Asphalt core sample analysis and detection method based on AI

Abstract

The invention relates to the technical field of AI, and particularly discloses an AI-based asphalt core sample analysis and detection method, which comprises the following steps: acquiring a side surface image of an asphalt core sample, and introducing a three-dimensional reconstruction technology and an image splicing synthesis technology to carry out three-dimensional reconstruction on the side surface image to obtain a core sample three-dimensional model; classifying each component in the asphalt core sample based on a semantic segmentation deep full convolutional neural network structure and the core sample three-dimensional model; and calculating the volume ratio of each component, and outputting the mix proportion in the asphalt core sample based on a pre-established mix proportion inversion model and the volume ratio. The method improves the efficiency of analyzing the mix proportion of the asphalt mixture.

Description

Technical Field

[0001] This invention relates to the field of AI technology, and more specifically to an AI-based method for analyzing and detecting asphalt core samples. Background Technology

[0002] Asphalt pavement is a seamless, continuous structural form with sufficient mechanical strength and excellent load-bearing capacity. It offers a smooth driving experience, low noise, no dust, and facilitates mechanized construction and subsequent maintenance, thus playing a vital role in modern high-grade highway construction.

[0003] Asphalt mixtures are made by mixing and compacting asphalt binder, mineral powder, and aggregates of different gradations in a specific ratio under certain temperature conditions. Due to the complexity of the material composition and structure, these mixtures exhibit heterogeneity, anisotropy, nonlinearity, and significant variability. Their mix proportions and key indicators such as porosity directly affect road performance characteristics such as high and low temperature performance, water stability, and fatigue life. Studies have shown that, under the premise of ensuring sufficient coating of the aggregate surface with structural asphalt and meeting temperature requirements for construction, optimizing the mix proportion can minimize porosity, achieve the densest aggregate packing, and form high structural strength at the aggregate-asphalt interface, thereby significantly improving road performance.

[0004] The main methods for testing asphalt mixtures include X-ray diffraction, centrifugal extraction, and combustion. Each method has its advantages and disadvantages. X-ray diffraction is simple to operate and quick, but the results are greatly affected by factors such as the source of asphalt and the type of aggregate. Centrifugal extraction can obtain gradation information, but the trichloroethylene used in this method is volatile and toxic, and some mineral powder will be lost into the mixture or adhere to the filter paper, which can easily cause test errors. Combustion is fast, highly automated, has fewer test steps, and is highly accurate, but the testing cost is high, the power consumption is large, the testing error is large for asphalt mixtures containing organic matter, and the error caused by aggregate loss on ignition cannot be ignored.

[0005] With the advancement of science and technology, people are gradually replacing traditional testing methods with non-destructive testing of asphalt core samples using techniques such as lasers, X-rays, and sound waves. However, cutting-edge research theories and methods face many bottlenecks in practical engineering applications, including complex processes, high levels of specialization, expensive equipment, and slow speed. Furthermore, some precise judgment methods are difficult to evaluate on-site. Therefore, a convenient, efficient, and feasible method for analyzing asphalt mixture proportions has become an urgent need in highway construction projects. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-based method for analyzing and detecting asphalt core samples, thereby solving the aforementioned technical problems.

[0007] The objective of this invention can be achieved through the following technical solutions: The AI-based method for analyzing and detecting asphalt core samples includes the following steps: Asphalt core samples were collected, and side surface images of the asphalt core samples were acquired using a line scan camera. Three-dimensional reconstruction technology and image stitching synthesis technology were introduced to perform three-dimensional reconstruction of the side surface images to obtain a three-dimensional model of the core sample. A deep fully convolutional neural network structure based on semantic segmentation and the three-dimensional model of the core sample are used to classify the components in the asphalt core sample, including aggregate, asphalt mortar and voids. Calculate the volume percentage of each component, and output the mix proportion in the asphalt core sample based on the pre-established mix proportion inversion model and the volume percentage. The mix proportion includes the amount of asphalt and the mineral powder content.

[0008] As a further aspect of the present invention: Three-dimensional reconstruction of the side surface image includes: The side surface of the asphalt core sample is scanned frame by frame using a line array camera with a precision that meets the preset requirements, ensuring that the entire circumferential direction is covered, thereby obtaining the image of the side surface. Feature points of two adjacent side surface images are extracted and matched. RANSAC is used to remove mismatches and the relative pose of each frame is estimated by combining rotation angle encoder information to achieve registration. The registration results are input into the feature point-based incremental SfM algorithm to generate a sparse point cloud through triangulation, and then Patch-MatchMVS is used to obtain a dense point cloud. After performing statistical outlier removal and Laplacian smoothing on the dense point cloud, a continuous mesh model is generated using Poisson, and the original side surface image is mapped back to the continuous mesh model via cylindrical projection to obtain high-resolution texture. The geometric-texture consistency is verified by reprojection error, point cloud density, and the root mean square error of overlapping areas, thus completing the three-dimensional reconstruction of the side surface image.

[0009] As a further aspect of the present invention: before extracting and matching feature points of two adjacent side surface images, the side surface images are further preprocessed, the preprocessing including denoising and enhancement.

[0010] As a further aspect of the present invention, it also includes: Aggregate voxels are classified according to their equivalent diameter, and the ratio of the number of aggregate voxels within a preset equivalent diameter range to the total number of aggregate voxels is calculated. The equivalent diameter is calculated based on the voxel connected domain. Using the equivalent diameter range as the horizontal axis and the ratio as the vertical axis, a gradation curve is generated for visualization.

[0011] As a further aspect of the present invention, it also includes: The three-dimensional model of the core sample is divided into several sub-regions along the axial and radial directions. The variance of the volume ratio of aggregate in each sub-region is calculated. The variance is used to measure the degree of segregation of the asphalt core sample. The larger the variance, the higher the degree of segregation. Texture contrast and uniformity of asphalt mortar areas are extracted using a gray-level co-occurrence matrix to evaluate the uniformity of the mixture.

[0012] As a further aspect of the present invention: the mix proportions output in the asphalt core sample include: A multiple regression model is established based on the volume percentage data to correlate the volume percentages of aggregate, asphalt mortar, and voids with the experimentally calibrated mix proportions, and the mix proportions in the asphalt core sample are output.

[0013] As a further aspect of the present invention: classifying the components in the asphalt core sample includes: A semantic segmentation network was used to distinguish aggregates from other components in the asphalt core sample. After identifying the aggregate, image binarization is performed, and deep learning image interpretation technology is used to identify voids and asphalt mortar.

[0014] The beneficial effects of this invention: Traditional asphalt mixture testing methods face many challenges: problems such as long testing time, low efficiency, and environmental pollution are becoming increasingly prominent, making it difficult to meet the current engineering needs for speed, accuracy, and environmental protection; In this invention: 1) By adopting advanced image acquisition and AI analysis technology, accurate mixing ratio information can be obtained without damaging the sample, which not only protects the valuable sample, but also leaves room for possible further testing. 2) By utilizing cutting-edge AI technologies such as deep learning, the detection results have a high confidence rate, providing reliable data support while greatly reducing human error, thus providing strong support for engineering quality control. 3) Compared with traditional methods, this invention avoids the use of toxic solvents and does not generate waste, which meets the environmental protection requirements of modern engineering. At the same time, its cost is much lower than that of high-end equipment such as X-rays, which can save enterprises a lot of testing costs. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart illustrating the AI-based asphalt core sample analysis and detection method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, this invention is an AI-based method for analyzing and detecting asphalt core samples, comprising the following steps: Asphalt core samples were collected, and side surface images of the asphalt core samples were acquired using a line scan camera. Three-dimensional reconstruction technology and image stitching synthesis technology were introduced to perform three-dimensional reconstruction of the side surface images to obtain a three-dimensional model of the core sample. In a preferred embodiment of the present invention, three-dimensional reconstruction of the side surface image includes: The side surface of the asphalt core sample is scanned frame by frame using a line array camera with a precision that meets the preset requirements, ensuring that the entire circumferential direction is covered, thereby obtaining the image of the side surface. Feature points of two adjacent side surface images are extracted and matched. RANSAC is used to remove mismatches and the relative pose of each frame is estimated by combining rotation angle encoder information to achieve registration. The registration results are input into the feature point-based incremental SfM algorithm to generate a sparse point cloud through triangulation, and then Patch-MatchMVS is used to obtain a dense point cloud. After performing statistical outlier removal and Laplacian smoothing on the dense point cloud, a continuous mesh model is generated using Poisson, and the original side surface image is mapped back to the continuous mesh model via cylindrical projection to obtain high-resolution texture. The geometric-texture consistency is verified by reprojection error, point cloud density, and mean square error of overlapping areas, thus completing the three-dimensional reconstruction of the side surface image; In the specific implementation process, the asphalt core sample is first fixed on a high-precision electrically controlled turntable with its geometric center as the axis. A linear array camera, in conjunction with a synchronously triggered ring LED light source, scans the sample frame by frame. An exposure is triggered every time the turntable rotates by a small angle. The linear array acquires narrow strip images in a line-by-line scanning manner, which are then unfolded into complete sidewall images during sample rotation. Subsequently, SIFT or ORB algorithms are used to extract scale- and rotation-invariant feature points from any two adjacent frames. Coarse matching is performed using mutual information or descriptor distance. Then, a RANSAC iterative model is used to eliminate mismatched point pairs. Simultaneously, the rotation encoder angle is read to map the feature pairs to the physical rotation coordinate system, and the relative pose between frames is solved using the least squares method to achieve precise registration. The continuous registration results are input into an incremental... The SfM (Sparse Point Cloud Model) process treats feature points as spatial anchors and generates a sparse point cloud through triangulation. Then, Patch-MatchMVS is used to compensate for dense disparity based on multi-view photometric consistency, generating a dense point cloud. The resulting point cloud undergoes statistical outlier removal to eliminate noise points, followed by Laplacian smoothing to suppress spikes. Next, the Poisson reconstruction algorithm is used to generate a closed mesh based on the point cloud normals, obtaining a topologically coherent continuous surface. Finally, the original side image is mapped to a cylindrical coordinate system, and cylindrical projection is performed based on pose information before being re-attached to the mesh to generate a high-resolution texture. Geometric and texture consistency is verified by calculating the reprojection residual, point cloud density, and the mean square error of grayscale in overlapping areas. This yields a high-quality core sample 3D model suitable for subsequent semantic analysis. It is important to note that using a linear array camera with synchronous rotation scanning allows for line-by-line acquisition of optical information while keeping the sample stationary relative to the focal plane. This avoids distortion caused by changes in the viewing angle of an area array camera and ensures consistent exposure. A ring light source reduces shadows on the aggregate surface, making feature extraction more stable. Matching images using scale-invariant features in conjunction with a random sampling consensus algorithm and removing erroneous points enhances the noise resistance of attitude estimation. Introducing an angle encoder to provide absolute angular displacement constraints limits error accumulation and eliminates closure uncertainty, resulting in unified and accurate point cloud coordinates. Incremental 3D reconstruction shortens the processing chain by acquiring and calculating simultaneously. Patch-matching dense reconstruction maintains fine texture while filling parallax holes, improving surface continuity. Statistical outlier removal and surface smoothing of the point cloud reduce noise and edge burrs. Global Poisson reconstruction uses normal information to ensure mesh closure, and cylindrical projection texture mapping achieves a one-to-one correspondence between geometric and optical data, providing high signal-to-noise ratio input for deep network recognition. Finally, multidimensional consistency checks promptly detect geometric fractures and illumination artifacts, enabling quality control during the reconstruction stage. To reduce subsequent semantic segmentation misjudgments and mix proportion inversion errors, the accuracy and reliability of core sample component classification, gradation curve generation, and segregation evaluation are improved overall. Consistency verification includes: first, reprojecting the vertices on the mesh model back to the original side surface images according to the camera extrinsic parameters; calculating the Euclidean distance between the projected points and their corresponding real feature points in the image coordinate system and averaging the values ​​across all cameras to generate reprojection residuals, which are used to verify the geometric solution accuracy; then, sampling the dense point cloud with a fixed voxel grid, counting the number of points contained in each voxel and generating a density histogram; if a local density is found to be significantly lower than the global median, it indicates that the geometric information in that area is insufficient and needs to be supplemented; next, mapping the overlapping textures of the same surface covered by two or more frames of images to the mesh and extracting the overlapping triangles; aligning the texture grayscale by pixels and calculating the root mean square error; if the root mean square error is large, it indicates that the illumination or registration error has caused texture inconsistency and requires re-exposure correction or pose optimization; once the three indicators meet the preset thresholds, the consistency between the geometric structure and texture mapping is considered to be up to standard, thus confirming that the model quality can support subsequent semantic segmentation and mix proportion inversion. In a preferred embodiment of this invention, before extracting and matching feature points of two adjacent side surface images, the side surface images are further preprocessed, including denoising and enhancement. It is understandable that Gaussian filtering is used to eliminate image noise, and histogram equalization is used to enhance contrast and highlight the boundary features of aggregates, asphalt mortar and voids. A deep fully convolutional neural network structure based on semantic segmentation and the three-dimensional model of the core sample are used to classify the components in the asphalt core sample, including aggregate, asphalt mortar and voids. In another preferred embodiment of the present invention, classifying the components in the asphalt core sample includes: A semantic segmentation network was used to distinguish aggregates from other components in the asphalt core sample. After identifying the aggregate, image binarization is performed, and deep learning image interpretation technology is used to identify voids and asphalt mortar. In the specific implementation, deep fully convolutional neural network structures such as U-Net or SegNet are adopted. These network structures are specifically designed for image segmentation tasks and can handle pixel-level classification problems well. Taking U-Net as an example, it achieves image segmentation through an encoder-decoder structure. The encoder part consists of a series of convolutional and pooling layers to extract image features, progressively reducing the spatial resolution of the image while increasing the number of channels to obtain richer semantic information. The decoder part uses upsampling and convolution operations to progressively restore the features extracted by the encoder to the resolution of the original image, thereby assigning a category label to each pixel; a large number of asphalt mixture images are collected as training data. These images need to cover different types of aggregates, asphalt mortar, and voids. Each image is labeled at the pixel level, that is, each pixel in the image is assigned a category label, for example, aggregate pixels are labeled as category 1, asphalt mortar pixels as category 2, void pixels as category 3, etc. The labeling work usually requires professional technicians to complete with labeling tools to ensure the accuracy of the labeling; the labeled image data is then input into the selected semantic segmentation network for training. During training, the network continuously adjusts its parameters through backpropagation to minimize the difference between the predicted results and the true labels. Commonly used loss functions include cross-entropy loss, which measures the similarity between the predicted and true class distributions. Through multiple iterations of training, the network gradually learns how to distinguish different components based on image features, ultimately obtaining a semantic segmentation model that can accurately classify aggregates and other components in asphalt mixtures. Since the collected asphalt mixture images may have different sizes and resolutions, cropping and scaling are necessary to adapt the images to the input requirements of the semantic segmentation network. The images are cropped to a suitable size and their resolution adjusted to ensure a consistent input format during training and prediction. The pixel values ​​of the images are normalized, adjusting their range to [0, 1] or [-1, 1]. Normalization can accelerate the convergence speed of the network and improve the model's generalization ability. Mean and standard deviation normalization methods are typically used, i.e., subtracting the mean of the image from each pixel value and then dividing by the standard deviation. The pre-processed asphalt mixture image is input into a trained semantic segmentation model for inference. The model outputs a segmentation result image of the same size as the input image, where the value of each pixel represents the category label to which that pixel belongs. By analyzing the segmentation result image, aggregates can be clearly distinguished from other components (such as asphalt mortar, voids, etc.). To further improve the accuracy and consistency of the segmentation results, post-processing can be performed on the model's output segmentation results. For example, morphological operations (such as dilation, erosion, etc.) can be used to eliminate small holes and noise in the segmentation results and smooth the segmentation boundaries; connected component analysis and other methods can also be used to merge or separate the segmented aggregate regions to ensure the integrity of the aggregate regions. Based on the output of the semantic segmentation model, pixel regions belonging to coarse aggregate are extracted. Since coarse aggregate is relatively large and easily identifiable, it can be filtered out by setting certain pixel or area thresholds. For example, connected components larger than a specific area can be identified as coarse aggregate regions. For the extracted coarse aggregate regions, the remaining parts (including asphalt mortar and voids) are considered background. To facilitate subsequent identification of voids and asphalt mortar components, image binarization is required. A suitable binarization algorithm (such as global thresholding or adaptive thresholding) is selected to divide the pixel values ​​in the image into two categories: foreground (coarse aggregate) and background (asphalt mortar and voids). In the binarized image, foreground pixel values ​​are 1, and background pixel values ​​are 0. Further analysis of the background portion in the binarized image is then performed using classification networks or object detection networks in deep learning. Since voids and asphalt mortar may exhibit different textures, colors, and shapes in images, a deep learning model can be trained to learn these feature differences, thereby achieving accurate identification of voids and asphalt mortar components. Samples, including image patches of voids and asphalt mortar, are extracted from the background region of the binarized image. Feature extraction is performed on these samples, with common methods including convolutional and pooling layer operations in convolutional neural networks (CNNs). These operations extract local and global features of the image. Then, the extracted features are input into a classification network (such as a support vector machine or a fully connected neural network) for training, resulting in a classification model capable of distinguishing between voids and asphalt mortar. The binarized image is then input into the trained classification model for inference. The model outputs the category (void or asphalt mortar) for each pixel or image patch in the image. By analyzing the model's output, the distribution of voids and asphalt mortar in the asphalt mixture can be obtained. In another preferred embodiment of the present invention, the volume percentage of each component is calculated, and the mix proportion in the asphalt core sample is output based on a pre-established mix proportion inversion model and the volume percentage. The mix proportion includes the amount of asphalt and the mineral powder content. The mix proportions output in the asphalt core sample include: A multiple regression model was established based on the volume percentage data to correlate the volume percentages of aggregate, asphalt mortar, and voids with the experimentally calibrated mix proportions, and the mix proportions in the asphalt core sample were output. The number of aggregate, asphalt mortar, and void voxels in the 3D voxel model is counted, and each voxel is multiplied by its actual volume to obtain the corresponding volume. Then, each volume is divided by the total volume of the core sample to obtain the volume percentage. Subsequently, data on the volume percentages of core samples with known mix proportions are retrieved from the experimental database. The samples are randomly divided into training and validation sets. A multiple regression model is constructed using sklearn or a similar machine learning framework, with the volume percentages of the three types of voxels as independent variables and asphalt content and mineral powder content as dependent variables. During the training phase, gradient descent is used to minimize the mean squared error, and the loss is monitored on the validation set to avoid overfitting. After the model converges, inputting the aggregate, mortar, and void volume percentage triplets into an unknown core sample will output the corresponding asphalt content and mineral powder content in real time. To increase the robustness of the model, cross-validation can be introduced, and the residuals are tested for normality. If heteroscedasticity or skewness is found, a Box-Cox transformation is performed on the volume percentages before regression. Finally, the inferred results are compared with experimental measurements to ensure the reliability of the output. Understandably, using volume percentage as a feature can directly reflect the actual material distribution inside the core sample without being affected by appearance color difference or light changes. The linear regression model parameters are highly interpretable, making it easy to quickly locate influencing factors when adjusting the formula later. Establishing a training library and improving the model's generalization ability through cross-validation can efficiently invert the amount of asphalt and mineral powder without destroying the specimen, providing an immediate reference for on-site testing, and thus helping the construction party to promptly correct the feeding ratio of the mixing plant to ensure the durability and economy of the road surface. Another preferred embodiment of the present invention further includes: Aggregate voxels are classified according to their equivalent diameter, and the ratio of the number of aggregate voxels within a preset equivalent diameter range to the total number of aggregate voxels is calculated. The equivalent diameter is calculated based on the voxel connected domain. Using the equivalent diameter range as the horizontal axis and the ratio as the vertical axis, a gradation curve is generated for visualization. It is important to note that after completing the three-dimensional segmentation of aggregates, asphalt mortar, and voids, the connected regions of the aggregates are marked in voxel space. The number of voxels contained in each connected region is counted and multiplied by the volume of a single voxel to obtain the volume of the connected region. Then, the volume is substituted into the formula for the volume of a sphere to calculate the equivalent diameter. The diameter value is matched with the particle size classification set in advance according to the engineering specifications. For example, fine, medium, and coarse particle ranges can be set. Each aggregate is assigned to the corresponding range according to its equivalent diameter and its voxel count is accumulated. Then, the number of voxels in each range is divided by the total number of voxels in the aggregate to obtain the proportion of that range. Subsequently, the drawing library is called to generate a line graph with the equivalent diameter range as the horizontal axis and the proportion as the vertical axis. The points are connected in sequence to form a gradation curve. The coordinate title and legend are automatically loaded, and one-click export of image files or embedding into the web interactive view is supported. Operators can switch to logarithmic coordinates or view the original statistical table in the interface to compare different core samples. The equivalent diameter classification based on voxel connected domains can make full use of the volume information obtained by 3D reconstruction and avoid the shape error of 2D projection measurement, making the particle size division more consistent with the true scale distribution of aggregates. The statistical voxel proportion by interval and the generation of gradation curves can intuitively reflect the particle size composition inside the core sample, helping engineers to quickly determine whether the aggregate gradation meets the design requirements and adjust the mix ratio accordingly. At the same time, it can provide reliable input data for subsequent segregation assessment and performance prediction, thus achieving efficient and visual detection of gradation quality on site without damaging the specimen. Another preferred embodiment of the present invention further includes: The three-dimensional model of the core sample is divided into several sub-regions along the axial and radial directions. The variance of the volume ratio of aggregate in each sub-region is calculated. The variance is used to measure the degree of segregation of the asphalt core sample. The larger the variance, the higher the degree of segregation. Texture contrast and uniformity of asphalt mortar areas are extracted using grayscale co-occurrence matrix to evaluate the uniformity of the mixture; It should be noted that a cylindrical coordinate system is established with the geometric axis of the core sample as the center. The model is cut into several horizontal layers along the axis, and then divided into concentric rings from the inside to the outside according to the radius. Each layer intersects with the ring to form an independent sub-region. Then, at the voxel level, the number of aggregate voxels in each sub-region is counted and multiplied by the volume of a single voxel to obtain the aggregate volume. Then, the aggregate volume is divided by the total volume of the sub-region to obtain the aggregate volume ratio. After that, the aggregate volume ratio of all sub-regions is put into the variance formula to calculate the overall variance to characterize the dispersion of aggregate distribution. At the same time, cross-sectional slices are selected from the classified asphalt mortar voxels and grayscale images are generated. A grayscale co-occurrence matrix is ​​constructed with a fixed step size and direction. The contrast and uniformity texture indices in the matrix are obtained to reflect the local grayscale changes and spatial distribution consistency of the asphalt mortar. This axial and radial bidirectional meshing method can decompose the originally monolithic core sample into fine-grained units, amplifying and quantifying the spatial non-uniformity of aggregate distribution. The larger the variance of aggregate volume ratio, the more obvious the enrichment and separation of coarse and fine materials at different locations, which helps to detect segregation risks early. The contrast of the gray-scale co-occurrence matrix can reveal the differences in brightness and darkness inside the asphalt mortar, and the uniformity can reflect the degree of texture repetition. These two indicators, combined with the variance evaluation results, can simultaneously examine the spatial coordination between aggregate and mortar, providing a reliable basis for judging the overall uniformity of the mixture, inferring construction quality, and predicting road performance.

[0019] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An AI-based method for analyzing and detecting asphalt core samples, characterized in that, Includes the following steps: Asphalt core samples were collected, and side surface images of the asphalt core samples were acquired using a line scan camera. Three-dimensional reconstruction technology and image stitching synthesis technology were introduced to perform three-dimensional reconstruction of the side surface images to obtain a three-dimensional model of the core sample. A deep fully convolutional neural network structure based on semantic segmentation and the three-dimensional model of the core sample are used to classify the components in the asphalt core sample, including aggregate, asphalt mortar and voids. Calculate the volume percentage of each component, and output the mix proportion in the asphalt core sample based on the pre-established mix proportion inversion model and the volume percentage. The mix proportion includes the amount of asphalt and the mineral powder content.

2. The AI-based asphalt core sample analysis and detection method according to claim 1, characterized in that, Three-dimensional reconstruction of the side surface image includes: The side surface of the asphalt core sample is scanned frame by frame using a line array camera with a precision that meets the preset requirements, ensuring that the entire circumferential direction is covered, thereby obtaining the image of the side surface. Feature points of two adjacent side surface images are extracted and matched. RANSAC is used to remove mismatches and the relative pose of each frame is estimated by combining rotation angle encoder information to achieve registration. The registration results are input into the feature point-based incremental SfM algorithm to generate a sparse point cloud through triangulation, and then Patch-MatchMVS is used to obtain a dense point cloud. After performing statistical outlier removal and Laplacian smoothing on the dense point cloud, a continuous mesh model is generated using Poisson, and the original side surface image is mapped back to the continuous mesh model via cylindrical projection to obtain high-resolution texture. The geometric-texture consistency is verified by reprojection error, point cloud density, and the root mean square error of overlapping areas, thus completing the three-dimensional reconstruction of the side surface image.

3. The AI-based asphalt core sample analysis and detection method according to claim 2, characterized in that, Before extracting and matching feature points of two adjacent side surface images, the side surface images are preprocessed, including denoising and enhancement.

4. The AI-based asphalt core sample analysis and detection method according to claim 1, characterized in that, Also includes: Aggregate voxels are classified according to their equivalent diameter, and the ratio of the number of aggregate voxels within a preset equivalent diameter range to the total number of aggregate voxels is calculated. The equivalent diameter is calculated based on the voxel connected domain. Using the equivalent diameter range as the horizontal axis and the ratio as the vertical axis, a gradation curve is generated for visualization.

5. The AI-based asphalt core sample analysis and detection method according to claim 1, characterized in that, Also includes: The three-dimensional model of the core sample is divided into several sub-regions along the axial and radial directions. The variance of the volume ratio of aggregate in each sub-region is calculated. The variance is used to measure the degree of segregation of the asphalt core sample. The larger the variance, the higher the degree of segregation. Texture contrast and uniformity of asphalt mortar areas are extracted using a gray-level co-occurrence matrix to evaluate the uniformity of the mixture.

6. The AI-based asphalt core sample analysis and detection method according to claim 1, characterized in that, The mix proportions output in the asphalt core sample include: A multiple regression model is established based on the volume percentage data to correlate the volume percentages of aggregate, asphalt mortar, and voids with the experimentally calibrated mix proportions, and the mix proportions in the asphalt core sample are output.

7. The AI-based asphalt core sample analysis and detection method according to claim 1, characterized in that, The classification of the components in the asphalt core sample includes: A semantic segmentation network was used to distinguish aggregates from other components in the asphalt core sample. After identifying the aggregate, image binarization is performed, and deep learning image interpretation technology is used to identify voids and asphalt mortar.

Images