A digital image multi-component segmentation method for asphalt mixture of phase separation and synergy

Abstract

The present application provides a kind of digital image multi-component segmentation method of asphalt mixture of phase separation cooperation, it is related to the field of asphalt mixture image segmentation, including: obtaining the digital image of asphalt mixture section, conversion is carried out into gray image and is preprocessed to obtain the image to be segmented;The image to be segmented is carried out void segmentation to obtain void phase image mask;Using RAN-UNet deep learning image segmentation model, the image to be segmented is segmented to obtain aggregate image mask;Three-dimensional voxel reconstruction is carried out to obtain three-dimensional digital model, and connectivity analysis and edge detection are carried out, to obtain connected void mask, open void mask and closed void mask;Critical particle size screening and structural void identification are carried out to obtain structural void mask and mortar internal void;The aggregate image mask is screened by particle size and the asphalt mortar area is extracted, and the asphalt mortar area is resliced and converted, to obtain mortar phase image.The present application improves the segmentation accuracy of asphalt mixture digital image.

Description

Technical Field

[0001] This invention relates to the field of asphalt mixture image segmentation technology, and in particular to a phase-coordinated multi-component segmentation method for digital images of asphalt mixtures. Background Technology

[0002] Asphalt mixtures are multiphase heterogeneous composite materials composed of coarse aggregates, fine aggregates, mineral powder, asphalt binder, and internal voids. Their microstructure significantly impacts the material's mechanical properties and durability. Accurately identifying and quantifying the geometric distribution and volume proportions of each phase in asphalt mixtures is fundamental and a prerequisite for conducting micromechanical simulations, gradation design evaluations, and void structure analysis. In recent years, X-ray computed tomography (CT) scanning technology, with its non-destructive and high-resolution advantages, has gradually become the primary means of acquiring information on the internal microstructure of asphalt mixtures. By identifying the grayscale differences of each component in CT scan images, the spatial distribution information of aggregates, voids, and mortar can be extracted without damaging the specimen.

[0003] However, component segmentation in CT images of asphalt mixtures faces several technical challenges. On the one hand, the grayscale difference between aggregates and mortar is influenced by multiple factors, including aggregate mineral composition, asphalt cement thickness, and CT scan parameters. The grayscale ranges of adjacent components overlap to varying degrees, making it difficult for simple global threshold segmentation methods to accurately distinguish aggregate boundaries. On the other hand, the definition of the mortar phase lacks a unified standard in existing studies, often confusing internal voids with structural voids, leading to systematic biases in the calculation of mortar phase volume proportions and related parameters. These problems limit the accuracy of digital imaging technology in the microscopic analysis of asphalt mixtures.

[0004] Chinese invention patent CN109211945B discloses a method for analyzing the void structure of asphalt mixtures based on discrete element method (DEM). This method uses CT scan images as a basis, employs MATLAB to perform thresholding on the images to extract the pixel coordinates of voids, then transforms the void structure into small spherical units in the DEM, and classifies voids into connected voids, semi-connected voids, and closed voids based on the position information of the spherical units. However, it relies solely on a single grayscale thresholding process to extract the void phase, failing to achieve complete and refined segmentation of the three-phase components of aggregate, mortar, and voids, thus making it difficult to meet the practical needs of full-component analysis of the microstructure of asphalt mixtures. Summary of the Invention

[0005] In view of this, the present invention provides a multi-component segmentation method for digital images of asphalt mixtures with phase synergy, to solve the technical problem in the prior art that it is impossible to achieve complete and refined extraction of the three-phase components when performing component segmentation on CT images of asphalt mixtures. This method achieves accurate segmentation of the three-phase components of voids, aggregates, and mortar, thereby improving the segmentation accuracy of digital images of asphalt mixtures.

[0006] The technical solution of this invention is implemented as follows:

[0007] On one hand, the present invention provides a multi-component segmentation method for digital images of asphalt mixtures with phase-coordinated processing, comprising:

[0008] S1. Obtain a digital image of the asphalt mixture cross-section, convert the digital image to a grayscale image, and preprocess the grayscale image to obtain the image to be segmented;

[0009] S2. Use the Otsu thresholding method to perform gap segmentation on the image to be segmented to obtain the gap phase image mask;

[0010] S3. Use the trained RAN-UNet deep learning image segmentation model to segment the image to be segmented and obtain the aggregate image mask;

[0011] S4. Obtain a grayscale image mask based on the grayscale image, and reconstruct a three-dimensional voxel model by combining the grayscale image mask with the void phase image and aggregate image to obtain a three-dimensional digital model containing the three phases of voids, aggregate, and mortar; perform connectivity analysis and edge detection on the three-dimensional digital model to obtain a connected void mask, an open void mask, and a closed void mask.

[0012] S5. Based on the connected void mask, open void mask and closed void mask, critical particle size screening and structural void identification are performed to obtain the structural void mask and the internal void of the mortar.

[0013] S6. Perform particle size screening on the aggregate image mask and extract the asphalt mortar region containing asphalt binder, mineral powder, fine aggregate and mortar internal voids through layer-by-layer Boolean subtraction operation. Re-slice the asphalt mortar region along the direction of the asphalt mixture cross-section and convert it into a binary mask format to obtain the mortar phase image.

[0014] Based on the above technical solutions, preferably, step S2 specifically includes:

[0015] Iterate through all gray values ​​in the image to be segmented, using the current gray value as a candidate threshold, and divide the image pixels into a foreground class representing the aggregate and mortar community and a background class representing voids. Calculate the inter-class variance between the two classes:

[0016]

[0017] in, This represents the inter-class variance between the foreground and background classes. This indicates the proportion of foreground pixels to the total number of pixels in the image. This represents the average grayscale value of the foreground pixels. This represents the global average grayscale value of the image. This indicates the proportion of background pixels to the total number of pixels in the image. This represents the average grayscale value of the background pixels;

[0018] inter-class variance The gray value corresponding to the maximum value in the middle is used as the optimal segmentation threshold. Based on the optimal segmentation threshold, the image to be segmented is binarized to obtain a foreground mask including white and black regions. The white region represents the aggregate and mortar community, and the black region represents the voids.

[0019] Set the white areas in the foreground mask to zero and the black areas to white to obtain the gap phase image mask.

[0020] Based on the above technical solutions, preferably, the trained RAN-UNet deep learning image segmentation model uses Nested-UNet as the backbone network. In the downsampling path of the backbone network, each convolutional module includes an initial convolutional layer, an improved convolutional layer, and a channel attention module.

[0021] The initial convolutional layer is the image to be segmented. The convolution operation, batch normalization, and ReLU activation function are performed sequentially to output the initial features;

[0022] The improved convolutional layer performs convolution operations and batch normalization on the initial features sequentially to generate features. ;

[0023] The channel attention module employs a squeezing and excitation structure to focus on features. The Perform global average pooling on the nth channel to obtain the nth channel description vector. Each component ,right The adaptive weights for each channel are calculated by sequentially passing the material through two fully connected layers and a sigmoid activation function. and will and Multiplying the corresponding channels yields the channel-weighted features. :

[0024]

[0025]

[0026]

[0027] in, Channel description vector The Middle The element, i.e., the th element The output value of global average pooling for each channel. Indicates the height of the feature map. Indicates the width of the feature map. The row index representing the spatial location of the feature map. Column index representing the spatial location of the feature map. To represent features No. The spatial location of each channel Pixel value at that location, This represents the adaptive weight vector for each channel. This represents the Sigmoid activation function. This represents a gate function consisting of two fully connected layers and an intermediate activation function. This represents the channel description vector obtained after global average pooling. This represents the learnable parameter matrix of the second fully connected layer. This represents the ReLU activation function following the first fully connected layer. This represents the learnable parameter matrix of the first fully connected layer. This indicates a channel scaling operation. Representation of features No. Complete feature map of each channel Indicates the first Each channel corresponds to an adaptive scalar weight. This represents the feature output after channel weighting;

[0028] Output the channel-weighted features With the image to be segmented After residual connection, the output is processed by the ReLU activation function to obtain the final output of the convolutional module. .

[0029] Based on the above technical solutions, preferably, when training the RAN-UNet deep learning image segmentation model, the weighted sum of the binary cross-entropy loss function and the Dice loss function is used as the total loss function, and the Adam optimizer is used to iteratively optimize the model parameters.

[0030] Based on the above technical solutions, preferably, step S4 specifically includes:

[0031] The grayscale image is binarized to obtain the overall mask of the grayscale image; the pixels in the asphalt mixture specimen body area are marked as 1, and the pixels in the scanned background area are marked as 0.

[0032] The overall mask, void phase image mask, aggregate image mask, and voxel size parameters corresponding to each slice are imported into the 3D visualization and analysis software. The 3D visualization and analysis software is used to construct a 3D digital model containing void phase, aggregate phase, and mortar phase. The voxel size parameters are determined by CT scanning equipment.

[0033] Connectivity analysis is performed on all voxels corresponding to the void phase in the 3D digital model. Void connected regions that have voxel connected paths to the outer boundary of the 3D digital model are marked as connected voids and connected void masks are generated. The remaining void connected regions that do not have voxel connected paths to the outer boundary are marked as non-connected voids.

[0034] Surface contact detection is performed on non-connected voids. Non-connected voids that are in contact with the outer surface of the asphalt mixture specimen are marked as open voids and open void masks are generated. Non-connected voids that are completely enclosed inside the asphalt mixture specimen body are marked as closed voids and closed void masks are generated.

[0035] Based on the above technical solutions, preferably, the connectivity analysis uses the 26-neighborhood connectivity criterion to mark the void voxels in the three-dimensional digital model as connected voids, and marks all void voxel sets that satisfy the two conditions of being mutually connected and having a connected path with the outer boundary of the three-dimensional digital model as connected voids, thereby generating a connected void mask.

[0036] Based on the above technical solutions, preferably, the surface contact detection employs a three-dimensional morphological method, specifically including:

[0037] Perform a three-dimensional morphological erosion operation on the overall mask to obtain the eroded mask. Subtract the eroded mask from the overall mask to obtain the boundary layer on the outer surface of the asphalt mixture specimen with a single voxel thickness.

[0038] Perform a logical AND operation on a voxel-by-voxel basis between the voxel data of the boundary layer on the outer surface of the asphalt mixture specimen and the voxel data of the non-connected voids:

[0039] If a non-connected void connected domain has at least one overlapping voxel with the boundary layer on the outer surface of the asphalt mixture specimen, then the void connected domain is determined to be in contact with the outer surface of the asphalt mixture specimen and is marked as an open void.

[0040] If no overlapping voxels exist, the void is determined to be completely enclosed within the asphalt mixture specimen and is marked as a closed void.

[0041] Based on the above technical solutions, preferably, step S5 specifically includes:

[0042] In the overall mask, the connected gap mask and the open gap mask are removed sequentially by Boolean subtraction to obtain the first intermediate mask which contains only closed gap and non-gap regions.

[0043] Based on the pixel physical dimensions of the asphalt mixture cross-section, the pixel area of ​​each aggregate particle's connected region in the aggregate image mask is converted into its actual physical area, and its equivalent diameter is calculated. Particles with an equivalent diameter greater than the critical diameter are then selected. Aggregate particles, generating coarse aggregate mask;

[0044] Traverse each closed void connected region in the first intermediate mask, perform three-dimensional morphological operations on the closed void connected region and the coarse aggregate particle connected region respectively, extract the individual voxel thick boundary layer of each, perform a logical "AND" operation on each voxel of the two boundary layers, and determine the category of each closed void according to the operation result, generate a structural void mask, and retain closed voids that do not belong to structural voids as mortar internal voids.

[0045] Based on the above technical solutions, preferably, the condition for determining the category of each closed gap is as follows:

[0046] If, after performing a logical AND operation between the single voxel thick boundary layer of a closed void connected domain and the single voxel thick boundary layer of a coarse aggregate mask, there is at least one overlapping voxel, then it is determined that the closed void boundary is in contact with the coarse aggregate boundary in three-dimensional space, and it is marked as a structural void and included in the structural void mask.

[0047] If no overlapping voxels are found in the result of the logical AND operation, the closed void is determined to be completely inside the asphalt mortar area, marked as an internal void of the mortar, and retained as a component of the asphalt mortar area.

[0048] Furthermore, step S6 specifically includes:

[0049] In the overall mask, the coarse aggregate mask is removed by Boolean subtraction to obtain a second intermediate mask that does not contain coarse aggregate. ;

[0050] Second intermediate mask In the middle, the structural void mask is further eliminated through Boolean subtraction to obtain a third intermediate mask that does not contain coarse aggregate and structural voids. The third intermediate mask The corresponding three-dimensional region is the asphalt mortar region that includes asphalt binder, mineral powder, fine aggregate and internal voids of mortar.

[0051] The asphalt mortar area is re-sliced ​​along the same direction as the asphalt mixture slice direction and at the same axial spacing as the slice spacing to obtain a two-dimensional mortar phase image for each layer.

[0052] The two-dimensional mortar phase image is converted into a binary mask format, with the mortar area assigned a value of 1 and the non-mortar area assigned a value of 0, to obtain the output mortar phase image.

[0053] The present invention has the following advantages over the prior art:

[0054] (1) Otsu threshold segmentation and RAN-UNet deep learning image segmentation model were used to extract void phase image mask and aggregate image mask. In the three-dimensional digital model, the voids were finely classified by connectivity analysis, surface contact detection and coarse aggregate contact judgment. Finally, the mortar phase was extracted by layer-by-layer Boolean subtraction operation, which realized the accurate segmentation of void, aggregate and mortar three-phase components and improved the segmentation accuracy of asphalt mixture digital image.

[0055] (2) Through three-dimensional voxel reconstruction and connectivity analysis, the voids are subdivided into connected voids, open voids and closed voids, and the structural voids surrounded by coarse aggregates are identified, making the void classification more refined.

[0056] (3) By eliminating coarse aggregate, structural voids and various voids through layer-by-layer Boolean subtraction, the asphalt mortar area containing asphalt binder, mineral powder, fine aggregate and internal voids of mortar is extracted, which breaks through the traditional method of simplifying the mortar area. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart of a multi-component segmentation method for digital images of asphalt mixtures with phase-coordination according to the present invention.

[0059] Figure 2 This is a schematic diagram of the computed tomography scan image and preprocessing effect of the present invention; wherein, (a) is the initial grayscale image, (b) is the image after contrast enhancement, and (c) is the image after noise reduction;

[0060] Figure 3 This is a schematic diagram of the gap effect obtained after Otsu threshold segmentation according to the present invention; wherein, (a) is the image to be segmented; and (b) is the gap phase image mask.

[0061] Figure 4 This is a schematic diagram of the RAN-UNet deep learning image processing model architecture of the present invention;

[0062] Figure 5 This is a schematic diagram of the aggregate segmentation effect obtained by post-processing the trained RAN-UNet model of the present invention; wherein, (a) is the image to be segmented, and (b) is the aggregate image mask;

[0063] Figure 6 A schematic diagram of manually annotating the aggregate outline and generating the aggregate mask for the present invention; wherein, (a) is the image to be segmented, (b) is the schematic diagram of the aggregate outline annotation, and (c) is the schematic diagram of the aggregate image mask;

[0064] Figure 7 This is a schematic diagram of the asphalt mortar extraction process of the present invention;

[0065] Figure 8 This is a schematic diagram illustrating the extraction effect of the asphalt-mixed mortar according to the present invention. Detailed Implementation

[0066] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0067] like Figure 1 As shown, this invention provides a multi-component segmentation method for digital images of asphalt mixtures with phase-coordination, comprising:

[0068] S1. Obtain a digital image of the asphalt mixture cross-section, convert the digital image to a grayscale image, and preprocess the grayscale image to obtain the image to be segmented; the preprocessing includes contrast enhancement and noise reduction.

[0069] like Figure 2 As shown, (a) is the initial grayscale image, (b) is the image after contrast enhancement, and (c) is the image after noise reduction. Specifically, step S1 includes:

[0070] Three-dimensional digital volume data of asphalt mixture specimens were obtained by computed tomography, and sliced ​​along the axial direction to obtain a series of two-dimensional digital image sequences.

[0071] Each image in a two-dimensional digital image sequence is converted into a grayscale image with a grayscale range of 0 to 255.

[0072] The grayscale image is preprocessed, including noise reduction and contrast enhancement, to obtain the image to be segmented.

[0073] Understandably, preprocessing is used to improve image quality, enhance the grayscale differences between aggregates, mortar, and voids, and lay the foundation for subsequent segmentation.

[0074] In this embodiment of the invention, the asphalt mixture specimens can be prepared according to standard methods such as Marshall compaction method and Superpave rotary compaction method, and the gradation type can be AC, SMA, cast-in-place asphalt mixture, etc., and the maximum nominal particle size can be selected from common specifications such as 13.2mm and 16mm.

[0075] In one embodiment of the present invention, a rotary compaction specimen of SMA-13 ​​asphalt mixture with a void ratio of 4% is selected. The specimen is a cylinder with a diameter of 100 mm and a height of 100 mm ± 5 mm.

[0076] In one embodiment of the present invention, the contrast enhancement adopts limited contrast adaptive histogram equalization, with parameters set to an image block size of 8×8 pixels and a contrast limit threshold of 3.0. By dividing the grayscale image into several sub-blocks and performing histogram equalization on each, and then eliminating the boundary effect between blocks with bilinear interpolation, the grayscale difference between each component is enhanced.

[0077] In one embodiment of the present invention, the noise reduction process uses median filtering with a filter window size of 3×3 pixels to remove salt-and-pepper noise generated during image acquisition and obtain the image to be segmented.

[0078] In one embodiment of the present invention, contrast enhancement can also be achieved by global histogram equalization, and noise reduction can also be achieved by Gaussian filtering. The appropriate method can be flexibly selected based on the actual application.

[0079] like Figure 3 As shown, (a) is the image to be segmented, and (b) is the gap phase image mask. S2. The Otsu thresholding method is used to perform gap segmentation on the image to be segmented to obtain the gap phase image mask;

[0080] Specifically, step S2 includes:

[0081] Iterate through all gray values ​​in the image to be segmented, using the current gray value as a candidate threshold, and divide the image pixels into a foreground class representing the aggregate and mortar community and a background class representing voids. Calculate the inter-class variance between the two classes:

[0082]

[0083] in, This represents the inter-class variance between the foreground and background classes. This indicates the proportion of foreground pixels to the total number of pixels in the image. This represents the average grayscale value of the foreground pixels. This represents the global average grayscale value of the image. This indicates the proportion of background pixels to the total number of pixels in the image. This represents the average grayscale value of the background pixels;

[0084] inter-class variance The gray value corresponding to the maximum value in the middle is used as the optimal segmentation threshold. Based on the optimal segmentation threshold, the image to be segmented is binarized to obtain a foreground mask including white and black regions. The white region represents the aggregate and mortar community, and the black region represents the voids.

[0085] Set the white areas in the foreground mask to zero and the black areas to white to obtain the gap phase image mask.

[0086] Understandably, voids appear as the areas with the lowest gray values ​​in grayscale images, with a significant difference in gray values ​​from aggregates and mortar, making them suitable for segmentation using a global thresholding method. However, aggregates and mortars have similar gray values, making it difficult to accurately distinguish them using a single global threshold. Therefore, step S2 only extracts the void phase and does not distinguish between aggregates and mortars at this stage.

[0087] S3. Use the trained RAN-UNet deep learning image segmentation model to segment the image to be segmented and obtain the aggregate image mask;

[0088] like Figure 4 and Figure 5 As shown, where, Figure 5 In the image, (a) is the image to be segmented, and (b) is the aggregate image mask. Specifically, the trained RAN-UNet deep learning image segmentation model uses Nested-UNet as the backbone network. In the downsampling path of the backbone network, each convolutional module includes an initial convolutional layer, an improved convolutional layer, and a channel attention module.

[0089] The initial convolutional layer is the image to be segmented. The convolution operation, batch normalization, and ReLU activation function are performed sequentially to output the initial features;

[0090] The improved convolutional layer performs convolution operations and batch normalization on the initial features sequentially to generate features. ;

[0091] The channel attention module employs a squeezing and excitation structure to focus on features. The Perform global average pooling on the nth channel to obtain the nth channel description vector. Each component ,right The adaptive weights for each channel are calculated by sequentially passing the material through two fully connected layers and a sigmoid activation function. and will and Multiplying the corresponding channels yields the channel-weighted features. :

[0092]

[0093]

[0094]

[0095] in, Channel description vector The Middle The element, i.e., the th element The output value of global average pooling for each channel. Indicates the height of the feature map. Indicates the width of the feature map. The row index representing the spatial location of the feature map. Column index representing the spatial location of the feature map. To represent features No. The spatial location of each channel Pixel value at that location, This represents the adaptive weight vector for each channel. This represents the Sigmoid activation function. This represents a gate function consisting of two fully connected layers and an intermediate activation function. This represents the channel description vector obtained after global average pooling. This represents the learnable parameter matrix of the second fully connected layer. This represents the ReLU activation function following the first fully connected layer. This represents the learnable parameter matrix of the first fully connected layer. This indicates a channel scaling operation. Representation of features No. A complete feature map of each channel, Indicates the first Each channel corresponds to an adaptive scalar weight. This represents the feature output after channel weighting;

[0096] Output the channel-weighted features With the image to be segmented After residual connection, the output is processed by the ReLU activation function to obtain the final output of the convolutional module. .

[0097] Understandably, aggregates have complex shapes, irregular edges, and similar gray levels to mortar, making them difficult to segment accurately using traditional methods. Therefore, deep learning methods are employed.

[0098] The backbone network of the RAN-UNet deep learning image segmentation model in this invention is Nested-Unet. The input image has a spatial size of 512×512 and 1 channel (grayscale image). It includes four downsampling operations, each of which halves the spatial resolution and increases the number of channels accordingly, forming the following five resolution levels: the first layer maintains 512×512 with 64 channels; the second layer decreases to 256×256 with 128 channels; the third layer decreases to 128×128 with 256 channels; the fourth layer decreases to 64×64 with 512 channels; and the fifth layer decreases to 32×32 with 1024 channels. Each downsampling is implemented using a 2×2 max pooling layer. The ordinary convolutional module performs standard 3×3 convolution, batch normalization, and ReLU activation function processing; the improved convolutional module additionally introduces a channel attention mechanism and residual connections. At the deepest 32×32 resolution layer of the network, the input feature map has 1024 channels. After convolution processing, semantic information is passed upwards level by level. At each decoding node, it receives not only upsampled features from the previous decoding layer, but also original features from the same resolution layer of the encoder, as well as intermediate node features that have undergone pre-fusion processing in the same layer. All inputs are concatenated along the channel dimension before being fed into the convolutional module for processing. The decoding path forms four columns of intermediate fusion nodes, corresponding to the outputs of the four layers L1 to L4. Each column of nodes outputs upwards to the corresponding mask generation end, namely Mask 1, Mask 2, Mask 3, and Mask 4. Each mask is a segmentation result image of aggregate phase. During the training phase, all four masks participate in the loss function L. seg The computation is performed to achieve deep supervision; during the inference stage, the last output mask 4 is usually taken as the final aggregate image mask output. The improved convolution module is as described above and will not be repeated here. All four output masks (masks 1 to 4) are connected to the total loss function L through the backpropagation path. seg L seg The loss function is a weighted sum of the binary cross-entropy loss function and the Dice loss function. The losses from the four-way masks are aggregated in an equally weighted or weighted manner and then uniformly participate in backpropagation. This allows shallow nodes of the network to directly receive gradients from the supervision signal, effectively alleviating the gradient vanishing problem in deep network training and accelerating model convergence. The network parameters are iteratively updated using the Adam optimizer until the optimal segmentation performance is achieved on the validation set. The corresponding weights are then saved as the final model for extracting the aggregate phase from all images to be segmented.

[0099] like Figure 6As shown, (a) is the image to be segmented, (b) is a schematic diagram of aggregate outline annotation, and (c) is a schematic diagram of aggregate image mask. In one embodiment of the present invention, several representative images are selected from the images to be segmented, and the edges of the aggregate in each image are manually outlined using the Labelme image annotation tool to generate a binary mask image containing only the aggregate region as a label. The original images are paired with the corresponding label images and divided into training set, validation set and test set in a ratio of 7:2:1. Data augmentation processing is performed on the training set images, including random rotation, flipping, cropping and other operations, to expand the training set to 20 to 60 times the original number to improve the generalization ability of the model. The RAN-UNet deep learning image segmentation model is constructed and trained. The trained RAN-UNet deep learning image segmentation model is applied to all images to be segmented, and the output is a binary mask image containing only the aggregate phase, i.e., the aggregate image mask.

[0100] In one embodiment of the present invention, when training the RAN-UNet deep learning image segmentation model, the weighted sum of the binary cross-entropy loss function and the Dice loss function is used as the total loss function, and the Adam optimizer is used to iteratively optimize the model parameters. The model performance is evaluated on the validation set and the optimal model weights are saved.

[0101] S4. Obtain a grayscale image mask based on the grayscale image, and reconstruct a three-dimensional voxel model by combining the grayscale image mask with the void phase image and aggregate image to obtain a three-dimensional digital model containing the three phases of voids, aggregate, and mortar; perform connectivity analysis and edge detection on the three-dimensional digital model to obtain a connected void mask, an open void mask, and a closed void mask.

[0102] In one embodiment of the present invention, Avizo 3D visualization and analysis software is used for 3D voxel reconstruction and void connectivity analysis.

[0103] Specifically, step S4 includes:

[0104] The grayscale images of the asphalt mixture specimens formed by CT scanning are binarized to obtain the overall mask of the grayscale image; the pixels in the asphalt mixture specimen body area are marked as 1, and the pixels in the scanned background area are marked as 0.

[0105] The overall mask, void phase image mask, aggregate image mask, and voxel size parameters corresponding to each slice are imported into the 3D visualization and analysis software. The 3D visualization and analysis software is used to construct a 3D digital model containing void phase, aggregate phase, and mortar phase. The voxel size parameters are determined by CT scanning equipment, including the physical spacing of pixels in the slice and the axial spacing of adjacent slices. Both ensure that the geometric dimensions after 3D reconstruction are consistent with the actual physical dimensions of the asphalt mixture specimen.

[0106] Connectivity analysis is performed on all voxels corresponding to the void phase in the 3D digital model. Void connected regions that have voxel connected paths to the outer boundary of the 3D digital model are marked as connected voids and connected void masks are generated. The remaining void connected regions that do not have voxel connected paths to the outer boundary are marked as non-connected voids.

[0107] Surface contact detection is performed on non-connected voids. Non-connected voids that are in contact with the outer surface of the asphalt mixture specimen are marked as open voids and open void masks are generated. Non-connected voids that are completely enclosed inside the asphalt mixture specimen body are marked as closed voids and closed void masks are generated.

[0108] This invention uses three-dimensional voxel reconstruction and connectivity analysis to subdivide voids into connected voids, open voids, and closed voids, and identifies structural voids surrounded by coarse aggregates, making void classification more precise.

[0109] In one embodiment of the present invention, the connectivity analysis uses the 26-neighborhood connectivity criterion to mark the void voxels in the three-dimensional digital model as connected components. All void voxels that satisfy the two conditions of being mutually connected and having a connected path to the outer boundary of the three-dimensional digital model are marked as connected voids, and a connected void mask is generated.

[0110] Understandably, for each void voxel, its connectivity encompasses all 26 adjacent voxels that share a plane, edge, and corner in three-dimensional space.

[0111] In one embodiment of the present invention, the surface contact detection employs a three-dimensional morphological method, specifically including:

[0112] Perform a three-dimensional morphological erosion operation on the overall mask to obtain the eroded mask. Subtract the eroded mask from the overall mask to obtain the boundary layer on the outer surface of the asphalt mixture specimen with a single voxel thickness.

[0113] Perform a logical AND operation on a voxel-by-voxel basis between the voxel data of the boundary layer on the outer surface of the asphalt mixture specimen and the voxel data of the non-connected voids:

[0114] If a non-connected void connected domain has at least one overlapping voxel with the boundary layer on the outer surface of the asphalt mixture specimen, then the void connected domain is determined to be in contact with the outer surface of the asphalt mixture specimen and is marked as an open void.

[0115] If no overlapping voxels exist, the void is determined to be completely enclosed within the asphalt mixture specimen and is marked as a closed void.

[0116] It is understandable that non-connected voids that are in contact with the outer surface of the asphalt mixture specimen, even if they do not constitute through channels in the overall connectivity analysis, are open voids that open on the surface of the specimen. They have different physical significance from closed voids in the analysis of material permeability and durability, and therefore need to be identified and marked separately.

[0117] S5. Based on the connected void mask, open void mask and closed void mask, critical particle size screening and structural void identification are performed to obtain the structural void mask and the internal void of the mortar.

[0118] Specifically, step S5 includes:

[0119] In the overall mask, the connected gap mask and the open gap mask are removed sequentially by Boolean subtraction to obtain the first intermediate mask which contains only closed gap and non-gap regions.

[0120] Based on the pixel physical dimensions of the asphalt mixture cross-section, the pixel area of ​​each aggregate particle's connected region in the aggregate image mask is converted into its actual physical area, and its equivalent diameter is calculated. Particles with an equivalent diameter greater than the critical diameter are then selected. Aggregate particles, generating coarse aggregate mask;

[0121] Traverse each closed void connected region in the first intermediate mask, perform three-dimensional morphological operations on the closed void connected region and the coarse aggregate particle connected region respectively, extract the individual voxel thick boundary layer of each, perform a logical "AND" operation on each voxel of the two boundary layers, and determine the category of each closed void according to the operation result, generate a structural void mask, and retain closed voids that do not belong to structural voids as mortar internal voids.

[0122] In one embodiment of the present invention, the critical particle size The value of 2.36 mm is chosen because the technical specification for asphalt mixtures uses 2.36 mm as the dividing particle size between coarse and fine aggregates.

[0123] Specifically, the equivalent diameter is calculated as follows: based on the pixel physical size determined by the CT scanning equipment calibration, the pixel area of ​​the connected region of each aggregate particle in the aggregate image mask is converted into the actual physical area. (Unit: mm²) and in Calculate its equivalent circle diameter Filter out The aggregate particles generate a coarse aggregate mask.

[0124] Understandably, in asphalt mixtures, closed cavities formed by spatial interlocking between coarse aggregate particles or between coarse aggregate and other components are part of the aggregate skeleton structure and not internal voids of the mortar phase, and should not be included in the volume of the mortar phase. By using three-dimensional morphological boundary layer extraction and logical AND operations for judgment, the existence of physical contact between closed voids and coarse aggregate can be accurately identified in three-dimensional space, thereby achieving a refined distinction between structural voids and internal mortar voids. Compared with traditional two-dimensional edge detection methods, this invention avoids misjudgments of contact relationships caused by differences in slicing angles, and the judgment results have higher consistency in three-dimensional space.

[0125] Furthermore, the condition for determining the category of each closed gap is as follows:

[0126] If, after performing a logical AND operation between the single voxel thick boundary layer of a closed void connected domain and the single voxel thick boundary layer of a coarse aggregate mask, there is at least one overlapping voxel, then it is determined that the closed void boundary is in contact with the coarse aggregate boundary in three-dimensional space, and it is marked as a structural void and included in the structural void mask.

[0127] If no overlapping voxels are found in the result of the logical AND operation, the closed void is determined to be completely inside the asphalt mortar area, marked as an internal void of the mortar, and retained as a component of the asphalt mortar area.

[0128] S6. Perform particle size screening on the aggregate image mask and extract the asphalt mortar region containing asphalt binder, mineral powder, fine aggregate and mortar internal voids through layer-by-layer Boolean subtraction operation. Re-slice the asphalt mortar region along the direction of the asphalt mixture cross-section and convert it into a binary mask format to obtain the mortar phase image.

[0129] like Figure 7 and 8 As shown, specifically, step S6 includes:

[0130] In the overall mask, the coarse aggregate mask is removed by Boolean subtraction to obtain a second intermediate mask that does not contain coarse aggregate. ;

[0131] Second intermediate mask In the middle, the structural void mask is further eliminated through Boolean subtraction to obtain a third intermediate mask that does not contain coarse aggregate and structural voids. The third intermediate mask The corresponding three-dimensional region is the asphalt mortar region that includes asphalt binder, mineral powder, fine aggregate and internal voids of mortar.

[0132] The asphalt mortar area is re-sliced ​​along the same direction as the asphalt mixture slice direction and at the same axial spacing as the slice spacing to obtain a two-dimensional mortar phase image for each layer.

[0133] The two-dimensional mortar phase image is converted into a binary mask format, with the mortar area assigned a value of 1 and the non-mortar area assigned a value of 0, to obtain the output mortar phase image.

[0134] Understandably, the logic of the entire Boolean subtraction operation is as follows: the overall mask obtained in step S4 includes the three phases of coarse aggregate, mortar, and all voids; subtracting the coarse aggregate mask from the overall mask yields the second intermediate mask. ,at this time It includes the mortar phase, fine aggregate, internal voids in the mortar, and all structural voids; in Further subtracting the structural gap mask (including the connected gap mask, the open gap mask, and the structural gap mask in the closed gap) yields the third intermediate mask. This is the final asphalt mortar region. Physically, this region encompasses asphalt binder, mineral powder, fine aggregate with a particle size no greater than 2.36 mm, and the internal voids of the mortar completely enclosed within it. It eliminates interference from the coarse aggregate skeleton and various structural voids, accurately reflecting the spatial distribution range of the asphalt mortar phase. During reslicing, the slicing direction is consistent with the slicing direction of the asphalt mixture in step S1, and the axial spacing of the slices is the same as the original slice spacing in step S1. This ensures that the two-dimensional mortar phase image obtained from the reslicing corresponds one-to-one with the original asphalt mixture slice in spatial coordinates, facilitating subsequent overlay analysis with other component masks. The final output mortar phase image is a binary mask sequence, with mortar regions assigned a value of 1 and non-mortar regions (including coarse aggregate, connected voids, open voids, and structural voids) assigned a value of 0.

[0135] By eliminating coarse aggregate, structural voids, and various other voids through layer-by-layer Boolean subtraction, the asphalt mortar region containing asphalt binder, mineral powder, fine aggregate, and internal voids of the mortar is extracted, breaking through the simplification of mortar region processing in traditional methods.

[0136] This invention employs Otsu thresholding and the RAN-UNet deep learning image segmentation model to extract void phase and aggregate image masks. Within a 3D digital model, it achieves refined void classification through connectivity analysis, surface contact detection, and coarse aggregate contact assessment. Finally, it extracts the mortar phase using layer-by-layer Boolean subtraction, achieving precise segmentation of the void, aggregate, and mortar phases, thus improving the segmentation accuracy of asphalt mixture digital images. Furthermore, the segmentation results of this invention have clear physical meaning and can directly serve asphalt mixture permeability analysis, mechanical simulation, and performance evaluation. It has high engineering application value, strong repeatability, clearly defined parameter settings for each step, and a standardized data processing flow. It can be easily applied to asphalt mixture images of different gradations and sources, demonstrating good adaptability and promotional value.

[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-component segmentation method for digital images of asphalt mixtures with phase-coordination, characterized in that, include: S1. Obtain a digital image of the asphalt mixture cross-section, convert the digital image to a grayscale image, and preprocess the grayscale image to obtain the image to be segmented; S2. Use the Otsu thresholding method to perform gap segmentation on the image to be segmented to obtain the gap phase image mask; S3. Use the trained RAN-UNet deep learning image segmentation model to segment the image to be segmented and obtain the aggregate image mask; S4. Obtain the grayscale image mask based on the grayscale image, and reconstruct the three-dimensional voxel together with the void phase image mask and the aggregate image mask to obtain a three-dimensional digital model containing the three phases of voids, aggregates and mortar; perform connectivity analysis and edge detection on the three-dimensional digital model to obtain the connected void mask, the open void mask and the closed void mask; S5. Based on the connected void mask, open void mask and closed void mask, critical particle size screening and structural void identification are performed to obtain the structural void mask and the internal void of the mortar. S6. Perform particle size screening on the aggregate image mask and extract the asphalt mortar region containing asphalt binder, mineral powder, fine aggregate and mortar internal voids through layer-by-layer Boolean subtraction operation. Re-slice the asphalt mortar region along the direction of the asphalt mixture cross-section and convert it into a binary mask format to obtain the mortar phase image. Step S5 specifically includes: Binarize the grayscale image to obtain the overall mask of the grayscale image; In the overall mask, the connected gap mask and the open gap mask are removed sequentially by Boolean subtraction to obtain the first intermediate mask which contains only closed gap and non-gap regions. Based on the pixel physical dimensions of the asphalt mixture cross-section, the pixel area of ​​each aggregate particle's connected region in the aggregate image mask is converted into its actual physical area, and its equivalent diameter is calculated. Particles with an equivalent diameter greater than the critical diameter are then selected. Aggregate particles, generating coarse aggregate mask; Traverse each closed void connected region in the first intermediate mask, perform three-dimensional morphological operations on the closed void connected region and the coarse aggregate particle connected region respectively, extract the individual voxel thick boundary layer of each, perform a logical "AND" operation on each voxel of the two boundary layers, and determine the category of each closed void according to the operation result, generate a structural void mask, and retain closed voids that do not belong to structural voids as mortar internal voids.

2. The method for multi-component segmentation of digital images of asphalt mixtures with phase-coordination as described in claim 1, characterized in that: Step S2 specifically includes: Iterate through all gray values ​​in the image to be segmented, using the current gray value as a candidate threshold, and divide the image pixels into a foreground class representing the aggregate and mortar community and a background class representing voids. Calculate the inter-class variance between the two classes: ； in, This represents the inter-class variance between the foreground and background classes. This indicates the proportion of foreground pixels to the total number of pixels in the image. This represents the average grayscale value of the foreground pixels. This represents the global average grayscale value of the image. This indicates the proportion of background pixels to the total number of pixels in the image. This represents the average grayscale value of the background pixels; inter-class variance The gray value corresponding to the maximum value in the middle is used as the optimal segmentation threshold. Based on the optimal segmentation threshold, the image to be segmented is binarized to obtain a foreground mask including white and black regions. The white region represents the aggregate and mortar community, and the black region represents the voids. Set the white areas in the foreground mask to zero and the black areas to white to obtain the gap phase image mask.

3. The method for multi-component segmentation of digital images of asphalt mixtures with phase-coordination as described in claim 1, characterized in that: The trained RAN-UNet deep learning image segmentation model uses Nested-UNet as its backbone network. In the downsampling path of the backbone network, each convolutional module includes an initial convolutional layer, an improved convolutional layer, and a channel attention module. The initial convolutional layer is the image to be segmented. The convolution operation, batch normalization, and ReLU activation function are performed sequentially to output the initial features; The improved convolutional layer performs convolution operations and batch normalization on the initial features sequentially to generate features. ; The channel attention module employs a squeezing and excitation structure to focus on features. The Perform global average pooling on the nth channel to obtain the nth channel description vector. Each component ,right The adaptive weights for each channel are calculated by sequentially passing the material through two fully connected layers and a sigmoid activation function. and will and Multiplying the corresponding channels yields the channel-weighted features. : ； ； ； in, Channel description vector The Middle The element, i.e., the th element The output value of global average pooling for each channel. Indicates the height of the feature map. Indicates the width of the feature map. The row index representing the spatial location of the feature map. Column index representing the spatial location of the feature map. To represent features No. The spatial location of each channel Pixel value at that location, This represents the adaptive weight vector for each channel. This represents the Sigmoid activation function. This represents a gate function consisting of two fully connected layers and an intermediate activation function. This represents the channel description vector obtained after global average pooling. This represents the learnable parameter matrix of the second fully connected layer. This represents the ReLU activation function following the first fully connected layer. This represents the learnable parameter matrix of the first fully connected layer. This indicates a channel scaling operation. Representation of features No. A complete feature map of each channel, Indicates the first Each channel corresponds to an adaptive scalar weight. This represents the feature output after channel weighting; Output the channel-weighted features With the image to be segmented After residual connection, the output is processed by the ReLU activation function to obtain the final output of the convolutional module. .

4. The method for multi-component segmentation of digital images of asphalt mixtures with phase-coordination as described in claim 3, characterized in that: When training the RAN-UNet deep learning image segmentation model, the weighted sum of the binary cross-entropy loss function and the Dice loss function is used as the total loss function, and the Adam optimizer is used to iteratively optimize the model parameters.

5. The method for multi-component segmentation of digital images of asphalt mixtures with phase-coordination as described in claim 1, characterized in that: Step S4 specifically includes: The pixel marker for the asphalt mixture specimen body area is 1, and the pixel marker for the scanned background area is 0. The overall mask, void phase image mask, aggregate image mask, and voxel size parameters corresponding to each slice are imported into the 3D visualization and analysis software. The 3D visualization and analysis software is used to construct a 3D digital model containing void phase, aggregate phase, and mortar phase. The voxel size parameters are determined by CT scanning equipment. Connectivity analysis is performed on all voxels corresponding to the void phase in the 3D digital model. Void connected regions that have voxel connected paths to the outer boundary of the 3D digital model are marked as connected voids and connected void masks are generated. The remaining void connected regions that do not have voxel connected paths to the outer boundary are marked as non-connected voids. Surface contact detection is performed on non-connected voids. Non-connected voids that are in contact with the outer surface of the asphalt mixture specimen are marked as open voids and open void masks are generated. Non-connected voids that are completely enclosed inside the asphalt mixture specimen body are marked as closed voids and closed void masks are generated.

6. The method for multi-component segmentation of digital images of asphalt mixtures with phase-coordination as described in claim 5, characterized in that: The connectivity analysis uses the 26-neighborhood connectivity criterion to label the void voxels in the 3D digital model as connected components. All void voxels that satisfy the two conditions of being mutually connected and having a connected path to the outer boundary of the 3D digital model are labeled as connected voids, generating a connected void mask.

7. The method for multi-component segmentation of digital images of asphalt mixtures with phase-coordination as described in claim 6, characterized in that: The surface contact detection employs a three-dimensional morphological method, specifically including: Perform a three-dimensional morphological erosion operation on the overall mask to obtain the eroded mask. Subtract the eroded mask from the overall mask to obtain the boundary layer on the outer surface of the asphalt mixture specimen with a single voxel thickness. Perform a logical AND operation on a voxel-by-voxel basis between the voxel data of the boundary layer on the outer surface of the asphalt mixture specimen and the voxel data of the non-connected voids: If a non-connected void connected domain has at least one overlapping voxel with the boundary layer on the outer surface of the asphalt mixture specimen, then the void connected domain is determined to be in contact with the outer surface of the asphalt mixture specimen and is marked as an open void. If no overlapping voxels exist, the void is determined to be completely enclosed within the asphalt mixture specimen and is marked as a closed void.

8. The method for multi-component segmentation of digital images of asphalt mixtures with phase-coordination as described in claim 7, characterized in that: The criteria for determining the category of each closed gap are as follows: If, after performing a logical AND operation between the single voxel thick boundary layer of a closed void connected domain and the single voxel thick boundary layer of a coarse aggregate mask, there is at least one overlapping voxel, then it is determined that the closed void boundary is in contact with the coarse aggregate boundary in three-dimensional space, and it is marked as a structural void and included in the structural void mask. If no overlapping voxels are found in the result of the logical AND operation, the closed void is determined to be completely inside the asphalt mortar area, marked as an internal void of the mortar, and retained as a component of the asphalt mortar area.

9. The method for multi-component segmentation of digital images of asphalt mixtures with phase-coordination as described in claim 1, characterized in that: Step S6 specifically includes: In the overall mask, the coarse aggregate mask is removed by Boolean subtraction to obtain a second intermediate mask that does not contain coarse aggregate. ; Second intermediate mask In the middle, the structural void mask is further eliminated through Boolean subtraction to obtain a third intermediate mask that does not contain coarse aggregate and structural voids. The third intermediate mask The corresponding three-dimensional region is the asphalt mortar region that includes asphalt binder, mineral powder, fine aggregate and internal voids of mortar. The asphalt mortar area is re-sliced ​​along the same direction as the asphalt mixture slice direction and at the same axial spacing as the slice spacing to obtain a two-dimensional mortar phase image for each layer. The two-dimensional mortar phase image is converted into a binary mask format, with the mortar area assigned a value of 1 and the non-mortar area assigned a value of 0, to obtain the output mortar phase image.

Images