An image detection and segmentation method for asphalt pavement aggregate gradation

Abstract

The application discloses an image detection and segmentation method for aggregate gradation of asphalt pavement, and comprises the following steps: step 1, aggregate image acquisition and data labeling; step 2, training and predicting aggregate contour data set based on U-Net network; and step 3, gradation generation, including determination of conversion parameters between image method and screening method and correction of aggregate error detected by image method, and the determination of conversion parameters between the image method and the screening method is specifically that the conversion parameters between the image method and the screening method are determined by three methods of aggregate projection contour area, contour perimeter and volume. The application completes the aggregate image segmentation task by using the U-Net deep learning network, establishes the corresponding relationship between the aggregate geometric information in the two-dimensional image and the screening quality in combination with the aggregate image features, and achieves the purpose of rapid and efficient detection. Different lithology and different particle size interval aggregates are taken as samples, a plurality of groups of screening method data and aggregate image processing data are compared, and the model training result is verified.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image detection, and in particular to an image detection and segmentation method for aggregate gradation of asphalt pavement. BACKGROUND

[0002] Nowadays, asphalt pavement is widely used, and aggregate gradation has a great influence on the road performance of asphalt pavement, so it is of great significance to detect aggregate gradation during aggregate transportation and before mixing. Whether in the laboratory or on the construction site, the highway engineering aggregate test specification obtains the aggregate gradation by the traditional screening method. In the laboratory, a small amount of aggregate is screened manually, which cannot meet the rapid and efficient detection requirements in the construction process.

[0003] With the mature development of computer image processing technology, more and more scholars use image recognition technology to replace the traditional screening experiment. Xue Yadong et al. use the Mask R-CNN instance segmentation network model to finely segment the tunnel leakage water disease image, and use the conditional convolution dynamic generation model parameter to replace the static parameter to improve the model segmentation speed and accuracy. Xu Wenjing uses the Yolov4 deep learning algorithm to detect and identify road cracks, so that the road crack recognition MAP value reaches 71.78%. Li Hongxiang et al. propose an ore image segmentation method based on U-Net network, which uses a generative adversarial network to segment the actual industrial production ore image, and the results show that the average relative error of the ore region is reduced to 8.20%. Hao Xueli et al. propose a bonded aggregate image segmentation model based on Inception network and residual connection optimization, which improves the accuracy by 30.46% compared with the traditional watershed segmentation algorithm. Based on the above-mentioned research of scholars, it can be seen that deep learning network has absolute advantages in image segmentation task. In this paper, the classical traditional image segmentation task algorithm is compared, and a mature deep learning network is selected to segment the particles in the aggregate image under different conditions.

[0004] However, there are still some problems in the process of aggregate image segmentation. For most of the agglomerated aggregate, effective segmentation cannot be achieved. Many scholars use high-speed cameras to shoot single images of falling aggregate to avoid aggregate and aggregate adhesion. However, this will result in less aggregate data in the image, increased image quantity, reduced calculation speed, and insufficient accuracy. SUMMARY

[0005] 1. Technical problem to be solved

[0006] The purpose of the present application is to solve the problem of less aggregate data in the image, increased image quantity, reduced calculation speed, and insufficient accuracy in the prior art, and to propose an image detection and segmentation method for aggregate gradation of asphalt pavement.

[0007] 2. Technical scheme

[0008] To achieve the above object, the present application adopts the following technical solutions:

[0009] An image detection and segmentation method for asphalt pavement aggregate gradation, comprising the following steps:

[0010] Step 1: aggregate image acquisition and data labeling, including data acquisition and data labeling and enhancement, data acquisition specifically includes collecting four kinds of aggregate image, respectively single distribution state, discrete distribution state, accumulation distribution state and vibration state;

[0011] Data labeling and enhancement specifically includes using an image acquisition device to collect a large amount of aggregate image data as a training data set for the network model;

[0012] Step 2: training and predicting aggregate contour data set based on U-Net network, including U-Net network training and model evaluation, U-Net network training specifically includes training the aggregate image data set labeled in the U-Net network model, using Python language to write U-Net network code, using neural network TensorFlow library and Keras library to build network structure, using image processing OpenCV library and Scikit-image library for image processing;

[0013] Model evaluation specifically includes that after U-Net network training is completed, a plurality of weight files will also be generated, and the weight file with the smallest loss value is used to predict the aggregate boundary during prediction;

[0014] Step 3: gradation generation, including determination of conversion parameters between image method and screening method and error correction of aggregate detection by image method, the determination of conversion parameters between image method and screening method specifically includes determining the conversion parameters between image method and screening method by aggregate projection contour area, contour perimeter and volume method, the calculation method is as follows:

[0015] S3.1 Aggregate projection contour area, contour perimeter is to calculate the cumulative geometric parameters of each particle size aggregate and the sum of the parameter values, and the equivalent screening method is used to calculate the percentage of the remaining screen;

[0016] S3.2 Volume method is that the same material source, the width-thickness ratio λ (the ratio of average thickness and width) of the aggregate particles is consistent; then the volume V of the aggregate particles is:

[0017] V=s×t=s×b×λ (1.1) In the formula:

[0018] s - aggregate particle projection area

[0019] t - average thickness of aggregate

[0020] b - width of aggregate particles;

[0021] then the percentage of each size fraction R i is:

[0022]

[0023] where:

[0024] p - density of aggregate

[0025] n i - number of aggregate particles in the i-th group

[0026] N - total number of aggregate particles

[0027] The test comparison shows that the projected area method has the smallest error as a conversion parameter between the image method and the sieve method. The error correction of the image method for aggregate detection is specifically to correct the error by the method of equation (1.3);

[0028] S i_amend = S i + S i × (P i_up + P i_down ) - S j × P j_down - S k × P k_up

[0029]

[0030] where:

[0031] S i_amend - corrected area result of the i-th size fraction

[0032] S i - total area of the i-th size fraction calculated by the image method before correction

[0033] p i_up - proportion of misidentification as the previous size fraction in the i-th size fraction

[0034] p i_down - proportion of misidentification as the next size fraction in the i-th size fraction

[0035] S j - total area of the j-th size fraction calculated by the image method before correction

[0036] p j_down - proportion of misidentification as the next size fraction in the j-th size fraction

[0037] S k - total area of the k-th size fraction calculated by the image method before correction

[0038] pk_up - the proportion of misidentification as the previous grade in the kth size fraction

[0039] i, j, k are adjacent sieve size fractions and the size relationship is j>i>k;

[0040] The batch aggregate is sampled by the above formula to perform multiple sieving and image methods, and the area results obtained by the image method are corrected according to the above formula, and the percentage of the classified oversize is recalculated.

[0041] Preferably, the discrete distribution state in step 1 is that different particle size aggregates are artificially placed and photographed, the stacking distribution state is that the aggregates are placed in a tray and gathered and then photographed, and the vibration state is that the aggregates are placed in the tray and then vibrated up and down at a fixed frequency to make the aggregates naturally placed without force as much as possible.

[0042] Preferably, the labeling work of the aggregate image needs to be completed before training in step 1, and the aggregate edge information that the deep learning network needs to learn is labeled.

[0043] Preferably, the aggregate image labeling software in step 1 is Labelme, version 3.16.2.

[0044] Preferably, the labeled samples are enhanced by image enhancement in step 1 to increase the amount of data, and the labeled aggregate images are processed by horizontal folding, vertical folding, random brightness enhancement and other enhancement methods using image enhancement software.

[0045] Preferably, the conversion parameter of the image method and the sieving method in step 3 is to compare the width of the minimum circumscribed rectangle of the aggregate particle projection with the size of the standard sieve screen; if it is greater than the size of the sieve, the geometric information or volume information of the aggregate particle is accumulated, and if it is less than the size of the sieve, it is compared with the size of the next grade sieve.

[0046] 3. Beneficial effects

[0047] Compared with the prior art, the advantages of the present application are:

[0048] In the present application, the U-Net deep learning network is used to complete the aggregate image segmentation task, the corresponding relationship between the aggregate geometric information in the two-dimensional image and the sieving quality is established by combining the aggregate image features, and the rapid and efficient detection purpose is achieved. Different lithology and different particle size range aggregates are used as samples, and the model training results are verified by comparing multiple groups of sieving data and aggregate image processing data. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A fixed light source aggregate image acquisition device according to the present application is shown in the figure;

[0050] Figure 2 The aggregate image enhancement schematic diagram proposed by the present application;

[0051] Figure 3 The U-Net model training result diagram proposed by the present application;

[0052] Figure 4 The U-Net network prediction effect diagram proposed by the present application;

[0053] Figure 5 The image method for calculating the aggregate particle size flowchart proposed by the present application;

[0054] Figure 6 The image method for screening process error generation schematic diagram proposed by the present application;

[0055] Figure 7 The image method and screening method gradation comparison curve diagram proposed by the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments.

[0057] Embodiment 1:

[0058] An image detection and segmentation method for asphalt pavement aggregate gradation, an image detection and segmentation method for asphalt pavement aggregate gradation, comprising the following steps:

[0059] Step 1: aggregate image acquisition and data labeling, including data acquisition and data labeling and enhancement, data acquisition specifically includes collecting four kinds of aggregate image, respectively single particle distribution state, discrete distribution state, accumulation distribution state and vibration state. Among them, the discrete distribution state is to artificially place different particle size aggregates for shooting, the accumulation distribution state is to place the aggregates on the tray and gather them again for shooting, and the vibration state is to place the aggregates on the tray and vibrate up and down at a fixed frequency, so that the aggregates are naturally placed without the action of force between the aggregates.

[0060] An aggregate image information acquisition device is designed to realize fixed light source image acquisition device for aggregate image information acquisition. The device can meet the rapid detection of indoor aggregate, and better simulate the real-time detection of aggregate at the site mixing station. The device is as follows Figure 1As shown, the device is composed of four 300x300mm 18W white LED light panels, a 300x300mm top light shield, and an AF12MP automatic focusing 1200MP industrial-grade camera. The four LED light panels and the enclosed shooting environment eliminate the influence of light intensity on the brightness of the aggregate image. Using a high-pixel, automatic focusing camera ensures the clarity of the aggregate image taken at this distance. Considering the dark color of the aggregate itself and the material of the bottom tray, to avoid the influence of background color and smooth tray reflection on the segmentation effect of the aggregate image, white card paper is used to wrap the aggregate tray. The camera USB interface is connected to a notebook computer, which is used to directly shoot and save the aggregate image.

[0061] Data annotation and enhancement involves using image acquisition devices to collect a large amount of aggregate image data as training data sets for the network model. Before training, the annotation of the aggregate image needs to be completed, and the edge information of the aggregate that the deep learning network needs to learn is labeled. The aggregate image annotation software is Labelme, version 3.16.2.

[0062] Since each aggregate image contains a large number of aggregate particle individuals, the annotation work requires a lot of time. To meet the requirement of the amount of data for deep learning network training, this paper enhances the labeled samples through image enhancement to increase the amount of data. The labeled aggregate images are processed using image enhancement software for horizontal folding, vertical folding, and random brightness enhancement. Image enhancement not only greatly reduces the workload of labeled images, but also through the enhanced aggregate images, which randomly increase noise information, can prevent overfitting to some extent during network training. Part of the results of image enhancement are shown in Figure 2 .

[0063] Step 2: Training and predicting aggregate contour data set based on U-Net network, including U-Net network training and model evaluation. U-Net network training is specifically using the labeled aggregate image data set to train in the U-Net network model, using Python language to write U-Net network code, using neural network TensorFlow library and Keras library to build network structure, using image processing OpenCV library and Scikit-image library for image processing.

[0064] The U-Net network model is a deep learning network modified and expanded on the basis of the fully convolutional neural network proposed by Olaf Ronneberger et al. in 2015. The network uses up-sampling layers to replace part of the operations of the pooling layer to increase the resolution of the output, so that the U-Net network has better recognition effect on details in the image. In the up-sampling layer, there are also a large number of feature channels, which allow the network to propagate feature information to higher resolution layers. Its unique "U" type structure well solves the transmission of context information and position information.

[0065] The U-Net network model is divided into three parts: the main feature extraction part, the enhanced feature extraction part, and the prediction part. The main feature extraction part is a stack of convolution and max pooling, similar to the feature extraction of the VGG16 network. Five preliminary effective feature layers are obtained through the main feature extraction part. The enhanced feature extraction part completes feature fusion after up-sampling the five preliminary effective feature layers, and the final effective feature layer is the feature information fused with all the feature layers. The prediction part uses the last fused effective feature layer to classify each pixel point.

[0066] Since the main feature extraction part of the U-Net network is similar in overall structure to the VGG16 network model, the VGG16 network is used instead of the U-Net network main extraction part in this paper. The VGG network is a convolutional neural network model proposed by Simonyan and Zisserman in the paper "Very Deep Convolutional Networks for Large Scale Image Recognition", and its name comes from the abbreviation of the Visual Geometry Group (VGG) of the University of Oxford where the authors are located. The VGG16 network structure is simple, composed of ordinary convolutional layers, max pooling and activation functions.

[0067] In the enhanced feature extraction network, the obtained 5 feature layers are repeatedly up-sampled and stacked to complete feature fusion. In order to increase the universality of the network in the aggregate image segmentation task, when up-sampling, no splicing operation is performed, but direct two times up-sampling is performed for stacking. This modification will make the input aggregate image size consistent with the output feature map size.

[0068] The model evaluation is specifically training in the U-Net network model using the labeled completed aggregate image dataset. The U-Net network code is written in Python language, and the network structure is built with the help of neural network TensorFlow library and Keras library, and image processing is carried out using image processing OpenCV library and Scikit-image library. Since the single particle aggregate image has a large number of background pixels (negative samples), it means that there are a lot of simple sample information in the aggregate image. In this way, the phenomenon of uneven positive and negative samples will appear. Uneven positive and negative samples will lead to the network not learning useful information, and the negative sample information will play a major role in the loss. In order to solve this problem, the Focal Loss loss function is used instead of the cross entropy loss function.

[0069] FL(p t )=-α t (1-p t ) γ log(p t ) (1.4)

[0070] In the formula:

[0071] p t —Predicted probability size

[0072] α t —Cross entropy weight, used to balance the proportion of positive and negative samples. α t Usually take 0.25

[0073] γ—used to reduce the loss of easy classification samples. 5>γ>0 usually take 2

[0074] The network training hyperparameter settings are shown in Table 1:

[0075] Table 1 U-Net network model hyperparameter settings in aggregate image segmentation

[0076]

[0077]

[0078] After 100 epochs, the model training loss function and MIoU value results are shown in Figure 3 . Mainly divided into training loss (train loss) and validation set loss (val loss). The model training is verified and evaluated once every 5 epochs on the validation set.

[0079] Through Figure 3It can be seen that when the model is trained to 60 generations, the model training loss and validation loss tend to be flat, and the U-Net network model also has good segmentation ability for aggregate edges, and there is no overfitting and underfitting phenomenon. MAccuracy, MPrecision, MIoU values reach good expected results.

[0080] Model evaluation is specifically after the U-Net network training is completed, and multiple weight files will also be generated. When predicting, use the weight file with the smallest loss value to predict the aggregate boundary, and the prediction result is as shown in Figure 4 By comparing the manually labeled mask image and the U-Net network prediction result, it can be seen that the U-Net network has good segmentation effect on aggregate images in different states, and the aggregate boundary is clear. For the case of multiple aggregate overlapping, the U-Net network can also well distinguish the boundary. Very few texture pixels are classified as boundary information.

[0081] Step 3: Gradation generation, including determination of conversion parameters between image method and screening method and error correction of image method detection aggregate, the conversion parameters between image method and screening method are determined by aggregate projection contour area, contour perimeter and volume method, and the calculation method is as follows:

[0082] S3.1 The aggregate projection contour area and contour perimeter are the cumulative geometric parameters of each particle size aggregate and the sum of the parameter values, which are equivalent to the percentage of the retained sieve in the screening method;

[0083] S3.2 The volume method is that the aggregate particles of the same material source have consistent width-thickness ratio λ (the ratio of average thickness to width); then the aggregate particle volume V is:

[0084] V=sxt=sxbxl (1.1) In the formula:

[0085] s—aggregate particle projection area

[0086] t—aggregate average thickness

[0087] b—width of aggregate particles

[0088] Then the percentage of each particle size R i is:

[0089]

[0090] In the formula:

[0091] ρ—aggregate density

[0092] n i —number of aggregate particles in the i group

[0093] N — total number of aggregate particles;

[0094] The conversion parameter of image method and sieving method is to compare the width of the minimum circumscribed rectangle of aggregate particle projection with the standard sieve mesh size. If it is greater than the mesh size, the geometric information or volume information of the aggregate particle is accumulated, and if it is less than the mesh size, it is compared with the next mesh size. The specific process of image method for determining aggregate particle size is shown in Figure 5 .

[0095] Three grades of limestone aggregate of 5-10 mm, 10-15 mm and 10-20 mm were used for sieving experiment, and the results of percentage of oversize obtained by various geometric parameter method and volume method were compared with the results of percentage of oversize obtained by sieving method to observe their error conditions. In terms of accuracy of identifying aggregate, the image method based on projection area conversion parameter has better performance than the image method based on projection perimeter parameter and volume geometric parameter. However, there is still some error in the identification rate, and the main reason is that from the method itself, the mass of three-dimensional aggregate particle is reflected by two-dimensional image geometric information, and there is bound to be error between them due to the lack of one dimension information, which is the inherent deficiency of two-dimensional image in gradation analysis. And this error is cumulative, the larger the particle size of aggregate, the greater the error.

[0096] Secondly, consider the sieving process of image method: as shown in Figure 6 , the aggregate surface is LW obtained by vibration method, and in this paper, the standard for whether a aggregate can pass through a certain size sieve is whether the longer side W of the circumscribed cube of the aggregate is less than the mesh size. However, in the actual sieving process, since the mesh size in the sieving experiment is the side length of the square hole sieve, whether a aggregate can pass through a certain size sieve is compared with the mesh in multiple directions. Figure 6 When W is greater than the side length of a certain sieve, it is identified as belonging to the aggregate on the sieve, but when the aggregate is rotated at a certain angle, it can actually pass through the sieve. Therefore, whether the aggregate can actually pass through a certain size sieve is determined by the short side and the longer side, but since the two-dimensional image cannot obtain the information of the third side and the shape of the aggregate particle is irregular, this error is inevitable.

[0097] The error correction of image method for detecting aggregate is to correct the error by formula (1.3);

[0098] S i_amend = S i +S i ×(P i_up +P i_down )-S j ×P j_down -S k ×P k_up

[0099]

[0100] In the formula:

[0101] S i_amend - the area result of the i-th size fraction after correction

[0102] S i - the area sum of the i-th size calculated by the image method before correction

[0103] p i_up - the proportion of misidentification as the previous size in the i-th size fraction

[0104] p i_down - the proportion of misidentification as the next size in the i-th size fraction

[0105] S j - the area sum of the j-th size calculated by the image method before correction

[0106] p j_down - the proportion of misidentification as the next size in the j-th size fraction

[0107] S k - the area sum of the k-th size calculated by the image method before correction

[0108] p k_up - the proportion of misidentification as the previous size in the k-th size fraction

[0109] i, j, k are adjacent sieve size fractions and the size relationship is j>i>k;

[0110] A plurality of parallel experiments are performed on each size fraction of the homologous aggregate using the above formula to solve the error. The correction coefficient is derived from the following table.

[0111] Table 2 Correction coefficient calculation process

[0112]

[0113] In the table, i, j, and k are the sieve sizes of the coarse aggregate of the asphalt mixture (j>i>k).

[0114] In the present application, through a plurality of verification experiments of error correction of different size fractions, it can be found that for the same aggregate of the same material source, the gradation calculated by the image method after correction according to formula (1.3) can better reproduce the gradation result of the sieving method. The percentage of sieve residue of each size fraction after correction and the actual difference is not more than 5%. It can be observed that Figure 7 It can be found that the gradation result after correction of the image method is highly coincident with the actual sieving gradation curve, further verifying that the correction parameters of the homologous aggregate are applicable to the mixture of the same material source.

[0115] In the application, the aggregate image segmentation task is completed by using a U-Net deep learning network, the corresponding relationship between the aggregate geometric information in the two-dimensional image and the screening quality is established in combination with the aggregate image features, and the fast and efficient detection purpose is achieved. With the aggregate of different lithology and different particle size interval as samples, the model training results are verified by comparing multiple sets of screening method data and aggregate image processing data.

[0116] The above merely describes a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for image detection and segmentation of aggregate gradation in asphalt pavement, characterized in that, Includes the following steps: Step 1: Aggregate image acquisition and data annotation, including data acquisition, data annotation and enhancement. Specifically, data acquisition involves acquiring images of four types of aggregates: single particle distribution, discrete distribution, piled distribution and post-vibration distribution. Data annotation and augmentation specifically involves using an image acquisition device to collect a large amount of aggregate image data as the training dataset for the network model; Step 2: Training and predicting aggregate contour dataset based on U-Net network, including U-Net network training and model evaluation. Specifically, U-Net network training is carried out in the U-Net network model using an annotated aggregate image dataset. U-Net network code is written in Python language, the network structure is built with the help of the TensorFlow and Keras libraries, and image processing is performed using the OpenCV and Scikit-image libraries. Specifically, after the U-Net network training is completed, multiple weight files will be generated. During prediction, the weight file with the smallest loss value will be used to predict the aggregate boundary. Step 3: Gradation generation, including determining the conversion parameters between the image method and the sieving method, and correcting the aggregate detection error in the image method. Specifically, the conversion parameters between the image method and the sieving method are determined using three methods: aggregate projected contour area, contour perimeter, and volume method. The calculation method is as follows: S3.1 The ratio of the projected contour area of ​​the aggregate to the sum of the areas of all aggregates is equivalent to the percentage of sieve residue in the screening method; the ratio of the projected contour perimeter of the aggregate to the sum of the perimeters of all aggregates is equivalent to the percentage of sieve residue in the screening method. S3.2 volumetric method uses the same material source and aggregate particle width-to-thickness ratio. If the ratio of average thickness to width remains consistent, then the aggregate particle volume V is: （1.1） In the formula: —Projected area of ​​aggregate particles —Average thickness of aggregate —Width of aggregate particles; Then there are percentages for each particle size. for: （1.2） In the formula: — Aggregate density ——No. Number of aggregate particles in the group —Total number of aggregate particles; Experimental comparisons show that the projected area method has the smallest error as a conversion parameter between the image method and the sieving method. The error correction for aggregate detection by the image method is specifically corrected by the method of formula (1.3). （1.3） In the formula: ——No. Area results after particle size correction; —The first image method calculated before correction The total area of ​​particles of the specified diameter; ——No. The percentage of items in the granularity level that are misidentified as the next higher level; ——No. The percentage of items at the [number] granularity level that are misidentified as the next lower level; —The first image method calculated before correction The total area of ​​particles of the specified diameter; ——No. The percentage of items at the [number] granularity level that are misidentified as the next lower level; —The first image method calculated before correction The total area of ​​particles of size 6 ——No. The percentage of items misidentified as the next higher level at the granularity level. , , The particle sizes are adjacent sieve openings and the particle size relationship is as follows: ; The same source aggregate was sampled using Equation (1.3) and subjected to multiple sieving and image methods. The projected area result obtained by the image method was corrected according to the above formula, and the percentage of residue in the grading sieve was recalculated.

2. The image detection and segmentation method for asphalt pavement aggregate gradation according to claim 1, characterized in that, In step 1, the discrete distribution state is achieved by artificially placing aggregates of different particle sizes for shooting; the stacked distribution state is achieved by placing the aggregates in a tray and gathering them together before shooting; and the vibrated state is achieved by placing the aggregates in a tray and vibrating them up and down at a fixed frequency to allow the aggregates to be naturally arranged between each other.

3. The image detection and segmentation method for asphalt pavement aggregate gradation according to claim 1, characterized in that, In step 1, the annotation of the aggregate image needs to be completed before training, and the aggregate edge information that the deep learning network needs to learn needs to be annotated.

4. The image detection and segmentation method for asphalt pavement aggregate gradation according to claim 1, characterized in that, The aggregate image annotation software used in step 1 is Labelme, version 3.16.

2.

5. The image detection and segmentation method for asphalt pavement aggregate gradation according to claim 1, characterized in that, In step 1, the labeled samples are enhanced by image enhancement to increase the amount of data. Image enhancement software is used to process the labeled aggregate images by horizontal folding, vertical folding, and random brightness enhancement.

6. The image detection and segmentation method for asphalt pavement aggregate gradation according to claim 1, characterized in that, In step 3, the conversion parameter between the image method and the sieving method is to compare the width of the minimum bounding rectangle of the aggregate particle projection with the standard sieve aperture size; if it is greater than the sieve aperture size, the geometric or volume information of the aggregate particle is accumulated; if it is less than the sieve aperture size, it is compared with the next sieve aperture size.

Images