Asphalt concrete ct image segmentation prediction method based on semantic segmentation

Abstract

The application provides an asphalt concrete CT image segmentation prediction method based on semantic segmentation. The method establishes an asphalt concrete CT image segmentation offline database based on CT scanning and digital image processing technology, develops an AEU-net neural network model based on semantic segmentation, an attention mechanism and a supervised learning theory, and finally trains the AEU-net neural network model by using the offline database to obtain an AEU-net-DIP proxy model. The application fully utilizes the efficient mapping capability of the neural network model, accurately extracts aggregate phase morphological information of the asphalt concrete CT image, and effectively solves the problems of low efficiency, weak universality and error segmentation of the existing digital image processing method.

Description

Technical Field

[0001] This invention relates to the field of tomographic image processing. Background Technology

[0002] A CT image set is a sequential collection of multiple homogeneous two-dimensional CT images that characterize the internal geometry of an object. It is typically obtained by rotating and processing computed tomography (CT) images along the surface contour of the specimen. CT images of asphalt concrete contain explicit morphological information such as the geometric features, density distribution, and material boundaries of the heterogeneous materials within the specimen. After image segmentation for information purification, these images can be used for research in areas such as microstructural feature analysis, mechanical behavior analysis, and prediction of damage self-healing behavior.

[0003] The pixel information of asphalt concrete CT images is represented by 8-bit codes "0-255". Pixels with a value of "0" appear dark, pixels with a value of "255" appear bright white, and pixels with values ​​of "1-254" show a linear distribution of features according to this rule. Among them, dark areas formed by clusters of "0" value pixels represent voids, gray areas formed by multiple clusters of low to medium value pixels represent asphalt slurry, and bright white areas formed by multiple clusters of medium to high value pixels represent aggregates.

[0004] CT imaging is essentially imaging based on differences in material density, and physical limitations lead to inherent defects in CT images. Specifically, CT images exhibit an overall brightness characteristic of "low inner brightness and high perimeter brightness, with randomly distributed noise," and there is a large-scale union of pixel grayscale values ​​between the aggregate phase and the asphalt mortar phase. This significantly increases the difficulty of image processing.

[0005] Digital image processing (DIP) is a technique that uses the pixel information features of two-dimensional images to process and analyze them through the integration of multiple algorithms, thereby achieving effects such as image enhancement, feature extraction, and noise reduction. In the field of asphalt concrete CT image processing, the key objective is to extract the aggregate phase morphological information from CT images.

[0006] Existing segmentation methods for asphalt concrete CT images suffer from low efficiency, weak versatility, and erroneous segmentation. In the current technological system, to effectively extract aggregate phases from asphalt concrete CT images, it is necessary to analyze the inherent defect characteristics of the asphalt concrete CT images to be processed beforehand, and then design a suitable morphological processing algorithm. Secondly, differences in physical conditions such as specimen quality, environmental factors, and CT power lead to significant differences in grayscale distribution between CT images from different sources, and existing morphological processing algorithms cannot meet the requirements for processing asphalt concrete CT images from different sources. Finally, the essence of digital image processing is to process the entire image or artificially defined regions using fixed algorithms. However, existing asphalt concrete CT image segmentation methods have varying sensitivities to aggregates with different shapes and contour features, leading to the erroneous segmentation of some aggregate phases, specifically manifested as incomplete aggregates, missing aggregates, and long strips of aggregate being segmented into multiple linearly distributed small aggregates. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention proposes a semantic segmentation-based CT image segmentation prediction method for asphalt concrete specimens. This method establishes an offline database for asphalt concrete CT image segmentation based on CT scanning and digital image processing technologies. Then, it develops an AEU-net neural network model based on semantic segmentation, attention mechanisms, and supervised learning theory. Finally, it trains a U-net using the offline database to establish an AEUnet-DIP surrogate model. Using the AEUnet-DIP surrogate model, efficient and accurate mapping from the original CT image to the segmentation result can be achieved. This invention fully utilizes the efficient mapping capability of the neural network model to accurately extract the aggregate phase morphological information of asphalt concrete CT images, effectively solving the problems of low efficiency, weak versatility, and incorrect segmentation in existing digital image processing methods.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A semantic segmentation-based segmentation prediction method for asphalt concrete CT images includes the following steps:

[0010] Step 1: Initial CT image acquisition: Prepare asphalt concrete CT scan specimens, determine the scanning direction, and use CT scanning technology to acquire its initial CT tomographic image set;

[0011] Step 2: Establishment and expansion of the offline image segmentation database;

[0012] Step 3: Neural Network Model Optimization and Development: Based on the data structure characteristics of the offline image segmentation database, a U-net neural network model considering the attention mechanism, namely the AEUnet neural network model, is constructed based on semantic segmentation, attention mechanism, and supervised learning theory.

[0013] Step 4: Neural Network Model Training: The AEUnet neural network model is trained using the image segmentation offline database established in Step 2 to obtain the AEUnet-DIP proxy model;

[0014] Step 5: Image segmentation prediction: The AEUnet-DIP surrogate model is used to perform aggregate phase segmentation prediction on the asphalt concrete CT image to be tested.

[0015] Preferably, the CT images in step one include 8-bit RGB color images and grayscale images.

[0016] Preferably, each CT image in step one contains all pixel information of the complete specimen on the cross-section, and redundant pixel positions are filled with pixels with a value of "0".

[0017] Furthermore, step two includes:

[0018] Based on the grayscale distribution characteristics and inherent defects of the initial CT tomographic images, a DIP processing flow is designed to segment different components of the image.

[0019] Using the initial CT original image and the corresponding image segmentation result as a set of data pairs, a high-quality offline image segmentation database is established based on the initial CT tomographic image set.

[0020] By using CT scan computer simulation algorithms, new tomographic images in different directions are extracted from the asphalt concrete CT tomographic image set obtained in step one, thereby expanding the offline image segmentation database.

[0021] Furthermore, the image segmentation results in step two include aggregate phase segmentation results and asphalt mortar phase segmentation results.

[0022] Furthermore, the implementation method of establishing and expanding the offline image segmentation database in step two includes using DIP technology to improve the quality of the offline image segmentation database and using CT scan computer simulation algorithm process to expand the quantity of the offline image segmentation database.

[0023] Preferably, the DIP processing flow in step two includes steps such as image filtering, pixel filling, maximum threshold segmentation, watershed segmentation, and pixel restoration.

[0024] Preferably, the CT scan computer simulation algorithm in step two includes steps such as voxel reconstruction, model coordinate space rotation, and voxel model reslicing.

[0025] Furthermore, the AEUnet neural network model, based on the basic semantic segmentation network U-net, implements feature fusion optimization under the receptive field of each layer of the neural network based on the channel attention mechanism, so as to avoid interference of irrelevant features on the prediction results.

[0026] Preferably, the training process of the AEUnet neural network model can reserve a transfer learning interface, which allows for the superposition and training of existing AEUnet neural network model training results using CT image scan results from different sources, further improving the image segmentation prediction accuracy and generalizability of the AEUnet neural network model. When training with new data, it is necessary to ensure that the image-label specifications are consistent with the training data.

[0027] The present invention has the following beneficial effects:

[0028] (1) High segmentation efficiency: The image segmentation efficiency of the AEUnet-DIP proxy model after training is greatly improved, and the efficiency ratio is further improved as the image resolution increases.

[0029] (2) Enhanced versatility: The trained AEUnet-DIP proxy model can achieve efficient and accurate segmentation of asphalt concrete CT images of any image format, shape, resolution and size.

[0030] (3) Improved accuracy: The AEUnet-DIP surrogate model obtained after training has higher accuracy in segmenting slender aggregate phases.

[0031] (4) No usage threshold: The AEUnet-DIP proxy model obtained after training requires no prior knowledge or skills from the operator (such as morphological processing knowledge and programming skills). Anyone can easily achieve high-quality asphalt concrete CT image segmentation. Attached Figure Description

[0032] Figure 1 This is a complete technical roadmap of the present invention.

[0033] Figure 2 This is a schematic diagram of the CT scan results.

[0034] Figure 3 This is a schematic diagram illustrating the inherent defects in the grayscale value distribution of asphalt concrete in CT images.

[0035] Figure 4 This is a flowchart of the DIP processing procedure.

[0036] Figure 5 Comparison of the original CT image with the corresponding image segmentation results.

[0037] Figure 6 A schematic diagram illustrating a method for expanding data in an offline database.

[0038] Figure 7 This is a diagram of a two-layer AEUnet neural network.

[0039] Figure 8This is the complete AEUnet neural network model architecture.

[0040] Figure 9 This is the loss function curve for the AEUnet neural network model training process.

[0041] Figure 10 A comparison of the segmentation results of CT images from the same source.

[0042] Figure 11 Comparison of details of segmentation results for CT images from the same source.

[0043] Figure 12 Comparison of segmentation results for CT images from different sources and with different shapes.

[0044] Figure 13 A comparison chart showing the segmentation efficiency of images at different resolutions. Detailed Implementation

[0045] The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0046] This invention is a segmentation and prediction method for asphalt concrete CT images based on semantic segmentation. It mainly includes steps such as establishing an asphalt concrete CT image segmentation database, building and training a U-net semantic segmentation neural network model, and the main steps can be completed by computer programming.

[0047] The specific steps of this method are as follows:

[0048] Step 1: Acquiring initial CT images.

[0049] refer to Figure 2 In accordance with the requirements of the "Test Procedures for Asphalt and Asphalt Mixtures in Highway Engineering" and the "Technical Specifications for Construction of Asphalt Pavement on Highways," parameters such as aggregate gradation, void ratio, and asphalt-aggregate ratio were determined. Cylindrical asphalt concrete specimens with a height of 75 mm and a diameter of 65 mm were prepared using a rotary compactor. Image acquisition was performed using a German Diondo d2 general-purpose micro nano-focused CT system equipped with a 270 kV and 72 μA X-ray source. 750 tomographic images were acquired along the height of the specimen, resulting in an initial CT tomographic image set. The cross-sectional spacing was 0.1 mm, the effective resolution of the tomographic images was 914 pixels × 914 pixels, and the resolution was set to 78 micrometers / pixel.

[0050] Furthermore, asphalt concrete specimens may include hexahedrons, cylinders, and specimens with other arbitrary shapes and sizes.

[0051] Furthermore, the CT images in step one include 8-bit RGB color images and grayscale images.

[0052] Furthermore, each CT image in step one contains all pixel information of the complete specimen on that cross-section, with redundant pixel positions filled with "0" value pixels.

[0053] Step 2: Establishment and expansion of the offline image segmentation database.

[0054] refer to Figure 3 Due to physical limitations, the initial CT images acquired in Step 1 inherently possess the defect of "low to medium perimeter height and random noise distribution" in grayscale values, and the degree of this inherent defect varies among CT images from different sources. This results in a large-scale union of grayscale value ranges for different components in actual asphalt concrete CT images, significantly increasing the difficulty of CT image segmentation. Specifically, when the segmentation threshold is underestimated, some asphalt mortar phases are incorrectly identified as aggregate phases, leading to severe aggregate adhesion; when the segmentation threshold is overestimated, some aggregate phases are incorrectly identified as asphalt mortar phases, resulting in over-segmentation of the aggregates and the appearance of numerous internal pores.

[0055] To ensure the quality and consistency of image segmentation results, this invention designs a DIP processing flow to segment different components of the image based on the grayscale distribution characteristics and inherent defects of the original CT image. (Reference) Figure 4 The DIP processing flow used in this invention involves the following algorithms in sequence: A pixel filling algorithm is used to expand the CT image to 3840 pixels × 3840 pixels; a grayscale algorithm is used to uniformly adjust the grayscale value range of the image to 8-bit; a local mean filtering algorithm is used to denoise the image, with the local filtering kernel function selected as the four-connected domain operator [0 0.3 0, 0.3 0.3 0.3, 0 0.3 0]; a maximum threshold segmentation algorithm based on annular partitioning is used to uniformly divide the CT image from the center into one circular region and four annular regions, and the OTSU algorithm is used to achieve maximum threshold segmentation and reassembly for each region; a watershed segmentation algorithm is used to remove incorrectly connected aggregate phases; and a pixel restoration algorithm is used to restore the image to 914 pixels × 914 pixels, finally obtaining a binary image containing aggregate phase segmentation results and asphalt mortar phase segmentation results.

[0056] refer to Figure 5 The original CT image and the image segmentation result are cropped to obtain a 512-pixel × 512-pixel original CT image and image segmentation result. The original CT image and the corresponding image segmentation result (label) are used as a data pair. An offline image segmentation database (containing a total of 750 original image-label data pairs) is established based on the initial CT tomographic image set.

[0057] refer to Figure 6(a) Based on voxel reconstruction, model coordinate space rotation, and voxel model reslicing, a multi-angle CT scan process is simulated in computer memory. The aforementioned 750 original image-label data pairs can be used to regenerate 650 original image-label data pairs containing "new image segmentation knowledge." (Reference) Figure 6 (b) For a 3D voxel model, the theoretically selectable CT scan directions are infinite. However, scan directions with very similar minimums result in CT scans with essentially identical microstructures, failing to provide the neural network with more "effective image segmentation knowledge" during subsequent training. Therefore, in this embodiment, the scan angle interval is set to 20°, and the model is rotated sequentially around the x-axis and y-axis, ultimately yielding 81 asphalt concrete CT image segmentation data pairs in different directions. Simultaneously, if the CT scan slice interval is too small, the geometric knowledge of adjacent CT images is essentially the same, failing to promote the neural network training effect. Therefore, in this embodiment, the scan interval is set to 1mm. This results in 5422 sets of asphalt concrete CT original image-label data pairs containing "new image segmentation knowledge." The pixels of the 5422 sets of CT original images and their corresponding image labels are adjusted to 512 pixels × 512 pixels using pixel augmentation or pixel restoration algorithms.

[0058] Step 3: Optimization and development of neural network models.

[0059] refer to Figure 5 The label essentially provides a description of the aggregate phase contour boundary (concrete features) and position coordinates (abstract features) based on the image. Semantic segmentation neural networks adjust the receptive field size through successive convolution and pooling processes, learning image features under different receptive fields. Specifically, the semantic segmentation neural network learns concrete features from shallow networks (smaller receptive fields) and abstract features from deep networks (larger receptive fields), ultimately achieving an organic combination of concrete and abstract features through feature fusion operations. This unique network architecture learning method is highly compatible with the data features of asphalt concrete CT images before and after segmentation. However, the unique learning process of semantic segmentation neural networks is inevitably affected by other irrelevant features in the image, such as the original background, noise, and gray-level circular distribution features. These negative impacts increase the difficulty of neural network training convergence and reduce prediction accuracy.

[0060] To address this, this invention proposes an Attention-Enhanced Semantic Segmentation Neural Network (AEUnet). This network defines feature weights through an attention mechanism module to highlight important features and suppress irrelevant features. The AEUnet network structure consists of three main parts: an encoder, a decoder, and a feature fusion operation based on a channel attention mechanism. Figure 7A simple two-layer AEUnet diagram is shown, where paths a, b, and c represent the encoder, decoder, and feature fusion operation based on the channel attention mechanism, respectively.

[0061] (a) Encoder

[0062] The encoder compresses the input image to increase the receptive field, specifically by increasing image depth (number of image channels) while reducing image size. Each encoder block consists of several convolutional layers and one max-pooling layer. Its operating principle is as follows:

[0063] X (l) =ReLU(Conv(X) (l-1) W (l-1) )) (1)

[0064] X (l+1) =ReLU(Conv(X) (l) W (l) (2)

[0065] Y (l) =MaxPool(X (l+1) (3)

[0066] Among them, X (l-1) The input features of the encoder; W (l-1) Y is the convolutional kernel of the (l-1)th layer of the encoder; (l) The output features of the encoder are: Conv() is the convolution operation; ReLU() is the ReLU linear activation operation; MaxPool() is the max pooling operation.

[0067] (b) Decoder

[0068] The decoder enhances the perception of image details by expanding the image to reduce the receptive field. This is achieved by increasing the image size while reducing the number of image channels. Each decoder consists of one deconvolutional layer and several convolutional layers. Its operating principle is as follows:

[0069] X (l) =ReLU(TransConv(X) (l-1) W (l) (4)

[0070] X (l+1) =ReLU(Conv(X) (l) W (l) (5)

[0071] Y (l) =ReLU(Conv(X) (l+1) W (l+1) (6)

[0072] Among them, X (l-1) For the input features of the decoder; W (l-1) Y is the convolutional kernel of the (l-1)th layer of the decoder; (l) The output features of the decoder are: Conv() is the convolution operation; ReLU() is the ReLU linear activation operation; TransConv() is the deconvolution operation.

[0073] (c) Feature fusion based on channel attention mechanism

[0074] Channel attention mechanisms are integrated into the feature fusion stage between the encoder and decoder to achieve effective feature selection. The method for calculating attention weights is as follows:

[0075]

[0076] Among them, Y enc (l) Y represents the output feature of the encoder's l-th layer; dec (l) α is the input feature of the decoder's layer l; W is the weight matrix; Q is the query vector; K is the key vector; V is the value vector; α ij The attention weights between positions i and j; d k Here, represents the dimension of the key vector; Softmax() is the Softmax function used for normalization. The attention mechanism Attention(Q,K,V) can be calculated by weighted summation of the value vectors:

[0077] Attention(Q,K,V)=∑ j α ij V j (11)

[0078] Feature fusion is achieved within the same receptive field based on the Concat operation:

[0079]

[0080] Verification showed that AEUnet, with a receptive field of 5 levels, achieved the best segmentation and prediction performance for asphalt concrete CT images in this embodiment. Its neural network architecture parameters are as follows: Figure 8 As shown, from top to bottom, there are five layers: A (A1, A2, A3, A4, A5, A6), B (B1, B2, B3, B4, B5, B6), C (C1, C2, C3, C4, C5, C6, C7), D (D1, D2, D3, D4, D5, D6, D7), and E (E1, E2, E3, E4). The forward propagation order is as follows:

[0081] Input layer [512×512]: Default 3 channels. If the input is a single channel, it will be copied to a three-channel version.

[0082] Layer A1 [512×512]: Obtained from the input layer via a single-channel 1×1 convolution, single-channel;

[0083] Layer A2 [512×512]: Obtained from Layer A1 by 64-channel 3×3 convolution + padding, 64 channels;

[0084] A3 layer [512×512]: obtained by A2 layer through 64-channel 3×3 convolution + padding, 64 channels;

[0085] B1 layer [256×256]: obtained from A3 layer through channel 2×2 max pooling, 64 channels;

[0086] B2 layer [256×256]: obtained from B1 layer through 128-channel 3×3 convolution + padding, 128 channels;

[0087] B3 layer [256×256]: obtained from B2 layer through 128-channel 3×3 convolution + padding, 128 channels;

[0088] C1 layer [128×128]: obtained from B3 layer through channel 2×2 max pooling, 128 channels;

[0089] C2 layer [128×128]: obtained by C1 layer through 256-channel 3×3 convolution + padding, 256 channels;

[0090] C3 layer [128×128]: obtained by C2 layer through 256-channel 3×3 convolution + padding, 256 channels;

[0091] C4 layer [128×128]: obtained from C3 layer through 256-channel 3×3 convolution + padding, 256 channels;

[0092] D1 layer [64×64]: obtained from C4 layer through channel 2×2 max pooling, 256 channels;

[0093] D2 layer [64×64]: obtained from D1 layer through 512-channel 3×3 convolution + padding, 512 channels;

[0094] D3 layer [64×64]: obtained by D2 layer through 512-channel 3×3 convolution + padding, 512 channels;

[0095] D4 layer [64×64]: obtained by D3 layer through 512-channel 3×3 convolution + padding, 512 channels;

[0096] E1 layer [32×32]: obtained from D4 layer through channel 2×2 max pooling, 512 channels;

[0097] E2 layer [32×32]: obtained by E1 layer through 512-channel 3×3 convolution + padding, 512 channels;

[0098] E3 layer [32×32]: obtained by E2 layer through 512-channel 3×3 convolution + padding, 512 channels;

[0099] E4 layer [32×32]: obtained by E3 layer through 512-channel 3×3 convolution + padding, 512 channels;

[0100] D5 layer [64×64]: obtained by fusing the 2×2 deconvolution result of E4 layer with the features of D4 layer + attention mechanism, 1024 channels.

[0101] D6 layer [64×64]: obtained by D5 layer through 512-channel 3×3 convolution + padding, 512 channels;

[0102] D7 layer [64×64]: obtained by D6 layer through 512-channel 3×3 convolution + padding, 512 channels;

[0103] C5 layer [128×128]: obtained by fusing the 2×2 deconvolution result of D7 layer with the features of E4 layer + attention mechanism, 768 channels.

[0104] C6 layer [128×128]: obtained from C5 layer through 256-channel 3×3 convolution + padding, 256 channels;

[0105] C7 layer [128×128]: obtained from C6 layer through 256-channel 3×3 convolution + padding, 256 channels;

[0106] B4 layer [256×256]: obtained by fusing the 2×2 deconvolution result of C7 layer with the features of B3 layer + attention mechanism, 384 channels.

[0107] B5 layer [256×256]: obtained from B4 layer through 128-channel 3×3 convolution + padding, 128 channels;

[0108] B6 layer [256×256]: obtained from B5 layer through 128-channel 3×3 convolution + padding, 128 channels;

[0109] Layer A4 [512×512]: obtained by fusing the 2×2 deconvolution result of layer B6 with the features of layer A3 and the attention mechanism, with 192 channels.

[0110] Layer A5 [256×256]: Obtained from layer A4 through 64-channel 3×3 convolution + padding, 64 channels;

[0111] A6 layer [256×256]: obtained by A5 layer through 64-channel 3×3 convolution + padding, 64 channels;

[0112] Output layer [512×512]: obtained by single-channel 3×3 convolution of layer A6, single channel.

[0113] Step 4: Training the neural network model.

[0114] The AEUnet neural network model was trained using the offline image segmentation database established in step two. During training, the cross-entropy loss function was used to judge the learning progress, and the AEUnet-DIP proxy model was obtained after training.

[0115] In this embodiment, to improve prediction accuracy, the AEUnet neural network model was trained for 100 rounds using the offline image segmentation database established in step two, totaling 5244 image-label data pairs. The learning rate was set to 0.0001, the batch size to 32, the training stride to 169, and the total number of training steps to 16900. (Reference) Figure 9 The cross-entropy loss function was used to assess the learning progress, and the loss function showed a good convergence trend during training. The training process was performed on a high-performance Windows 10 computer with a core configuration of i5-13600KF 3.5GHz CPU, NVIDIA GeForce RTX3080TI 12G GPU, and 32GB 6000MHz DDR5 RAM.

[0116] Step 5: Image segmentation prediction and verification. The AEUnet-DIP proxy model obtained after training in Step 4 can directly achieve fast and accurate mapping from the original CT image of asphalt concrete to the segmentation result.

[0117] To verify the image segmentation prediction performance of the AEUnet-DIP surrogate model, a controlled experiment was designed with CT image source and shape as independent variables. The verification results are as follows:

[0118] (1) Comparison of segmentation results of CT images from the same source

[0119] refer to Figure 10Because the grayscale value ranges of the asphalt mortar phase and aggregate phase pixels in the asphalt concrete CT image set obtained in step one have a union, the image segmentation effect is unsatisfactory when the grayscale value threshold is overestimated or underestimated. The DIP processing flow designed and developed in step two achieves image segmentation very well, which proves the effectiveness and reliability of the offline image segmentation database established in this invention. The AEUnet-DIP proxy model reproduces the label results of the offline image segmentation database very well, proving that the AEUnet architecture designed and developed in step three is reasonable and can effectively "learn" the mapping relationship between the original image and the label.

[0120] refer to Figure 11 In terms of restoring aggregate contour details, the AEUnet-DIP surrogate model surpasses the accuracy of the DIP processing workflow. For example, randomly distributed slender aggregates are more susceptible to the influence of surrounding asphalt mortar phase pixels during morphological processing. The DIP processing workflow essentially performs global or predefined region morphological processing on the entire image, failing to handle randomly distributed slender aggregates locally. Therefore, the DIP processing workflow often segments slender aggregates into two or more aggregates distributed linearly. The essence of the AEUnet-DIP surrogate model lies in first establishing probes for identifying aggregate and asphalt mortar phase pixels through a training process, and then using these probes to distinguish aggregate and asphalt mortar phase pixels based on local grayscale distribution patterns in the image. Therefore, the AEUnet-DIP surrogate model can accurately handle aggregates of any shape.

[0121] (2) Comparison of segmentation effects of CT images from different sources and with different shapes

[0122] refer to Figure 12 The results of underestimating and overestimating the thresholds show that there are significant differences in the union interval of pixel gray values ​​among CT image samples from three different sources. Specifically, the union interval range of the same-source image (sample I), different-source image (sample II), and different-source image (sample III) gradually increases, indicating that the difficulty of image segmentation gradually increases.

[0123] The DIP processing flow demonstrated good segmentation results in samples I and II. However, as the pixel grayscale values ​​of the aggregate and asphalt mortar phases overlapped, the segmentation process lost geometric information in the central part of sample III, specifically manifested as incomplete aggregates and missing aggregates.

[0124] The AEUnet-DIP surrogate model demonstrated superior image segmentation results across three different sample sets. Compared to CT image segmentation procedures, the AEUnet-DIP surrogate model performed better on sample III, effectively segmenting the aggregate phase in the central region of the CT image. This indicates that the AEUnet-DIP surrogate model possesses excellent generalization ability on asphalt concrete CT images from different sources and with varying shapes.

[0125] (3) Comparison of segmentation efficiency for images at different resolutions

[0126] refer to Figure 13 The AEUnet-DIP surrogate model can directly process image files of any format, size, or color space, avoiding the preprocessing steps such as format adjustment, grayscale conversion, and image denoising required in the DIP processing workflow, thereby further improving segmentation efficiency. It has been verified that in segmenting AC CT images of different specifications, the AEUnet-DIP surrogate model achieves an exponential improvement in processing efficiency compared to the DIP processing workflow.

[0127] Furthermore, the CT images from different sources in step five include those from different asphalt concrete specimens, different CT scanning devices, and different CT scanning environments.

[0128] Furthermore, the different shaped CT images in step five include circular, square, and other arbitrary shapes and size characteristics, and the image resolutions include 1507×914, 1920×1080, 3840×2160, and other size ratio types.

[0129] Furthermore, when processing CT images of different sizes, CT image segmentation programs require code modifications based on image characteristics, which poses a significant challenge for users lacking image processing and programming skills. However, the robust versatility of the AEUnet-DIP proxy model perfectly solves this problem, allowing users to easily achieve efficient and accurate segmentation of CT images of asphalt concrete of varying sizes.

[0130] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments, and all fall within the scope of protection of the technical solution of this application.

Claims

1. A method for segmenting and predicting asphalt concrete CT images based on semantic segmentation, characterized in that, Includes the following steps: Step 1: Initial CT image acquisition: Prepare asphalt concrete CT scan specimens, determine the scanning direction, and use CT scanning technology to acquire its initial CT tomographic image set; Step 2: Establishment and expansion of the offline image segmentation database; First, based on the grayscale distribution characteristics and inherent defects of the initial CT tomographic images, a DIP processing flow is designed to segment different components of the image; Then, using the initial CT original image and the corresponding image segmentation result as a set of data pairs, a high-quality offline image segmentation database is established based on the initial CT tomographic image set. Finally, using CT scan computer simulation algorithms, new tomographic images in different directions are extracted from the asphalt concrete CT tomographic image set obtained in step one, thereby expanding the number of offline image segmentation databases. Step 3: Neural Network Model Optimization and Development: Based on the data structure characteristics of the offline image segmentation database, a U-net neural network model considering the attention mechanism, namely the AEUnet neural network model, is constructed based on semantic segmentation, attention mechanisms, and supervised learning theory. The AEUnet neural network model includes: The AEUnet neural network model is based on the basic semantic segmentation network U-net. It realizes feature fusion optimization under the receptive field of each layer of the neural network based on the channel attention mechanism. The attention mechanism module defines feature weights to highlight important features and suppress irrelevant features. The AEUnet neural network architecture includes: an encoder, a decoder, and a feature fusion operation based on a channel attention mechanism, wherein: The encoder compresses the input image to increase the receptive field, specifically by increasing the image depth (i.e., the number of image channels) while reducing the image size. Each encoder consists of several convolutional layers and a max-pooling layer, and its operating principle is as follows: (1) (2) (3) in, X (l-1) These are the input features of the encoder; W (l) For encoder number l The convolution kernel of the layer, W (l-1) For encoder number l -1 layer convolutional kernel; Y (l) The encoder output features; Conv() is the convolution operation; ReLU() is the ReLU linear activation operation; MaxPool() is the max pooling operation; The decoder enhances the perception of image details by expanding the image to reduce the receptive field; this is achieved by increasing the image size while reducing the number of image channels. Each decoder includes a deconvolution layer and several convolutional layers, and its operating principle is as follows: (4) (5) (6) in, X (l-1) These are the input features for the decoder; W (l-1) For decoder number l -1 layer convolutional kernel; Y (l) The output features of the decoder; Conv() is the convolution operation; ReLU() is the ReLU linear activation operation; TransConv() is the deconvolution operation; The feature fusion operation based on the channel attention mechanism integrates the channel attention mechanism into the feature fusion stage between the encoder and decoder to achieve effective feature selection; the method for calculating attention weights is as follows: (7) (8) (9) (10) in, Y enc (l) For encoder number l The output features of the layer; Y dec (l) For decoder number l The input features of the layer; W This is the weight matrix; Q For query vector, It is a location i The query vector; K The key vector; V It is a value vector; α ij For position i and j Attention weights between them; d k The dimension of the key vector; Softmax() is the Softmax function used for normalization; Attention mechanism (... Q , K , V It is calculated by weighted summation of the value vectors: (11) in, It is a location j The value vector; Feature fusion is achieved within the same receptive field based on the Concat operation: (12)； Step 4: Neural Network Model Training: The AEUnet neural network model is trained using the image segmentation offline database established in Step 2 to obtain the AEUnet-DIP proxy model; Step 5: Image segmentation prediction: The AEUnet-DIP surrogate model is used to perform aggregate phase segmentation prediction on the asphalt concrete CT image to be tested.

2. The semantic segmentation-based CT image segmentation and prediction method for asphalt concrete as described in claim 1, characterized in that, The CT images in step one include 8-bit RGB color images and grayscale images.

3. The method for segmenting and predicting asphalt concrete CT images based on semantic segmentation as described in claim 1, characterized in that, Each CT image in step one contains all pixel information of the complete specimen on that cross-section, with extra pixel positions filled with "0" value pixels.

4. The method for segmenting and predicting asphalt concrete CT images based on semantic segmentation as described in claim 1, characterized in that, The image segmentation results in step two include aggregate phase segmentation results and asphalt mortar phase segmentation results.

5. The method for segmenting and predicting asphalt concrete CT images based on semantic segmentation as described in claim 1, characterized in that, The DIP processing flow includes pixel filling, image filtering, maximum threshold segmentation, watershed segmentation, and pixel restoration.

6. The method for segmenting and predicting asphalt concrete CT images based on semantic segmentation as described in claim 1, characterized in that, The CT scan computer simulation algorithm includes voxel reconstruction, model coordinate space rotation, and voxel model reslicing.

7. The method for segmenting and predicting asphalt concrete CT images based on semantic segmentation as described in claim 1, characterized in that, The DIP processing flow involves the following algorithms in sequence: A pixel-filling algorithm is used to expand the CT image to 3840 pixels × 3840 pixels; a grayscale algorithm is used to uniformly adjust the image grayscale value range to 8-bit; a local mean filtering algorithm is used to denoise the image, with the local filtering kernel function selected being the four-connected domain operator [0 0.3 0, 0.3 0.3 0.3, 0 0.3 0]; a maximum threshold segmentation algorithm based on annular partitioning is used to uniformly divide the CT image from the center into one circular region and four annular regions, and the OTSU algorithm is used to achieve maximum threshold segmentation and reassembly; a watershed segmentation algorithm is used to remove incorrectly connected aggregate phases. The image was restored to 914 pixels × 914 pixels using a pixel restoration algorithm, resulting in a binary image containing aggregate phase segmentation results and asphalt mortar phase segmentation results. The original CT image and the image segmentation result were cropped to obtain a 512-pixel × 512-pixel original CT image and image segmentation result. The original CT image and the corresponding image segmentation result were used as a data pair. An offline image segmentation database was established based on the initial CT tomographic image set, which contained a total of 750 data pairs. Based on voxel reconstruction, model coordinate space rotation algorithm, and voxel model reslicing algorithm, a CT scanning process from multiple angles was simulated in computer memory. The scanning interval was set to 1 mm, resulting in 5422 sets of data pairs containing "new image segmentation knowledge". The pixels of the 5422 sets of original CT images and corresponding image segmentation results were adjusted to 512 pixels × 512 pixels through pixel augmentation or pixel restoration algorithms.

Images