Method for predicting particle roundness distribution of densely stacked asphalt mixture

Abstract

The invention relates to the technical field of crossing of road engineering and computer vision, and discloses a densely-stacked asphalt mixture particle roundness distribution prediction method, which comprises the following steps of: performing instance segmentation on particles in an image by adopting a visual basic model to obtain a mask image of each particle in the image; analyzing each particle mask image, extracting morphological parameters including the area, the equivalent circle diameter, the equivalent ellipse long axis and the like, constructing a particle geometric feature set, constructing a distribution curve for each morphological parameter according to the number of particles, and performing equal-interval sampling on the curve to form a multi-dimensional feature vector for machine learning input; and constructing a recurrent neural network model, and training the model by taking the multi-dimensional feature vector as an input and the known particle roundness distribution as a label to realize prediction of the particle roundness number distribution in the new image. And the segmentation applicability and the roundness distribution prediction precision of the densely-stacked asphalt mixture image are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road engineering and computer vision, in particular to a dense packing asphalt mixture particle roundness distribution prediction method. BACKGROUND

[0002] Asphalt mixture is an important material in road construction, and its performance is largely influenced by the particle shape characteristics, especially the roundness distribution of particles, which plays an important role in the mechanical properties and density of the mixture. Traditional particle roundness measurement methods mainly rely on manual screening and three-dimensional geometric modeling. These methods not only consume time and effort, but also have difficulty in dealing with the complex morphology of densely packed particles, resulting in limited accuracy and efficiency of the measurement results. With the development of computer vision technology, image analysis-based particle feature extraction methods have emerged, which can realize automatic identification and morphological parameter calculation of particles. However, existing methods are mostly aimed at relatively dispersed particle images and have difficulty in effectively processing the segmentation problem of densely packed particles. In addition, traditional image analysis methods often rely on empirical formulas or simple models, which cannot fully explore the complex nonlinear relationship between particle morphological parameters and roundness distribution, limiting the accuracy and scope of the prediction. Therefore, there is an urgent need for a new particle roundness distribution prediction method that combines advanced visual basic models and machine learning techniques to achieve efficient and accurate automatic analysis of densely packed asphalt mixture particles. SUMMARY

[0003] TECHNICAL PROBLEM SOLVED To solve the problems of the prior art, the present application provides a dense packing asphalt mixture particle roundness distribution prediction method, which has the advantages of significantly improving the segmentation applicability and roundness distribution prediction accuracy of dense packing asphalt mixture images, and has good automation processing potential and engineering application value.

[0004] (II) TECHNICAL SCHEME To achieve the above-mentioned purpose, the present application provides the following technical scheme: a dense packing asphalt mixture particle roundness distribution prediction method, comprising the following steps: S1, particle instance segmentation: using a SegmentAnythingModel (SAM) visual basic model to perform instance segmentation on the particles in the image, and obtaining a mask image of each particle in the image; S2, morphological parameter extraction: analyzing each particle mask image and extracting morphological parameters including area, equivalent circle diameter, and equivalent ellipse major axis, etc., to construct a particle geometric feature set; S3, feature vector construction: constructing a distribution curve for each morphological parameter according to the number of particles, and equally sampling the curve to form a multi-dimensional feature vector for machine learning input; S4, roundness distribution prediction: a recurrent neural network (RNN) model is constructed, a multi-dimensional feature vector is taken as input, a known particle roundness distribution is taken as label, the model is trained, and the prediction of the number distribution of particles in a new image is realized.

[0005] Preferably, the specific manner of performing instance segmentation of particles in the image by using the visual base model in step S1 includes the following steps: S1-1, obtaining an original image of dense packed asphalt mixture , , , respectively represent the height, width and channel number of the image. The image is input into the image encoder in the visual base model, and the high-dimensional semantic feature representation of the image is output , wherein is the feature dimension, , is the spatial size after downsampling; S1-2, based on the image edge gradient or prior morphological processing result, the potential particle boundary region is identified, and the spatial prompt information is constructed , the prompt information includes single-point prompt , multi-point prompt set , edge point set and point grid prompt structure , the above prompt information is input into the prompt decoder , and the corresponding prompt feature is generated; S1-3, the semantic feature output by the image encoder is fused with the prompt feature generated by the prompt decoder to construct a joint feature representation , wherein represents a channel dimension splicing or weighted fusion operation. The fused joint feature is input into the mask decoder , and the instance-level segmentation mask is generated, wherein represents the predicted number of particles.

[0006] Preferably, the specific manner of extracting morphological parameters from each particle mask image in step S2 includes the following steps: S2-1, extracting the boundary contour curve of each mask image , ensuring that is a closed boundary, which is used as the basic data for geometric quantity calculation; S2-2, count the number of pixels contained in each particle mask, convert to actual projection area, denoted as , used to represent the geometric projection range of the particle on the image plane; S2-3, construct the convex hull of the boundary point set of each particle, calculate the area of the smallest convex hull formed , used to measure the compactness of the particle shape and the complexity of the boundary; S2-4, count the length of the particle boundary curve , denoted as the perimeter , which can be used to derive the subsequent roundness or shape factor; S2-5, based on the particle area , calculate the equivalent circle radius equal to its area , that is: for unified scale particle comparison; S2-6, use the least squares method to fit the equivalent ellipse to the boundary curve , extract the length of the major axis of the ellipse and the length of the minor axis , used to describe the geometric size in the main direction of the particle and its shape anisotropy; S2-7, record the above morphological parameters according to the particle number, construct the particle geometric feature set: for subsequent feature vector construction and roundness distribution prediction tasks.

[0007] Preferably, the specific way of constructing a multi-dimensional input feature vector in step S3 includes the following steps: S3-1, sort each type of parameter in the particle geometric feature set extracted in step S2 independently, form an ascending order sequence in all particles in the image, each type of parameter includes area , convex hull area , perimeter , equivalent circle radius , ellipse major axis , ellipse minor axis , each type of parameter ascending order sequence is: S3-2, normalize each ranking sequence, and construct a normalized cumulative distribution function (CDF) curve based on the ranking position of each data point in the sequence in the whole. The curve describes the distribution trend of the parameter in all particles, which helps to reveal the global statistical characteristics; S3-3, set a fixed number of equidistant sampling positions in each cumulative distribution curve, extract the values at the corresponding positions from the curve to form a fixed-length numerical sequence. Taking the area parameter as an example, the extracted sample values are represented as: Similarly, the sampling sequences of other parameters can be obtained ; S3-4, concatenate the sampling numerical sequences corresponding to each parameter in order to form the final multi-dimensional input feature vector: wherein represents the vector concatenation operation, , is the total number of parameters, is the number of sampling points for each type of parameter. The input feature vector comprehensively represents the geometric morphological distribution characteristics of all particles in the image; S3-5, normalize the multi-dimensional input feature vector constructed to ensure that the numerical ranges of each dimension are consistent, enhance the numerical stability and convergence speed in the subsequent neural network model training process, and improve the prediction accuracy.

[0008] Preferably, the specific method of using a recurrent neural network (RNN) for roundness distribution prediction in step S4 includes the following steps: S4-1, construct an RNN model structure for roundness distribution prediction. The structure includes an input layer, several hidden recurrent layers (such as LSTM or GRU units), and an output layer. The network as a whole is used to model the nonlinear mapping relationship between morphological features and particle roundness statistical distribution. Let the input feature vector dimension be , and the output roundness distribution discrete sequence length be , then the model is represented as: wherein, is the input feature vector, is the predicted roundness distribution discrete value sequence, represents the model parameters; S4-2, the multi-dimensional input feature vector Input to the input layer of the RNN model. The RNN model regards the feature vector as a sequence feature input unfolded by time steps, and sends it to the recurrent layer for sequential processing; S4-3, by using the time sequence recursive structure of the RNN, the potential correlation and sequence distribution characteristics between the dimensions of the feature vector are modeled through the state transmission mechanism thereof. Specifically, the recurrent structure processes the feature vector at each time step Receives the input feature and combines the previous state Calculates the current state , which is formally expressed as follows: wherein represents the update function of the recurrent unit, represents the parameter set of the recurrent layer; S4-4, in the output layer, the state vector output by the recurrent layer is mapped to the prediction result of the roundness distribution. The output form is a fixed-length discrete value sequence: wherein each represents the proportion of particles in a certain roundness interval, which constitutes the number distribution curve of the roundness of the asphalt mixture particles.

[0009] S4-5, the RNN model is trained in a supervised learning manner.

[0010] Preferably, the image encoder adopts a Transformer structure with global context modeling capability.

[0011] Preferably, the instance-level segmentation mask is used to separate different particle regions in the image, providing a segmentation basis for subsequent extraction and distribution prediction of the roundness feature.

[0012] Preferably, the RNN model is trained using the real roundness distribution data of known image samples as labels, and the loss function is set to measure the difference between the predicted results and the real labels: Preferably, after the RNN model is trained, the loss function is minimized through the back propagation algorithm to update the model parameters , and finally the RNN model that can be used for roundness distribution prediction of new image samples is obtained.

[0013] Compared with the prior art, the present application provides a dense packing asphalt mixture particle roundness distribution prediction method, which has the following beneficial effects: The application first adopts a Segment Anything Model (SAM) visual basic model to perform instance segmentation on particles in a dense packed asphalt mixture image, to obtain a single particle mask image. Subsequently, through semantic feature engineering, morphological parameters such as area, equivalent circle diameter, and equivalent ellipse major axis are extracted from each particle mask as input features for predicting the roundness distribution. With a multi-dimensional feature vector as input and known roundness distribution data as labels, a recurrent neural network (RNN) model is constructed and trained to realize roundness distribution prediction. The application significantly improves the segmentation applicability and roundness distribution prediction accuracy of dense packed asphalt mixture images, and has good automation processing potential and engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The application implements an overall framework diagram of an asphalt mixture particle roundness distribution recognition algorithm based on SAM and machine learning; Figure 2 For example 1 of the application, an original image of dense packed asphalt mixture; Figure 3 For example 1 of the application, a point grid prompt image; Figure 4 For example 2 of the application, a diagram of quantitative features of particle contour extraction; Figure 5 For example 2 of the application, a cumulative distribution diagram of a particle morphological parameter set; Figure 6 For example 2 of the application, area morphological parameter cumulative distribution data sampling; Figure 7 For example 3 of the application, an RNN network structure diagram; Figure 8 For example 3 of the application, asphalt mixture aggregate particle roundness distribution prediction; Figure 9 For the method implementation steps of the application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0016] Please refer to Figures 1-9 A dense packed asphalt mixture particle roundness distribution prediction method, comprising the following steps: S1, particle instance segmentation: using a Segment Anything Model (SAM) visual base model to perform instance segmentation on the particles in the image to obtain a mask image of each particle in the image; S2, morphological parameter extraction: analyzing each particle mask image to extract morphological parameters including area, equivalent circle diameter, equivalent ellipse major axis, etc., to construct a particle geometric feature set; S3, feature vector construction: constructing a distribution curve for each morphological parameter according to the number of particles, and equally sampling the curve to form a multi-dimensional feature vector for machine learning input; S4, roundness distribution prediction: constructing a recurrent neural network (RNN) model and inputting the multi-dimensional feature vector, with the known particle roundness distribution as the label, to train the model and achieve prediction of the particle roundness distribution in a new image.

[0017] Example 1: (1) First, collect an RGB color image with a resolution of 1024x1024 as the original image of dense packed asphalt mixture , where , , represent the height and width of the input image. The image records the spatial distribution and boundary information of each particle in the dense packed asphalt mixture, as shown in Figure 2 .

[0018] (2) Input the image to the image encoder for feature extraction. The encoder structure uses a Vision Transformer (ViT), and its core calculation process is as follows: where represents dividing the input image into fixed size image blocks and mapping them into a vector sequence , and is the position encoding; the Transformer module models the self-attention of each image block feature and outputs a semantic feature map , where is the channel number, , is the patch size.

[0019] (3) Construct the point grid hint information , defined as: where , , represents the grid spacing, is the number of rows and columns of the grid, usually . The total number of grid points is , the point grid prompts .

[0020] The prompt information is input to the prompt decoder , and the prompt feature is generated: Each point in the prompt decoder is embedded and aligned with the attention of the corresponding position in the image space, as shown in Figure 3 .

[0021] (4) Fuse the semantic feature and the prompt feature to obtain the joint feature : Or use a weighted fusion method: The fused feature is input to the mask decoder for pixel-level instance segmentation, and the mask tensor is output: Wherein, is the predicted number of particles, is the mask of the th particle. Each channel in the mask is a normalized probability map representing the location distribution of the particle corresponding to the channel in the image.

[0022] Finally, the maximum channel response is used for mask discretization: The final pixel-level particle instance segmentation image is obtained, and the value of each pixel position represents the particle number to which it belongs, thereby completing the instance-level separation of each particle in the asphalt mixture as shown in Figure 4 .

[0023] Example 2: (1) Analyze the mask image of each particle in the instance segmentation result, and extract the particle geometric feature set parameters area, convex hull area, perimeter, equivalent circle diameter, equivalent ellipse major axis and minor axis as shown in Figure 5 .

[0024] Area : Count the number of all pixels in the particle mask. where represents the particle mask region.

[0025] Convex hull area : Calculate the convex polygon area of the minimum envelope particle boundary. Perimeter : Calculate the length of the particle boundary by edge detection. where represents the coordinates of the continuous pixel points on the boundary, is the number of boundary points.

[0026] Equivalent circle diameter : refers to the diameter of a circle with the same area, defined as: Equivalent ellipse major axis and minor axis : Calculate the length of the two major axes of the minimum fitting ellipse based on the principal axis of inertia of the particle.

[0027] (2) For each morphological parameter, extract all the corresponding numerical values of the particles at the image level, and sort them in ascending order to construct the distribution sequence of the parameter. For example, for the parameter , the sorted sequence is: (3) Normalize the above sorted sequence to the interval, and construct the cumulative distribution curve: where represents the normalized cumulative distribution function of the parameter , and the cumulative distribution is shown in Figure 6 .

[0028] (4) Set 10 equally spaced sampling points on the cumulative distribution curve, and take the corresponding function values as representative statistics. For the area parameter, for example Figure 7 is shown: For each parameter, a feature sub-vector with a length of 10 is extracted.

[0029] (5) The feature sub-vectors corresponding to the particle geometric feature set parameters area, convex hull area, perimeter, equivalent circle diameter, major axis, and minor axis are sequentially concatenated to construct a multi-dimensional input feature vector : (6) Normalization or standardization is performed on the final constructed feature vector, for example, using the Z-score standardization method: wherein and are the mean and standard deviation in the training set, respectively.

[0030] Example 3 (1) The RNN network structure is shown in FIG. 1, which is used to model the nonlinear mapping relationship between the particle morphological features and the roundness distribution. The structure includes an input layer, several hidden recurrent layers, and an output layer. The overall structure of the network can be represented as: Figure 8 wherein is the input feature sub-vector at time , is the hidden state, is the output, are the input weight matrix, hidden state transition matrix, and output weight matrix, respectively, is the bias term, is the activation function (such as tanh or ReLU).

[0031] (2) The input normalized multi-dimensional feature vector is input to the input layer of the RNN model as a time series input.

[0032] (3) Capture the correlation between the feature sequences. The RNN processes the input feature sequence through the recursive mechanism, explicitly models the time series correlation and overall distribution trend between the morphological parameters, and thus learns the statistical mapping relationship between them and the roundness distribution.

[0033] (4) Output the roundness distribution curve. The network outputs a discrete value sequence of fixed length at the output layer: wherein represents the value of the predicted roundness cumulative distribution at the th sampling point, is the set output sequence length, forming an approximation of the roundness distribution curve.

[0034] (5) Supervised training and model optimization. The known roundness distribution sequence is used as the label data to train the RNN model using supervised learning. The loss function can be in the form of mean square error (MSE):​​​​​ The network parameters are iteratively updated by the back propagation algorithm until the loss converges. The trained RNN model can be used to predict the statistical distribution curve of the roundness of new asphalt mixture image samples, as shown in Figure 9 .

[0035] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications, changes, omissions, substitutions and adaptations can be made by those skilled in the art without departing from the application, which is defined by the following claims and their equivalents.

Claims

1. A method for predicting the particle roundness distribution of densely packed asphalt mixtures, characterized in that, Includes the following steps: S1. Particle instance segmentation: The SegmentAnythingModel (SAM) visual basic model is used to segment particles in the image to obtain the mask image of each particle in the image. S2. Morphological parameter extraction: Analyze each particle mask image and extract morphological parameters including area, equivalent circle diameter, and equivalent ellipse major axis to construct a set of particle geometric features; S3. Feature Vector Construction: Construct distribution curves for each morphological parameter according to the number of particles, and sample the curves at equal intervals to form a multidimensional feature vector for machine learning input. S4. Roundness Distribution Prediction: Construct a recurrent neural network (RNN) model, and use the multidimensional feature vector as input and the known particle roundness distribution as label to train the model to predict the distribution of particle roundness in a new image.

2. The method for predicting the particle roundness distribution of densely packed asphalt mixtures according to claim 1, characterized in that, The specific method for performing instance segmentation of particles in the image using a visual basic model in step S1 includes the following steps: S1-1. Obtain the original image of the densely packed asphalt mixture. ,in , , These represent the image's height, width, and number of channels, respectively. The image is then input into the image encoder within the visual fundamental model. The high-dimensional semantic feature representation of the output image ,in For feature dimension, , This refers to the spatial dimensions after downsampling. S1-2. Based on image edge gradient or prior morphological processing results, identify potential particle boundary regions and construct spatial cue information. The prompt information Including single sign-on Multi-point prompt collection Edge point set and dot grid hint structure The above prompt information is input into the prompt decoder. In the process, corresponding prompt features are generated. ; S1-3, Semantic features output by the image encoder cue features generated by the cue decoder Perform fusion to construct a joint feature representation ,in This indicates a channel-level concatenation or weighted fusion operation. It combines the fused joint features. Input to mask decoder Generate instance-level segmentation mask ,in This indicates the predicted number of particles.

3. The method for predicting the particle roundness distribution of densely packed asphalt mixtures according to claim 2, characterized in that, The specific method for extracting morphological parameters for each particle mask image in step S2 includes the following steps: S2-1, For each mask image Extract its boundary contour curve ,make sure It is a closed boundary, used as the basic data for geometric quantity calculations; S2-2. Calculate the number of pixels contained in each particle mask, convert it into the actual projected area, and denot it as... , used to characterize the geometric projection range of particles on the image plane; S2-3. Construct the convex hull of the boundary point set for each particle, and calculate the area of ​​the minimum convex hull region formed. It is used to measure the compactness of particle shape and boundary complexity; S2-4, Statistical Particle Boundary Curve The length of the perimeter is denoted as the circumference. This parameter can be used to derive subsequent roundness or shape factor; S2-5, based on particle area Based on this, calculate the radius of the equivalent circle with the same area. ,Right now: Used for particle comparison at a uniform scale; S2-6. Use the least squares method to analyze the boundary curves. Perform an equivalent ellipse fitting and extract the length of the major axis of the ellipse. and minor axis length It is used to describe the geometric dimensions and shape anisotropy of particles along their principal direction; S2-7. Record the above morphological parameters according to particle number to construct a set of particle geometric features: Used for subsequent feature vector construction and roundness distribution prediction tasks.

4. The method for predicting the particle roundness distribution of densely packed asphalt mixtures according to claim 3, characterized in that, The specific method for constructing the multidimensional input feature vector in step S3 includes the following steps: S3-1. The set of particle geometric features extracted in step S2 Each type of parameter is independently sorted to form an ascending sequence of all particles in the image. Each type of parameter includes area. convex hull area ,perimeter Equivalent circle radius ellipse major axis minor axis of the ellipse The ascending sequence of each type of parameter is as follows: S3-2. Normalize each sorted sequence and construct a normalized cumulative distribution function (CDF) curve based on the sorted position of each data point in the sequence within the whole. This curve describes the distribution trend of this parameter across all particles, helping to reveal global statistical characteristics; S3-3. Set a fixed number of equally spaced sampling positions in each cumulative distribution curve, and extract the values ​​at the corresponding positions from the curve to form a fixed-length numerical sequence. Taking the area parameter as an example, the extracted sample values ​​are expressed as follows: Similarly, sampling sequences for other parameters can be obtained. ; S3-4. Concatenate the sampled numerical sequences corresponding to each parameter in order to form the final multidimensional input feature vector: in This represents a vector concatenation operation. , For the total number of parameters, This represents the number of sampling points for each parameter class. This input feature vector comprehensively characterizes the geometric distribution properties of all particles in the image. S3-5. Process the constructed multidimensional input feature vector Normalization is performed to ensure that the numerical range of each dimension is consistent, thereby enhancing the numerical stability and convergence speed during the subsequent training of the neural network model and improving prediction accuracy.

5. The method for predicting the particle roundness distribution of densely packed asphalt mixtures according to claim 4, characterized in that, The specific method for predicting the roundness distribution using a recurrent neural network (RNN) in step S4 includes the following steps: S4-1. Construct an RNN model structure for predicting particle roundness distribution. This structure includes an input layer, several hidden recurrent layers (such as LSTM or GRU units), and an output layer. The network as a whole models the nonlinear mapping relationship between morphological features and the statistical distribution of particle roundness. Let the dimension of the input feature vector be... The length of the output circularity distribution discrete sequence is The model is then represented as: in, For the input feature vector, For the predicted discrete value sequence of roundness distribution, Indicates model parameters; S4-2. The multidimensional input feature vector constructed and normalized in step S3 is... The feature vector is input to the input layer of the RNN model. The RNN model treats this feature vector as a sequence of features expanded over time steps and feeds it into the recurrent layer for sequential processing. S4-3. Utilize the time-series recursive structure of RNNs to model the potential correlations and sequence distribution characteristics between different dimensions of the feature vectors through their state propagation mechanism. Specifically, the recursive structure at each time step... Receive input features and in combination with the previous state Calculate the current state The formal expression is as follows: in This represents the update function for the loop unit. This represents the parameter set of the loop layer; S4-4. In the output layer, the state vector output by the recurrent layer is mapped to the predicted result of a rounded distribution. The output form is a fixed-length discrete value sequence: Each of them This represents the proportion of particles within a predicted range of roundness, and together they form the quantity distribution curve of asphalt mixture particle roundness. 6.S4-5. Train the RNN model using supervised learning.

7. The method for predicting the particle roundness distribution of densely packed asphalt mixtures according to claim 2, characterized in that: The image encoder adopts a Transformer structure with global context modeling capabilities.

8. The method for predicting the particle roundness distribution of densely packed asphalt mixtures according to claim 2, characterized in that: The instance-level segmentation mask It is used to separate different particle regions in an image, providing a segmentation basis for subsequent extraction and distribution prediction of roundness features.

9. The method for predicting the particle roundness distribution of densely packed asphalt mixtures according to claim 5, characterized in that: The RNN model is trained using the true roundness distribution data of known image samples. As a label, the mean squared error loss function is defined to measure the difference between the predicted result and the true label:

10. The method for predicting the particle roundness distribution of densely packed asphalt mixtures according to claim 5, characterized in that: After training, the RNN model minimizes the loss function using the backpropagation algorithm and updates the model parameters. Finally, an RNN model that can be used to predict the roundness distribution of new image samples is obtained.

Images