Intelligent evaluation method for permeability coefficient of asphalt pavement based on residual network
Through the intelligent evaluation method based on residual network, the water permeability coefficient of asphalt pavement is quickly evaluated using texture images, solving the problems of cumbersome operation and low efficiency of traditional methods, and achieving fast and accurate water permeability coefficient evaluation.
Patent Information
- Application Number
- CN202411860864.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-17
- Publication Date
- 2025-04-11
AI Technical Summary
The traditional asphalt pavement water permeability coefficient evaluation method is cumbersome to operate and inefficiently, which hinders traffic.
Using an intelligent evaluation method based on residual network, the texture image of the asphalt pavement is obtained, the trained residual network is constructed, and the texture image is input to obtain the water permeability level to achieve intelligent evaluation of the water permeability coefficient.
This method does not require on-site equipment testing, and can quickly and accurately evaluate the water permeability coefficient, reduce the occupation of roads, and improve evaluation efficiency and accuracy.
Smart Images
Figure CN119810048A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of pavement permeability analysis, and in particular to an intelligent evaluation method for asphalt pavement permeability coefficient based on a residual network. Background Art
[0002] The permeability of asphalt pavement is affected by many factors, and the water permeability coefficient can directly reflect the permeability of asphalt pavement. It helps to monitor the anti-skid performance of the pavement and timely discover changes in the pavement drainage capacity and related diseases. It plays a great role in extending the life of the pavement and reducing traffic accidents.
[0003] The traditional permeability coefficient evaluation method has the disadvantages of complicated operation, low efficiency, and obstruction of traffic, which is not conducive to practical operation. Summary of the invention
[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an intelligent evaluation method for water permeability coefficient of asphalt pavement based on residual network, which solves the problems of cumbersome operation and low efficiency of the existing water permeability coefficient evaluation method.
[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:
[0006] A method for intelligently evaluating the water permeability coefficient of asphalt pavement based on a residual network is provided, which comprises the following steps:
[0007] S1. Acquire a texture image of the asphalt road surface to be tested;
[0008] S2, construct the trained residual network;
[0009] S3. Using the texture image of the asphalt pavement to be tested as the input of the trained residual network, obtaining the water permeability grade of the asphalt pavement to be tested through the output of the trained residual network, and completing the intelligent evaluation of the water permeability coefficient of the asphalt pavement.
[0010] Furthermore, the specific method for obtaining the texture data of the asphalt pavement to be tested in step S1 includes the following sub-steps:
[0011] S1-1, use the ls-40 3D laser scanner to scan the data collection points of the asphalt pavement to be tested, and capture the texture of each data collection point as a 2048×2048 relative elevation matrix;
[0012] S1-2, calculating the median value of the elevation in each relative elevation matrix and setting the threshold elevation;
[0013] S1-3, denoising the corresponding relative elevation matrix based on the median value of the elevation and the threshold elevation to obtain a denoised relative elevation matrix;
[0014] S1-4, normalizing the denoised relative elevation matrix and converting it into a grayscale image;
[0015] S1-5, separating the macro texture and micro texture in the grayscale image by Fourier transform, and obtaining a macro texture image and a micro texture image;
[0016] S1-6, using Gaussian filtering to denoise the macro texture image to obtain a denoised macro texture image; using wavelet denoising to denoise the micro texture image to obtain a denoised micro texture image;
[0017] S1-7, reconstructing the denoised macro texture image and the denoised micro texture image to obtain a reconstructed texture image;
[0018] S1-8. Perform Gaussian blur on the reconstructed texture image to obtain a Gaussian blurred texture image, that is, a texture image of the asphalt pavement to be tested.
[0019] Furthermore, in step S1-3, the specific method for denoising the corresponding relative elevation matrix based on the median value of the elevation and the threshold elevation is:
[0020] Data points in the relative elevation matrix that exceed the threshold elevation are identified as outliers, and the values of the outliers are replaced with the median elevation of the relative elevation matrix in which they are located; the threshold elevation is 2000.
[0021] Furthermore, the specific method of normalizing the relative elevation matrix after denoising in step S1-4 is:
[0022] Normalize the denoised relative elevation matrix to between (-1,1).
[0023] Furthermore, in step S1-6, the expression for reconstructing the denoised macro texture and the denoised micro texture is:
[0024] Z1=H1+W1
[0025] Among them, Z1 is the reconstructed texture data; H1 is the macro texture after denoising; W1 is the micro texture after denoising.
[0026] Furthermore, the residual network includes a convolutional layer, a batch normalization layer, a ReLU activation layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer, a random dropout layer, a global average pooling layer, a flattening layer and a fully connected layer, which are connected in sequence.
[0027] Further, the first residual layer includes three first residual modules connected in sequence;
[0028] Each first residual module includes a first branch and a second branch; the first branch of the first residual module includes a first convolution unit, a first batch normalization unit, a first ReLU activation unit, a second convolution unit, a second batch normalization unit, and a second ReLU activation unit connected in sequence; the second branch of the first residual module includes a third convolution unit and a third batch normalization unit connected in sequence;
[0029] In the same first residual module, the output of the second ReLU activation unit and the output of the third batch normalization unit are respectively used as the input of the splicing unit; the output of the splicing unit is connected to the fourth batch normalization unit and the third ReLU activation unit in sequence; the input end of the first convolution unit and the input end of the third convolution unit are jointly used as the input end of the first residual module to which they are located; the output end of the third ReLU activation unit is the output end of the first residual module to which it is located.
[0030] Further, the second residual layer includes four second residual modules connected in sequence; the third residual layer includes six second residual modules connected in sequence; the fourth residual layer includes three second residual modules connected in sequence;
[0031] Each second residual module adds a downsampling unit in the second branch of the first residual module, and the downsampling unit is added between the input end of the first residual module and the input end of the third convolution unit.
[0032] Furthermore, the specific method for training the residual network includes the following steps:
[0033] A1. Select a test point on the experimental asphalt road section and obtain a texture image of the test point using the same method as step S1;
[0034] A2. Place the HDSS-Ⅲ asphalt pavement water permeability tester on the test point for testing to obtain the true water permeability coefficient at the test point;
[0035] A3. Classify the real water permeability coefficient at the test point according to the water permeability and remove the outliers to obtain the real label data;
[0036] A4. Perform data enhancement on the texture image of the test point to obtain several training samples;
[0037] A5. Use the training samples to train the residual network and obtain the water permeability level output by the residual network;
[0038] A6. Obtain the loss value of the residual network according to the water permeability level output by the residual network and the true label data;
[0039] A7. Based on the loss value of the residual network, the parameters to be learned of the residual network are optimized until the training is completed to obtain the trained residual network.
[0040] Furthermore, during the training of the residual network, the initial learning rate was 0.001. When the performance indicators of the residual network did not improve in 5 training cycles, the learning rate was adjusted by 0.5, and the minimum learning rate was 0.000001. The number of training samples input into the residual network in a single training cycle was 8; the maximum number of training times was 50.
[0041] The beneficial effects of the present invention are:
[0042] 1. This method only needs to obtain the texture image of the asphalt pavement to be tested, and the permeability coefficient can be evaluated through the trained residual network. There is no need to test the permeability coefficient through equipment on site, which can reduce the occupation of the road to a greater extent and realize the rapid and accurate evaluation of the permeability coefficient.
[0043] 2. This method first denoises the scanning data of the LS-40 3D laser scanner, and then converts it into a grayscale image after normalization. The macro-texture and micro-texture in the grayscale image are separated by Fourier transform. Different methods are used to denoise the macro-texture image and the micro-texture image respectively, which can achieve smoothing of the macro-texture image and highlight the details of the micro-texture image. The denoised macro-texture image and the denoised micro-texture image are then reconstructed and Gaussian blurred, so that the texture image of the asphalt pavement to be tested has a clearer feature expression capability, which is convenient for the residual network to classify and identify, and improves the classification accuracy of the residual network.
[0044] 3. The residual network has a simple structure, involves relatively small network parameters and computational complexity, and helps improve evaluation speed.
[0045] 4. The training process of the residual network by this method can make the accuracy of the residual network increase and tend to be stable with the training iteration process, so that the residual network has good evaluation ability for the texture images of different asphalt pavements to be tested. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 is a schematic diagram of the process of this method;
[0047] Figure 2 is a grayscale picture of an asphalt road surface in the embodiment;
[0048] Figure 3 It is a schematic diagram of the structure of the residual network;
[0049] Figure 4 is a schematic diagram of the structure of the second residual module;
[0050] Figure 5 A schematic diagram of data augmentation of texture images of test points by random cutting during the training of the residual network;
[0051] Figure 6 This is a graph showing changes in the accuracy of water permeability prediction of the residual network during the training process. DETAILED DESCRIPTION
[0052] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0053] like Figure 1 As shown in FIG. 1 , the intelligent evaluation method for asphalt pavement permeability coefficient based on residual network includes the following steps:
[0054] S1. Acquire a texture image of the asphalt road surface to be tested;
[0055] S2, construct the trained residual network;
[0056] S3. Using the texture image of the asphalt pavement to be tested as the input of the trained residual network, obtaining the water permeability grade of the asphalt pavement to be tested through the output of the trained residual network, and completing the intelligent evaluation of the water permeability coefficient of the asphalt pavement.
[0057] The specific method for obtaining the texture data of the asphalt pavement to be tested in step S1 includes the following sub-steps:
[0058] S1-1, use the ls-40 3D laser scanner to scan the data collection points of the asphalt pavement to be tested, and capture the texture of each data collection point as a 2048×2048 relative elevation matrix;
[0059] S1-2, calculating the median value of the elevation in each relative elevation matrix and setting the threshold elevation;
[0060] S1-3, denoising the corresponding relative elevation matrix based on the median value of the elevation and the threshold elevation to obtain a denoised relative elevation matrix;
[0061] S1-4, the denoised relative elevation matrix is normalized and converted into a grayscale image, such as Figure 2 As shown;
[0062] S1-5, separating the macro texture and micro texture in the grayscale image by Fourier transform, and obtaining a macro texture image and a micro texture image;
[0063] S1-6, using Gaussian filtering to denoise the macro texture image to obtain a denoised macro texture image; using wavelet denoising to denoise the micro texture image to obtain a denoised micro texture image;
[0064] S1-7, reconstructing the denoised macro texture image and the denoised micro texture image to obtain a reconstructed texture image;
[0065] S1-8. Perform Gaussian blur on the reconstructed texture image to obtain a Gaussian blurred texture image, that is, a texture image of the asphalt pavement to be tested.
[0066] In the specific implementation process, the specific method for denoising the corresponding relative elevation matrix based on the median elevation and the threshold elevation in step S1-3 is: the data points in the relative elevation matrix that exceed the threshold elevation are judged as outliers, and the values of the outliers are replaced by the median elevation of the relative elevation matrix in which they are located; the threshold elevation is 2000.
[0067] The specific method of normalizing the denoised relative elevation matrix in step S1-4 is: normalizing the denoised relative elevation matrix to between (-1, 1).
[0068] The expression for reconstructing the denoised macro texture and the denoised micro texture in step S1-6 is:
[0069] Z1=H1+W1
[0070] Among them, Z1 is the reconstructed texture data; H1 is the macro texture after denoising; W1 is the micro texture after denoising.
[0071] like Figure 3 and Figure 4 As shown in the figure, the residual network includes sequentially connected convolutional layers (Conv7×7, c64, s2, p3), batch normalization layers (BatchNorm1), ReLU activation layers (ReLU), the first residual layer (Layer1), the second residual layer (Layer2), the third residual layer (Layer3), the fourth residual layer (Layer4), a random inactivation layer (Dropout), a global average pooling layer (AvgPool), a flattening layer (Flatten) and a fully connected layer (FC).
[0072] The first residual layer includes three first residual modules connected in sequence;
[0073] Each first residual module includes a first branch and a second branch; the first branch of the first residual module includes a first convolution unit, a first batch normalization unit, a first ReLU activation unit, a second convolution unit, a second batch normalization unit, and a second ReLU activation unit connected in sequence; the second branch of the first residual module includes a third convolution unit and a third batch normalization unit connected in sequence;
[0074] In the same first residual module, the output of the second ReLU activation unit and the output of the third batch normalization unit are respectively used as the input of the splicing unit; the output of the splicing unit is connected to the fourth batch normalization unit and the third ReLU activation unit in sequence; the input end of the first convolution unit and the input end of the third convolution unit are jointly used as the input end of the first residual module to which they are located; the output end of the third ReLU activation unit is the output end of the first residual module to which it is located.
[0075] The second residual layer includes four second residual modules connected in sequence; the third residual layer includes six second residual modules connected in sequence; the fourth residual layer includes three second residual modules connected in sequence;
[0076] Each second residual module adds a downsampling unit in the second branch of the first residual module, and the downsampling unit is added between the input end of the first residual module and the input end of the third convolution unit.
[0077] In the specific implementation process, the texture image of the asphalt pavement to be tested is input into the convolutional layer of the residual network in the shape of (512, 512, 3) for convolution, batch normalization and ReLU activation, and then processed through four residual layers. The features output by the last residual layer are globally average pooled and flattened once, and classified by the fully connected layer to output the permeability coefficient.
[0078] In this embodiment, the convolution unit and the convolution layer are both used to perform convolution operations; the batch normalization layer and the batch normalization unit are both used to perform batch normalization operations; the ReLU activation layer and the ReLU activation unit are both used to perform ReLU activation operations.
[0079] In this embodiment, the specific method for training the residual network includes the following steps:
[0080] A1. Select a test point on the experimental asphalt road section and obtain a texture image of the test point using the same method as step S1;
[0081] A2. Place the HDSS-Ⅲ asphalt pavement water permeability tester on the test point for testing to obtain the true water permeability coefficient at the test point;
[0082] A3. Classify the real water permeability coefficient at the test point according to the water permeability and remove the outliers to obtain the real label data;
[0083] A4. Perform data enhancement on the texture image of the test point to obtain several training samples;
[0084] A5. Use the training samples to train the residual network and obtain the water permeability level output by the residual network;
[0085] A6. Obtain the loss value of the residual network according to the water permeability level output by the residual network and the true label data;
[0086] A7. Based on the loss value of the residual network, the parameters to be learned of the residual network are optimized until the training is completed to obtain the trained residual network.
[0087] In the specific implementation process, the texture image of the test point is enhanced by random cutting, rotation, flipping, etc. Figure 5 shown.
[0088] During the training of the residual network, the initial learning rate is 0.001. When the performance index of the residual network does not improve in 5 training cycles (epochs), the learning rate is adjusted to 0.5, and the minimum learning rate is 0.000001; the training samples (batch_size) input to the residual network in a single training cycle are 8; the maximum number of training times (Epochs) is 50. When the residual network reaches the maximum number of training times or converges to the set value, the training can be considered to be over.
[0089] In this embodiment, a part of the acquired training samples can be used as a training set, and the other part can be used as a verification set. The residual network is trained by the training set, and the residual network is verified by the verification set. Figure 6 As shown in the figure, the permeability level prediction accuracy (Train Accuracy) of the residual network during the training process increases and tends to be stable with the training iteration process, and the final validation set accuracy (Validation Accuracy) is as high as 81.37%, proving that this method has excellent prediction effect.
[0090] In conclusion, the present invention only needs to obtain the texture image of the asphalt pavement to be tested, and can evaluate the permeability coefficient through the trained residual network. There is no need to test the permeability coefficient through equipment on site, which can reduce the occupation of the road to a greater extent and achieve a fast and accurate evaluation of the permeability coefficient.
Claims
1. An intelligent evaluation method for asphalt pavement permeability coefficient based on residual network, characterized in that: The following steps are involved: S1. Acquire a texture image of the asphalt road surface to be tested; S2, construct the trained residual network; S3. Using the texture image of the asphalt pavement to be tested as the input of the trained residual network, obtaining the water permeability grade of the asphalt pavement to be tested through the output of the trained residual network, and completing the intelligent evaluation of the water permeability coefficient of the asphalt pavement.
2. The method for intelligent evaluation of asphalt pavement permeability coefficient based on residual network according to claim 1 is characterized in that: The specific method for obtaining the texture data of the asphalt pavement to be tested in step S1 includes the following sub-steps: S1-1, use the ls-40 3D laser scanner to scan the data collection points of the asphalt pavement to be tested, and capture the texture of each data collection point as a 2048×2048 relative elevation matrix; S1-2, calculating the median value of the elevation in each relative elevation matrix and setting the threshold elevation; S1-3, denoising the corresponding relative elevation matrix based on the median value of the elevation and the threshold elevation to obtain a denoised relative elevation matrix; S1-4, normalizing the denoised relative elevation matrix and converting it into a grayscale image; S1-5, separating the macro texture and micro texture in the grayscale image by Fourier transform, and obtaining a macro texture image and a micro texture image; S1-6, using Gaussian filtering to denoise the macro texture image to obtain a denoised macro texture image; using wavelet denoising to denoise the micro texture image to obtain a denoised micro texture image; S1-7, reconstructing the denoised macro texture image and the denoised micro texture image to obtain a reconstructed texture image; S1-8. Perform Gaussian blur on the reconstructed texture image to obtain a Gaussian blurred texture image, that is, a texture image of the asphalt pavement to be tested.
3. The intelligent evaluation method for asphalt pavement permeability coefficient based on residual network according to claim 2 is characterized in that: The specific method for denoising the corresponding relative elevation matrix based on the median value and threshold elevation of the elevation in step S1-3 is: Data points in the relative elevation matrix that exceed the threshold elevation are identified as outliers, and the values of the outliers are replaced with the median elevation of the relative elevation matrix in which they are located; the threshold elevation is 2000.
4. The method for intelligent evaluation of asphalt pavement permeability coefficient based on residual network according to claim 2 is characterized in that: The specific method for normalizing the relative elevation matrix after denoising in step S1-4 is: Normalize the denoised relative elevation matrix to between (-1,1).
5. The method for intelligent evaluation of asphalt pavement permeability coefficient based on residual network according to claim 2 is characterized in that: The expression for reconstructing the denoised macro texture and the denoised micro texture in step S1-6 is: Z1=H1+W1 Among them, Z1 is the reconstructed texture data; H1 is the macro texture after denoising; W1 is the micro texture after denoising.
6. The method for intelligent evaluation of asphalt pavement permeability coefficient based on residual network according to claim 1 is characterized in that: The residual network includes a convolutional layer, a batch normalization layer, a ReLU activation layer, a first residual layer, a second residual layer, a third residual layer, a fourth residual layer, a random dropout layer, a global average pooling layer, a flattening layer, and a fully connected layer, which are connected in sequence.
7. The method for intelligent evaluation of asphalt pavement permeability coefficient based on residual network according to claim 6 is characterized in that: The first residual layer includes three first residual modules connected in sequence; Each first residual module includes a first branch and a second branch; the first branch of the first residual module includes a first convolution unit, a first batch normalization unit, a first ReLU activation unit, a second convolution unit, a second batch normalization unit, and a second ReLU activation unit connected in sequence; the second branch of the first residual module includes a third convolution unit and a third batch normalization unit connected in sequence; In the same first residual module, the output of the second ReLU activation unit and the output of the third batch normalization unit are respectively used as the input of the splicing unit; the output of the splicing unit is connected to the fourth batch normalization unit and the third ReLU activation unit in sequence; the input end of the first convolution unit and the input end of the third convolution unit are jointly used as the input end of the first residual module to which they are located; the output end of the third ReLU activation unit is the output end of the first residual module to which it is located.
8. The method for intelligently evaluating water permeability of asphalt pavement based on residual network according to claim 7 is characterized in that: The second residual layer includes four second residual modules connected in sequence; the third residual layer includes six second residual modules connected in sequence; the fourth residual layer includes three second residual modules connected in sequence; Each second residual module adds a downsampling unit in the second branch of the first residual module, and the downsampling unit is added between the input end of the first residual module and the input end of the third convolution unit.
9. The intelligent evaluation method for asphalt pavement permeability coefficient based on residual network according to claim 8 is characterized in that: The specific method for training the residual network includes the following steps: A1. Select a test point on the experimental asphalt road section and obtain a texture image of the test point using the same method as step S1; A2. Place the HDSS-Ⅲ asphalt pavement water permeability tester on the test point for testing to obtain the true water permeability coefficient at the test point; A3. Classify the real water permeability coefficient at the test point according to the water permeability and remove the outliers to obtain the real label data; A4. Perform data enhancement on the texture image of the test point to obtain several training samples; A5. Use the training samples to train the residual network and obtain the water permeability level output by the residual network; A6. Obtain the loss value of the residual network according to the water permeability level output by the residual network and the true label data; A7. Based on the loss value of the residual network, the parameters to be learned of the residual network are optimized until the training is completed to obtain the trained residual network.
10. The intelligent evaluation method for asphalt pavement permeability coefficient based on residual network according to claim 9 is characterized in that: During the training of the residual network, the initial learning rate is 0.
001. When the performance indicators of the residual network do not improve in 5 training cycles, the learning rate is adjusted to 0.5, and the minimum learning rate is 0.000001. The number of training samples input into the residual network in a single training cycle is 8; the maximum number of training times is 50.