Texture parameter detection method suitable for asphalt pavement

Abstract

The invention mainly relates to the technical field of pavement detection. In order to improve the accuracy and simplicity of an asphalt pavement texture detection result, the invention provides a texture parameter detection method suitable for an asphalt pavement, and the method comprises the following steps: S1, building an asphalt texture image data set; s2, establishing an asphalt pavement texture parameter detection model, and training the established asphalt pavement texture parameter detection model based on the asphalt texture image data set established in the step S1; and S3, acquiring a to-be-detected asphalt pavement texture image, inputting the to-be-detected asphalt pavement texture image into the asphalt pavement texture parameter detection model established in the step S2, and outputting a pavement texture detection result, so that the asphalt pavement texture parameters can be simply and quickly detected, and the method can be suitable for texture parameter detection of various types of indoor and outdoor asphalt pavements.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of road surface detection, and in particular to a texture parameter detection method suitable for asphalt road surfaces. Background Art

[0002] Existing asphalt pavement texture detection methods can be divided into two categories: contact and non-contact. Existing asphalt pavement texture detection methods can be divided into two categories: contact and non-contact.

[0003] Traditional contact measurement methods such as needle combs and surface profilers have high requirements on the test platform, long test time, and are prone to damage to the original texture of the road surface.

[0004] The main non-contact measurement methods include laser scanning, photometric stereo, digital image, binocular stereo vision, etc.

[0005] Among them, laser detection equipment has high precision and good stability, but it is usually used for indoor detection and the equipment is expensive. Because its laser beam is emitted vertically downward, it is often unable to accurately detect the curved channels and cracks inside the asphalt mixture texture. At the same time, the diode detector receiving board is easily disturbed by ambient light, causing measurement fluctuations, thereby affecting the measurement accuracy; the digital image method has low equipment requirements and short detection time, but the accuracy is not high and it is easily affected by the external environment, such as natural light intensity and road color; the photometric stereo method applies multiple directional illumination to the road surface to be tested, and then takes a two-dimensional image. The directional gradient data of each point is calculated through the light intensity reflected by the image to obtain the three-dimensional information of the road surface texture, and then reconstruct the three-dimensional model. Although the accuracy is high, it is easily disturbed by ambient light and the detection takes a long time; binocular stereo vision completes high-precision distance measurement by simulating human eyes, with moderate cost and high detection efficiency, but the detection environment is harsh, with high requirements for programming foundation, and a series of post-production software is required to reconstruct the three-dimensional model. Summary of the invention

[0006] Technical Problems to be Solved by the Invention

[0007] A texture parameter detection method suitable for asphalt pavement is provided to improve the efficiency and simplicity of asphalt pavement texture detection results.

[0008] The technical solution adopted by the present invention to solve the above technical problems

[0009] A texture parameter detection method applicable to asphalt pavement comprises the following steps:

[0010] Step S1: Establishing an asphalt pavement texture image dataset;

[0011] Step S2: establishing an asphalt pavement texture parameter detection model, and training the established asphalt pavement texture parameter detection model based on the asphalt pavement texture image dataset established in step S1;

[0012] Step S3: Collect the asphalt pavement texture image to be detected, input it into the trained asphalt pavement texture parameter detection model, and output the pavement texture detection result.

[0013] Furthermore, step S1 specifically includes: using a camera to collect asphalt pavement texture images of different gradations and different wear stages, performing image segmentation processing on the collected asphalt pavement texture images, using a laser scanner to perform texture detection on the segmented images, recording texture parameters of the corresponding images, and setting different labels for different texture parameters.

[0014] Furthermore, the method also includes: in step S1, when using a camera to collect asphalt pavement texture images of different gradations and different wear stages, setting a shadowless lamp to perform brightness correction on the asphalt pavement area to be collected.

[0015] Furthermore, the establishment of the asphalt pavement texture parameter detection model in step S2 specifically includes:

[0016] Step S21: Use Python language to call the PyTorch tool library in the PyCharm interpreter;

[0017] Step S22: Create an ImageFolder dataset of PyTorch, which contains images in the asphalt texture image dataset;

[0018] Step S23: define helper function;

[0019] Step S24: Create a base class of the PyTorch model named FingerModelBase, where the FingerModelBase base class includes the nn.Module class in PyTorch;

[0020] Step S25: Setting the asphalt pavement texture parameter detection model structure;

[0021] Step S26: setting an optimizer, and using the optimizer to optimize the asphalt pavement texture parameter detection model;

[0022] Step S27: Based on the asphalt pavement texture image dataset established in step S1, the established asphalt pavement texture parameter detection model is trained, and the training results are evaluated by confusion matrix and TOP-K accuracy. If the evaluation results meet the settings, the asphalt pavement texture parameter detection model training is completed.

[0023] Furthermore, step S3 includes: pre-processing the asphalt pavement texture image to be detected and inputting the trained asphalt pavement texture parameter detection model, outputting the prediction probability of various labels, and the highest probability label is the asphalt pavement texture parameter detection result value.

[0024] Beneficial Effects of the Invention

[0025] The texture parameter detection method for asphalt pavement described in the present invention selects two-dimensional images of texture morphology of various graded asphalt mixtures and their corresponding texture parameter values ​​to construct a deep learning data set, builds an initial model based on the PyTorch framework, and sets an optimizer during the training process to reduce the model training gradient, so that the model can finally achieve accurate measurement of asphalt pavement texture parameters. Compared with the existing technology, the cost is extremely low, the operation is simple, the detection efficiency is high, and the detection accuracy can be controlled by the number of data set labels. It is suitable for texture parameter detection of various types of indoor and outdoor asphalt pavements. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The texture image dataset of the partial asphalt road surface;

[0027] Figure 2 (a), (b), (c), and (d) are the AC dataset accuracy graphs of asphalt pavement texture parameter detection models using Adam, RMSprop, SGD, and Adadelta optimizers, respectively;

[0028] Figure 3 (a), (b), (c), and (d) are the accuracy graphs of the SMA dataset using the Adam, RMSprop, SGD, and Adadelta optimizers for asphalt pavement texture parameter detection models, respectively;

[0029] Figure 4 (a), (b), (c), and (d) are the accuracy graphs of the asphalt pavement texture parameter detection model using the Adam, RMSprop, SGD, and Adadelta optimizers for the OGFC dataset, respectively. DETAILED DESCRIPTION

[0030] The texture parameter detection method applicable to asphalt pavement described in the present invention specifically includes the following steps.

[0031] Step S1: Establish an asphalt pavement texture image dataset.

[0032] First, a high-resolution camera was used to obtain the texture images of asphalt pavement areas with different gradations and different wear stages. At the same time, a shadowless lamp was introduced to improve the brightness stability of the asphalt pavement and avoid the influence of changes in light intensity on the quality of the collected texture images. During the image acquisition process, the vertical height between the camera and the specimen was kept at 35 cm. The camera's built-in software was used to correct the shooting angle and ensure that the lens plane was parallel to the asphalt mixture specimen as much as possible. At the same time, before acquiring the image, the emery on the asphalt surface and the residues such as rubber tire powder produced by wear and tear need to be carefully processed.

[0033] In order to increase the number of samples and ensure the balance of samples, further image segmentation processing is required based on the texture images of asphalt pavement areas with different wear degrees and different gradations collected by the camera. Each specimen photo is segmented into 6 images with a size of 300×100. Some images in the asphalt pavement texture image dataset are as follows: Figure 1 shown.

[0034] Finally, a laser scanner is used to detect the texture parameters of each segmented asphalt pavement area, and the texture parameter value of each image is recorded. Labels are divided according to needs. The more labels there are, the higher the accuracy of the asphalt pavement texture parameter detection model.

[0035] Step S2: Establishing an asphalt pavement texture parameter detection model, and training the established asphalt pavement texture parameter detection model based on the asphalt texture image dataset established in step S1.

[0036] (1) Use Python language to call PyTorch and other tool libraries in the PyCharm interpreter

[0037] (2) Create a PyTorch ImageFolder dataset, which will contain images located in a given folder. To expand the dataset and maintain the balance of data samples, a series of transformations such as random rotation, flipping, and grayscale are performed to enhance the images, and finally these transformations are combined together through the tt.Compose function.

[0038] (3) Define some helper functions, then call the device depending on whether the GPU is available, and finally define new functions to move the data and model to the device.

[0039] (4) Create a base class of the PyTorch model named FingerModelBase. This class inherits the nn.Module class in PyTorch and defines the following main functions:

[0040] ① training_step: used to load batches of training images and output their loss, accuracy, and predictions. This function accepts training data and weight parameters and uses the model to generate predictions for training images. Then, the loss is calculated using the cross entropy loss function with weights. Finally, the function calculates the accuracy by calling the accuracy function.

[0041] ②train_epoch_end: used to calculate the average loss and accuracy of training for each epoch. This function extracts the loss from the output of each batch and then merges them into the loss of one batch. Similarly, this function also extracts the accuracy of each batch and merges them into the accuracy of one batch.

[0042] ③validation_step: used to load batches of validation / test images and output their loss, accuracy, predictions, and labels. This function is similar to the training_step function, but no weight parameters are required when calculating the loss.

[0043] ④validation_epoch_end: used to calculate the validation average loss and accuracy of each epoch. The implementation of this function is similar to the train_epoch_end function.

[0044] (5) Set the network structure, where the conv1 convolution layer uses 64 7x7 convolution kernels with a step size of 2 to reduce the 224x224 color image to 112x112. The conv2_x layer performs 3x3 maximum pooling with a step size of 2 to further reduce the dimension to 56x56, followed by three residual learning modules, each of which consists of three convolution layers with convolution kernel sizes of 1x1 for 64 channels, 3x3 for 64 channels, and 1x1 for 256 channels. The conv3_x layer consists of 4 residual learning modules. Similarly, the conv4_x layer consists of 6 residual learning modules, and the conv5_x layer consists of 3 residual learning modules. The last one is a fully connected layer, which sets the number of output categories according to the accuracy.

[0045] (6) Four optimizers, Adam, RMSprop, SGD, and Adadelta, were set to optimize the pre-trained model for asphalt pavement texture parameter detection. The accuracy of the asphalt pavement texture parameter detection model using the AC (dense-graded asphalt concrete) dataset under each optimizer is as follows: Figure 2 As shown in (a), (b), (c), and (d), the accuracy of the asphalt pavement texture parameter detection model using the SMA (mashy asphalt macadam mixture) dataset under each optimizer is as follows Figure 3As shown in (a), (b), (c), and (d), the accuracy of the asphalt pavement texture parameter detection model using the OGFC (open-graded asphalt mixture) dataset under each optimizer is as follows Figure 4 (a), (b), (c), and (d). The optimizers and initial hyperparameters are shown below:

[0046] ①Adam optimizer: class torch.optim.Adam(params, lr=0.001, betas=(0.9,0.999), eps=1e-08, weight_decay=0.0001, amsgrad=False)

[0047] ②RMSprop optimizer: class torch.optim.RMSprop(params,lr=0.001,alpha=0.99,eps=1e-08,weight_decay=0.001,momentum=0,centered=False)

[0048] ③SGD ​​optimizer: class torch.optim.SGD(params, lr=0.001, momentum=0, dampening=0, weight_decay=0.0001, nesterov=False)

[0049] ④Adadelta optimizer: class torch.optim.Adadelta(params,lr=0.001,rho=0.9,eps=1e-08,weight_decay=0.0001)

[0050] (7) Train and output results, and evaluate the model through confusion matrix and TOP-K accuracy.

[0051] Step S3: Collect the asphalt pavement texture image to be detected, input the asphalt pavement texture image to be detected into the asphalt pavement texture parameter detection model established in step S2, and output the pavement texture detection result.

[0052] Use a mobile phone or digital camera to collect the texture of the asphalt pavement to be tested. The shooting height is 20 to 50 cm. Transfer the image to the test folder, load the asphalt pavement texture parameter detection model pre-trained in the second step to perform image classification processing, and output the predicted probability of each type of label according to the preset label type. The label with the highest probability is the asphalt pavement texture parameter detection value.

Claims

1. A texture parameter detection method suitable for asphalt pavement, characterized in that: The following steps are involved: Step S1: Establishing an asphalt pavement texture image dataset; Step S2: establishing an asphalt pavement texture parameter detection model, and training the established asphalt pavement texture parameter detection model based on the asphalt pavement texture image dataset established in step S1; Step S3: Collect the asphalt pavement texture image to be detected, input it into the trained asphalt pavement texture parameter detection model, and output the pavement texture detection result.

2. A texture parameter detection method suitable for asphalt pavement according to claim 1, characterized in that: Step S1 specifically includes: using a camera to collect asphalt pavement texture images of different gradations and different wear stages, performing image segmentation processing on the collected asphalt pavement texture images, using a laser scanner to perform texture detection on the segmented images, recording texture parameters of the corresponding images, and setting different labels for different texture parameters.

3. A texture parameter detection method suitable for asphalt pavement according to claim 2, characterized in that: The method further comprises: in step S1, when using a camera to collect asphalt pavement texture images of different gradations and different wear stages, setting a shadowless lamp to perform brightness correction on the asphalt pavement area to be collected.

4. A texture parameter detection method for asphalt pavement according to claim 1, characterized in that: The step S2 of establishing the asphalt pavement texture parameter detection model specifically includes: Step S21: Use Python language to call the PyTorch tool library in the PyCharm interpreter; Step S22: Create an ImageFolder dataset of PyTorch, which contains images in the asphalt texture image dataset; Step S23: define helper function; Step S24: Create a base class of the PyTorch model named FingerModelBase, where the FingerModelBase base class includes the nn.Module class in PyTorch; Step S25: Setting the asphalt pavement texture parameter detection model structure; Step S26: setting an optimizer, and using the optimizer to optimize the asphalt pavement texture parameter detection model; Step S27: Based on the asphalt pavement texture image dataset established in step S1, the established asphalt pavement texture parameter detection model is trained, and the training results are evaluated by confusion matrix and TOP-K accuracy. If the evaluation results meet the settings, the asphalt pavement texture parameter detection model training is completed.

5. A texture parameter detection method for asphalt pavement according to claim 4, characterized in that: Step S3 includes: pre-processing the asphalt pavement texture image to be detected and inputting the trained asphalt pavement texture parameter detection model, outputting the prediction probability of various labels, and the highest probability label is the asphalt pavement texture parameter detection result value.

Images