; w nm represents the parameter of the nth row and mth column of the convolutional filter.

[0017] To speed up the training speed and alleviate the over-fitting situation, the average pooling method is adopted, that is, the pooling layer only calculates the average value of the feature points in the neighborhood to better retain information.

[0018] Using the ReLU activation function: f(x) = max(0, x), the ReLU activation function makes the output of some neurons 0, which creates the sparsity of the network, reduces the interdependence of parameters, and alleviates the occurrence of overfitting problems.

[0019] The bottom layer of the first-level neural network and the second-level neural network is a fully connected layer, and the fully connected layer uses a fully connected output, that is, a weighted sum is performed on each element of the matrix output by the M-1 layer, where is the size of the output of the M-1 layer, and the subscript of x i is used to identify the probabilities corresponding to different classification results of the same training sample, and k nm is the parameter of the nth row and mth column of the output layer of the Mth layer, and e nm is the element of the nth row and mth column of the output matrix of the M-1 layer, and the classification probability matrix of the same training sample obtained at the Mth layer is [x1 x2 x3... x p T , where p represents the number of types required for each training sample.

[0020] The target objects of the present invention are voids (holes), looseness, pipelines, manhole covers, and health.

[0021] In the step S3 of the present invention, a rectangular window with a width of W1 is taken, and a sliding window process is performed with the target object in the C-san map as the center, and a class label is added to the target object in the C-san map; a rectangular window with a width of W2 is taken, and a sliding window process is performed with the target object in the B-san map as the center, and a class label is added to the target object in the B-san map.

[0022] In the step S3 of the present invention, after adding class labels to the target objects in the C-san map and the B-san map, a training set, a test set, and a validation set are determined. Among them, the training samples in the training set are input into the neural network for training.

[0023] Compared with the prior art, the present invention has the following remarkable effects:

[0024] The present invention constructs a cascaded model based on the AlexNet deep network, and obtains a deep learning model of the first-level neural network and the second-level neural network through neural network training. After inputting the data to be predicted into the deep learning model, accurate prediction results can be obtained, which not only improves the detection accuracy, but also improves the detection efficiency, and can realize intelligent identification of road underground disease bodies with high precision and high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0026] ​Figure 1 It is a schematic diagram of the AlexNet neural network of the present invention;

[0027] Figure 2 It is a three-dimensional radar map of the present invention (showing void diseases);

[0028] Figure 3 It is a schematic diagram of the model evaluation accuracy of the present invention. Specific implementation manners

[0029] The present invention will be described in detail below in conjunction with the embodiments and their accompanying drawings to help those skilled in the art better understand the inventive concept of the present invention. However, the protection scope of the claims of the present invention is not limited to the following embodiments. For those skilled in the art, all other embodiments obtained without creative labor under the premise of not departing from the inventive concept of the present invention belong to the protection scope of the present invention.

[0030] An intelligent road underground disease identification method based on the AlexNet framework of the present invention includes the following steps:

[0031] S1. Collect three-dimensional ground penetrating radar image data and perform preprocessing to obtain radar maps in two-dimensional directions, namely C-scan maps and B-scan maps; the three-dimensional ground penetrating radar image data is the three-dimensional radar maps corresponding to road diseases verified by drilling in various cities and counties in Guangdong Province.

[0032] S2. Classify the radar maps corresponding to different target bodies to obtain a radar map dataset corresponding to different target bodies; target body identification: identification of pipelines (level 0), identification of looseness (level 1), identification of voids (level 2), and identification of manhole covers (level 3).

[0033] The radar maps corresponding to different target bodies refer to the radar maps corresponding to common road underground disease bodies such as voids (see Figure 2 ) or cavities, looseness, pipelines, manhole covers, etc., as well as the radar maps of healthy roads, that is, the target bodies include 5 types: healthy, pipelines, manhole covers, voids (cavities), and soil looseness. Regularize and classify the radar maps of these different target bodies, including the C-scan maps and B-scan maps corresponding to each target body.

[0034] S3. Perform preprocessing on the radar maps in the dataset, that is, add class labels to the C-scan maps and B-scan maps of each radar map dataset;

[0035] Specifically, take a rectangular window with a width of W1 and perform a sliding window process centered on the target object in the C-san atlas, and add YoLo labels to the target objects in the C-san atlas; take a rectangular window with a width of W2 and perform a sliding window process centered on the target object in the B-san atlas, and add category labels to the target objects in the B-san atlas.

[0036] After adding category labels to the target objects in the C-san atlas and the B-san atlas, determine the training set, test set, and validation set. Among them, the samples in the training set are training samples, and the samples in the test set and validation set are used to evaluate the quality of the model. This is a common practice in this field. See Figure 3 , the evaluation accuracy of this model. The accuracy in the training set is 100%, and the accuracies in the test set and validation set are both 95%. Therefore, when the data to be predicted is input into the deep learning model completed after training in the present invention, the accuracy of the output prediction result is very high.

[0037] S4. Construct a cascaded model based on the AlexNet deep network, which includes two levels of neural networks, namely the first-level neural network and the second-level neural network, as Figure 1 shown. The figure shows the structures of the convolutional layer Conv.Layer and the pooling layer Maxpool Layer, where:

[0038] Input the training samples of the C-scan atlas with added category labels into the first-level neural network, and train according to the parameters of each layer of this neural network to output predicted values. Update the parameters of each layer of the first-level neural network by continuously reducing the error between the predicted value and the true value, and finally obtain the trained first-level neural network. Specifically:

[0039] The first-level neural network has L layers, including convolutional layers, pooling layers, and fully connected layers. In the first-level neural network, the convolutional layer is the L-2 layer, and the last two layers are fully connected layers. The Lth layer finally outputs the predicted target.

[0040] Convolution operation formula:

[0041] In the formula: Convolve the training sample S ij to obtain the output S at the corresponding position i'j' , i represents the ith echo, j represents the jth point of the echo; w nm represents the parameter of the nth row and mth column of the convolution filter; adjusting the size of w controls the size of the convolution kernel. Since the deep network model has a large number of convolutional layers, a large number of parameters will be generated, which not only has a large amount of redundant information affecting the calculation efficiency, but also easily causes overfitting of the neural network. To speed up the training speed and alleviate the overfitting situation, the method of average pooling is adopted, that is, only the average of the feature points in the neighborhood is calculated, and the information is better retained.

[0042] The function used in this level of network is the ReLu activation function, and its expression is as follows:

[0043] f(x) = max(0, x) Formula ⑵

[0044] ReLu will make the output of some neurons be 0, which results in the sparsity of the network, reduces the interdependence of parameters, and alleviates the occurrence of overfitting problems.

[0045] The bottom layer (the Mth layer and the (M - 1)th layer) of the deep model uses a fully connected output, that is, a weighted sum is performed on each element of the matrix output by the (M - 1)th layer, where is the size of the output of the (M - 1)th layer, and the subscript of x i is used to identify the probabilities corresponding to different classification results of the same training sample, and k nm is the parameter of the nth row and mth column of the output layer of the Mth layer, and e nm is the element of the nth row and mth column of the output matrix of the (M - 1)th layer. The classification probability matrix of the same training sample obtained at the Mth layer is [x1 x2 x3... x p T , where p represents the number of types required for each training sample.

[0046] Input the training samples of the C-scan atlas and B-scan atlas with added class labels into the second-level neural network, and train according to the parameters of each layer of this neural network to output a predicted value. Update the parameters of each layer of the second-level neural network by continuously reducing the error between the predicted value and the true value. Finally, obtain the trained second-level neural network. The difference from obtaining the trained first-level neural network is that the training samples of the C-scan atlas and B-scan atlas with added class labels are input into the second-level neural network, and the input is the radar atlas in two dimensions, that is, the B-Scan and C-scan atlases are spliced along the length direction and then input.

[0047] Before inputting the training samples into the neural network, first set the training parameters: the number of training steps epoch, the learning rate learning_rate, and the sample batch size batch_size.

[0048] The training samples are trained according to the initialized parameters of each layer of the neural network (such as the convolution kernel size, etc.). Finally, a predicted matrix X is output through the fully connected layer, and the error F - X from the true value F is obtained. Use the gradient descent method to reduce the error and update the parameters of each layer of the neural network.

[0049] By continuously iterating the number of training steps and continuously updating the parameters of each convolution layer, after obtaining iteration stability, finally obtain the trained deep learning model.

[0050] ​S5. Input the C-scan map of the data to be predicted into the trained first-level neural network to preliminarily screen for diseases. If there are no diseases, output the healthy prediction result;

[0051] If there are diseases, splice the C-scan map and B-scan map of the data to be predicted along their length directions, then input them into the trained second-level neural network to determine the disease type and output the prediction result.

Claims

1. An intelligent identification method for underground road diseases based on the AlexNet model, characterized in that It includes the following steps: S1. Collect three-dimensional ground penetrating radar image data and perform preprocessing to obtain radar maps in two-dimensional directions, namely C-scan maps and B-scan maps; S2. Classify the radar maps corresponding to different target bodies to obtain radar map data sets corresponding to different target bodies; S3. Add class labels to the C-scan maps and B-scan maps of each radar map data set; S4. Construct a cascaded model based on the AlexNet deep network: Use the C-scan maps with added class labels as training samples to input into the first-level neural network, and train according to the parameters of each layer of this neural network to output prediction values. Continuously reduce the error between the prediction value and the true value to update the parameters of each layer of the first-level neural network, and finally obtain the trained first-level neural network; Use the C-scan maps and B-scan maps with added class labels as training samples to input into the second-level neural network, and train according to the parameters of each layer of this neural network to output prediction values. Continuously reduce the error between the prediction value and the true value to update the parameters of each layer of the second-level neural network, and finally obtain the trained second-level neural network; S5. Input the C-scan map of the data to be predicted into the trained first-level neural network for preliminary screening of diseases. If there is no disease, output a healthy prediction result; If there is a disease, splice the C-scan map and B-scan map of the data to be predicted along their length directions and input them into the trained second-level neural network to determine the disease type and output the prediction result.

2. The intelligent identification method for underground road diseases based on the AlexNet model according to claim 1, wherein: In the step S4, both the first-level neural network and the second-level neural network include a convolutional layer, a pooling layer, and a fully connected layer, where Convolution operation formula: Where: for the training sample S ij perform convolution to obtain the output S at the corresponding position i'j' , i represents the i-th echo, j represents the j-th point of the echo; w nm represents the parameter of the n-th row and m-th column of the convolution filter; the pooling layer calculates the average value of the feature points in the neighborhood; Use the reLu activation function: f(x) = max(0, x) formula ⑵ The bottom layer of the first-level neural network and the second-level neural network is a fully connected layer, and the fully connected layer uses a fully connected output, that is, each element of the matrix output by the M-1 layer is weighted and summed, where is the size of the output of the M-1 layer, and the subscript of x i is used to identify the probabilities corresponding to different classification results of the same training sample, and k nm is the parameter of the nth row and mth column of the output layer of the Mth layer, and e nm is the element of the nth row and mth column of the output matrix of the M-1 layer. The classification probability matrix of the same training sample obtained at the Mth layer is [x1 x2 x3... x p T , where p represents the number of types required for each training sample.​ 3. The intelligent road underground disease identification method based on the AlexNet model according to claim 2, characterized in that: The target bodies are voids or cavities, looseness, pipelines, manhole covers, and health.

4. The intelligent road underground disease recognition method based on the AlexNet model according to claim 3, characterized in that: In the step S3, take a rectangular window with a width of W1, perform a sliding window process with the target body in the C-san map as the center, and add class labels to the target body in the C-san map; take a rectangular window with a width of W2, perform a sliding window process with the target body in the B-san map as the center, and add class labels to the target body in the B-san map.

5. The intelligent identification method for underground road diseases based on the AlexNet model according to claim 4, characterized in that: In the step S3, after adding class labels to the target bodies in the C-san map and B-san map, determine the training set, test set, and validation set, where the training samples in the training set are input into the neural network for training.

Images