Asphalt mixture CT image small sample multi-scale classification method based on deep reasoning decision

Abstract

The invention discloses an asphalt mixture CT (Computed Tomography) image small sample multi-scale classification method based on a deep reasoning decision, belongs to the technical field of traffic infrastructure digitization, and aims to solve the problems that a traditional threshold segmentation method is insufficient in multiphase material processing capability and a pure deep learning model is insufficient in mesoscopic interface segmentation precision. The method comprises the following steps: S1, constructing an asphalt mixture CT (Computed Tomography) image mixed data set; s2, a multi-scale expansion convolution model CNN-S is built, and aggregate and asphalt mortar edge multi-scale classification is achieved; the CNN-S captures macro-micro characteristics of aggregates, gaps and mortar through multi-expansion-rate parallel convolution; a causal feature correction module is introduced, a high confidence region and a low confidence region are divided, a backbone network processes a macrostructure, and an edge branch decoder identifies a mesoscopic interface; s3, dynamic reasoning optimization: deploying a DeepSeek-R1 reasoning model as an auxiliary dynamic decision maker, and recording a misjudgment mode; and based on training indexes and expert knowledge, dynamically correcting CNN-S model weight and a loss function to realize self-optimization training.

Description

Technical Field

[0001] This invention relates to the field of digital technology for transportation infrastructure, and in particular to a multi-scale material classification method for small-sample asphalt mixture CT images that integrates deep reasoning decision-making. This method is used to solve the problems of low recognition accuracy and model overfitting in industrial CT image segmentation caused by insufficient sample size and similar boundary gray levels. Background Technology

[0002] Digital information models are a fundamental element in the construction of physical entities and an important aspect of the digital transformation of transportation infrastructure. In recent years, with the rise of emerging technologies such as artificial intelligence, liquid neural networks, and virtual reality, utilizing computer vision and artificial intelligence image recognition technologies to analyze the internal structural properties of physical entities has proven technically feasible for constructing digital twin models of asphalt mixtures, enabling the analysis, control, and optimization of the material's unique properties.

[0003] Asphalt mixtures, as the core load-bearing material in road engineering, directly determine the service performance and durability of pavements due to their multi-scale, heterogeneous composite structure. Accurate identification and quantitative characterization of their internal aggregate distribution, void structure, and interface properties are crucial technological foundations for material performance prediction, optimized design, and life-cycle management. Therefore, achieving high-precision image segmentation of asphalt mixtures can ensure the extraction of mesoscopic parameters and the construction of information models that capture the mesoscopic structural characteristics of asphalt mixtures. Currently, mainstream image segmentation techniques are divided into threshold-based segmentation and deep learning-based segmentation. Traditional threshold-based segmentation techniques, such as those using watersheds, are insufficient for handling various types of mixtures and are prone to aggregate adhesion. While deep learning image segmentation has good model generalization capabilities, it is prone to edge recognition misjudgment in areas where the asphalt mortar and aggregate boundaries are close to the threshold, leading to model overfitting. Therefore, it is necessary to improve existing image segmentation models to address their shortcomings, such as insufficient data sample size, inadequate accuracy in identifying complex boundaries, and overfitting caused by edge recognition errors. This improvement is of great significance for exploring the multi-scale evolution behavior of asphalt mixture macroscopic properties and for developing multi-scale mechanical analysis methods for asphalt mixtures.

[0004] In summary, current technology suffers from three major pain points:

[0005] 1. Small sample size problem: Scarcity of labeled data leads to poor generalization ability of the model;

[0006] 2. Blurred boundaries: The similar grayscale of aggregates and asphalt mortar leads to misjudgment of edges;

[0007] 3. Overfitting risk: Misjudgment of edges can lead to local optima traps in the model.

[0008] It is evident that traditional threshold segmentation methods (such as watersheds) are insufficient for handling multiphase materials, while pure deep learning models lack sufficient accuracy in segmenting mesoscopic interfaces. Summary of the Invention

[0009] To address the shortcomings of traditional threshold segmentation methods in handling multiphase materials, and the insufficient segmentation accuracy of pure deep learning models at micro-interfaces, this invention provides a small-sample multi-scale classification method for asphalt mixture CT images based on deep inference decision-making.

[0010] The present invention describes a multi-scale classification method for small samples of asphalt mixture CT images based on deep reasoning decision-making. This method includes the following steps:

[0011] S1. Construct a hybrid dataset of CT tomographic images of asphalt mixtures;

[0012] S2. Construct a multi-scale dilated convolutional model CNN-S to achieve multi-scale classification of aggregate and asphalt mortar edges;

[0013] The multi-scale dilated convolutional model CNN-S captures the macro- and micro-features of aggregates, voids, and mortar through parallel convolution with multiple dilation rates; it introduces a causal feature correction module to divide high and low confidence regions, with the backbone network processing the macro structure and the edge branch decoder identifying the micro interface.

[0014] S3. Dynamic Inference Optimization: Deploy the DeepSeek-R1 inference model as an auxiliary dynamic decision-maker to record misjudgment patterns; dynamically adjust the weights and loss function of the CNN-S model based on training metrics and expert knowledge to achieve self-optimization training.

[0015] Preferably, step S1, which involves constructing a mixed dataset of CT tomographic images of asphalt mixtures, includes:

[0016] S11. Select N asphalt specimens under various working conditions, and label each group with the true value labels of M CT images.

[0017] S12. Use the initial CNN-S model to generate pseudo-labels for unlabeled images, and mix the real values ​​and pseudo-labels at a ratio of 1:9 to form an initial mixed training set.

[0018] S13. Substitute the initial mixed training set into the model for secondary training. By strengthening contrastive learning and consistency regularization, labels with confidence > 90% are selected to construct a mixed dataset of asphalt mixture CT tomographic scan images.

[0019] Preferably, the truth label in step S11 is to mark the voids, aggregates, asphalt mortar and boundaries of the asphalt specimen.

[0020] Preferably, the multi-scale dilated convolutional model CNN-S in step S2 employs a dual-stream feature decoder, and the specific process of multi-scale classification is as follows:

[0021] S21. Using the causal feature correction module, the test image is traversed through a lightweight parallel convolutional layer to obtain the confidence level at each position of the image. Based on the confidence level threshold, high and low confidence regions are divided, and the macroscopic structure region and microscopic interface distribution region of the image are defined.

[0022] S22. The backbone feature extraction algorithm Encoder is used to capture the spatial distribution of each macroscopic structure in the scanned image through high dilation rate convolution of the aggregate core area, the void core area and the mortar uniform area.

[0023] S23. An edge detection branch algorithm is adopted. A single-channel boundary probability distribution map is obtained through low dilation rate convolution. The Sigmoid activation function is used for mapping to obtain the edge probability of pixels at the fine interface. The complex fine interface structure features are identified through the edge probability map.

[0024] Preferably, in step 23, pixels with an edge probability threshold ≥ 0.85 are determined to be valid edges.

[0025] Preferably, the void ratio of high dilation rate convolution is ≥6, and the void ratio of low dilation rate convolution is ≤2.

[0026] Preferably, step S3, deploying the DeepSeek-R1 inference model as an auxiliary dynamic decision-maker, includes:

[0027] S31. Deploy the local runtime environment of the DeepSeek-R1 inference model, introduce domain knowledge and physical equation constraints to fine-tune the language model, and use the policy generation algorithm to limit the data representation of evaluation indicators and model decision commands during the training process.

[0028] S32. Establish a model connection port, record the changes in the misjudgment pattern of aggregate-mortar-void in the segmentation model and the improvement trend of boundary recognition accuracy, and transmit them to the auxiliary decision model. Use linguistic logical reasoning and expert experience constraints to correct the model segmentation weights and adjust the training loss function, and correct the model's recognition accuracy of fine interfaces in real time to achieve self-optimization training of the image segmentation model.

[0029] Preferably, step S3 uses the cross-entropy function loss rate, accuracy, center-IOU, edge-IOU, and single-class block ratio generated every ten training iterations as evaluation indicators for parameter fine-tuning and loss function correction of the DeepSeek-R1 inference model through an interval update strategy.

[0030] The beneficial effects of this invention are:

[0031] This invention proposes a training scheme for a hybrid dataset of CT tomographic images of asphalt mixtures. The original training set is expanded through secondary iterative model training. High-quality hybrid image training datasets are generated by strengthening contrastive learning and using consistency regularization criteria. A dual-stream architecture edge branch decoder with a main feature extraction and edge detection branch is constructed. Addressing the challenge of subtle grayscale differences in CT images at fine interfaces, the Sigmoid activation function is used to map the edge probability map of a single channel, obtaining the edge probability of pixels at fine interfaces and achieving high-confidence recognition of complex fine interface structural features. Simultaneously, a DeepSeek-R1 inference model (RLM) is introduced as an auxiliary dynamic decision-maker. By recording changes in the misjudgment pattern of aggregate-mortar-void and monitoring the improvement trend of boundary recognition accuracy, key indicators and expert experience constraints during training are used to correct the model segmentation weights and adjust the training loss function. This avoids model overfitting and local optima traps caused by misjudgments in complex edge recognition, achieving self-optimization training of the image segmentation model. Specifically, the scheme includes the following points:

[0032] 1. Hybrid datasets improve model robustness: The intersection-union ratio of aggregate segmentation increases by 0.82, which is better than the traditional 0.68;

[0033] 2. Dual-stream structure solves the grayscale similarity boundary problem: Confidence level for fine-grained interface recognition >90%;

[0034] 3. The RLM decision generator avoids overfitting and achieves self-optimization in the training process: the training cycle is shortened by 50%. Attached Figure Description

[0035] Figure 1 This is a true dataset of asphalt mixtures under multiple working conditions; where Figure 1 (a) is a real image. Figure 1 (b) is Figure 1 (a's image label (boundary)) Figure 1 (c) is a real image. Figure 1 (d) for Figure 1 (c) Image labels (aggregate + void);

[0036] Figure 2 This is a boundary probability distribution diagram of asphalt mixtures; where Figure 2 (a) is a real image. Figure 2 (b) is the confidence distribution image; (b) is the confidence distribution image, in which the main structure is represented by a green solid and the boundary is represented by a yellow border;

[0037] Figure 3 This is a schematic diagram of the DS-R1-CNN-S model of the present invention;

[0038] Figure 4This is a comparison chart of the training results of different image segmentation models for the same sample. Figure 4 (a) is a real image. Figure 4 (b) shows the results after training the CNN model for 45 minutes. Figure 4 (c) Results of training the CNN-DNN model for 8.5 hours. Figure 4 (d) shows the results after 6 hours of training using the proposed DS-R1-CNN-S model;

[0039] Figure 5 A comparison chart of the accuracy evaluation metrics (F1 score) for image segmentation models;

[0040] Figure 6 This is a comparison chart of image segmentation model accuracy evaluation metrics (model accuracy). Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0043] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0044] Specific Implementation Method 1: The following is combined with... Figures 1 to 6This embodiment describes a multi-scale classification method for small samples of asphalt mixture CT images based on deep inference decision-making. The invention employs a DeepSeek-R1 model combined with a multi-objective dilated convolutional neural network (CNN-S) as the asphalt mixture tomographic image segmentation model framework (DS-R1-CNN-S). This model first modifies the traditional convolutional neural network by replacing the traditional convolutional-connector layer with a dilated convolutional layer. By dilating the CT-annotated image, it captures small gaps and difficult-to-identify aggregate-mortar boundaries with a low dilation rate, and identifies large aggregate structures with a high dilation rate, thus achieving refined capture of macro- and micro-scale structural features. Furthermore, during the construction of the image segmentation model framework, the reinforcement learning and logical reasoning capabilities of the DeepSeek-R1 model are utilized. Through an interval update strategy, the cross-entropy function loss rate, accuracy, center-IOU, edge-IOU, and the proportion of single-class blocks generated every ten training iterations are used as evaluation indicators for parameter fine-tuning and loss function correction of the large inference model. Logical judgment is made on statistical data through experience backtracking and supervised learning modules. Combined with the historical backtracking capability of the DeepSeek-R1 model, the model's recognition attention to difficult-to-segment blocks (aggregate-asphalt mortar boundary) is dynamically improved. Based on key indicators and expert experience constraints during the training process, the model segmentation weights are corrected and the training loss function is adjusted to achieve refined material classification of voids, aggregates, asphalt mortar, and mixture boundaries in industrial CT tomographic images of asphalt mixtures.

[0045] Specifically, the method includes the following steps:

[0046] S1. Construct a hybrid dataset of CT tomographic images of asphalt mixtures;

[0047] S2. Construct a multi-scale dilated convolutional model CNN-S to achieve multi-scale classification of aggregate and asphalt mortar edges;

[0048] The multi-scale dilated convolutional model CNN-S captures the macro- and micro-features of aggregates, voids, and mortar through parallel convolution with multiple dilation rates; it introduces a causal feature correction module to divide high and low confidence regions, with the backbone network processing the macro structure and the edge branch decoder identifying the micro interface.

[0049] S3. Dynamic Inference Optimization: Deploy the DeepSeek-R1 inference model as an auxiliary dynamic decision-maker to record misjudgment patterns; dynamically adjust the weights and loss function of the CNN-S model based on training metrics and expert knowledge to achieve self-optimization training.

[0050] Step S1 involves constructing a mixed dataset of CT tomographic images of asphalt mixtures, which includes:

[0051] S11. Select N asphalt specimens under various working conditions, and label each group with the true value labels of M CT images.

[0052] S12. Use the initial CNN-S model to generate pseudo-labels for unlabeled images, and mix the real values ​​and pseudo-labels at a ratio of 1:9 to form an initial mixed training set.

[0053] S13. Substitute the initial mixed training set into the model for secondary training. By strengthening contrastive learning and consistency regularization, labels with confidence > 90% are selected to construct a mixed dataset of asphalt mixture CT tomographic scan images.

[0054] The truth labels in step S11 are used to mark the voids, aggregates, asphalt mortar, and boundaries of the asphalt specimen.

[0055] This step utilizes reinforced contrastive learning and the uniformity criterion to perform secondary iterative training of the model on unlabeled images, retains high-confidence images for training labels, and integrates manually labeled ground truth images into the training set, proposing a training scheme for a hybrid dataset of asphalt mixture CT tomographic scan images.

[0056] This scheme first uses asphalt Marshall specimens under various working conditions for manual image annotation (N=100, M=10). Ten ground truth images are labeled for each sample group as the initial training set. An initial classification model for the mixture is obtained through iterative supervised training. Then, pseudo-labels for unlabeled images are obtained from the initial classification model. A dynamic threshold strategy is used to proportionally mix ground truth and pseudo-label images to form an initial mixed training set, which is then substituted into the model for weakly supervised training. By embedding reinforcement contrastive learning and consistency regularization criteria into the training model, pseudo-labels are selectively filtered. A confidence scoring mechanism is used to retain high-confidence image labels, forming a representative mixed CT image training set. This scheme effectively solves the problems of weak model generalization ability and poor classification accuracy in this image classification algorithm under iterative training with small sample data, significantly improving model robustness and reducing the material classification error rate of asphalt mixture tomographic images by approximately 18.5%.

[0057] Step S2 of the multi-scale dilated convolutional model CNN-S employs a two-stream feature decoder. The specific process of multi-scale classification is as follows:

[0058] S21. Using the causal feature correction module, the test image is traversed through a lightweight parallel convolutional layer to obtain the confidence level at each position of the image. Based on the confidence level threshold, high and low confidence regions are divided, and the macroscopic structure region and microscopic interface distribution region of the image are defined.

[0059] S22. The backbone feature extraction algorithm Encoder is used to capture the spatial distribution of each macroscopic structure in the scanned image through high dilation rate convolution of the aggregate core area, the void core area and the mortar uniform area.

[0060] S23. An edge detection branch algorithm is adopted. A single-channel boundary probability distribution map is obtained through low dilation rate convolution. The Sigmoid activation function is used for mapping to obtain the edge probability of pixels at the fine interface. The complex fine interface structure features are identified through the edge probability map.

[0061] In step 23, pixels with an edge probability threshold ≥ 0.85 are determined to be valid edges.

[0062] High dilation rate convolution has a void ratio ≥ 6, while low dilation rate convolution has a void ratio ≤ 2.

[0063] This step addresses the problem of blurred boundaries and unclear aggregate edge recognition caused by the similar grayscale of aggregates and asphalt mortar during the recognition of tomographic scan images of asphalt mixtures. It proposes a confidence-guided feature fusion and causal feature correction algorithm to achieve multi-scale classification of aggregate and asphalt mortar edges.

[0064] This step employs an edge branch decoder with a dual-stream architecture of backbone feature extraction and edge detection branches. It captures the macroscopic structure, primarily consisting of the aggregate core region, void core region, and mortar homogeneous region, and the microscopic interface, primarily consisting of the aggregate-mortar transition region and micropore boundaries. First, a causal correction module uses lightweight convolutional layers to calculate the confidence level at each location in the image in real time, thus dividing the image into macroscopic structure and microscopic interface distribution regions. The macroscopic structure, with its homogeneous material, appears as clearly defined solid color blocks in the tomographic image; its spatial distribution in the scan image can be quickly captured by the backbone feature extraction algorithm. At the microscopic interface, the grayscale differences in the CT image are weak. Therefore, the edge detection branch algorithm obtains a single-channel boundary probability distribution map, which is then mapped using the Sigmoid function to obtain the edge probability of pixels at the microscopic interface. This edge probability map is used to identify complex microscopic interface structural features, achieving accurate multi-scale material classification.

[0065] Step S3, deploying the DeepSeek-R1 inference model as an auxiliary dynamic decision maker, includes:

[0066] S31. Deploy the local runtime environment of the DeepSeek-R1 inference model, introduce domain knowledge and physical equation constraints to fine-tune the language model, and use the policy generation algorithm to limit the data representation of evaluation indicators and model decision commands during the training process.

[0067] S32. Establish a model connection port, record the changes in the misjudgment pattern of aggregate-mortar-void in the segmentation model and the improvement trend of boundary recognition accuracy, and transmit them to the auxiliary decision model. Use linguistic logical reasoning and expert experience constraints to correct the model segmentation weights and adjust the training loss function, and correct the model's recognition accuracy of fine interfaces in real time to achieve self-optimization training of the image segmentation model.

[0068] Step S3 uses the cross-entropy function loss rate, accuracy, center-IOU, edge-IOU, and single-class block ratio generated every ten training iterations as evaluation indicators for parameter fine-tuning and loss function correction of the DeepSeek-R1 inference model through an interval update strategy.

[0069] This step addresses the overfitting problem caused by the deep learning model's misjudgment of small-scale aggregate edges during iterative training. It proposes a material classification model for asphalt mixture tomographic scan images, using the DeepSeek-R1 inference model as an auxiliary dynamic decision-making mechanism.

[0070] This model employs a dilated convolutional neural network model with a void space pyramid pooling framework. By running parallel convolutional layers with different dilation rates, it can simultaneously capture the macroscopic distribution of aggregates, voids, and mortar, as well as the fine texture of the aggregate-mortar transition zone and microporous interface. To achieve the organic integration of deep inference decision-making and small-sample training, the DeepSeek-R1 inference model (RLM) is introduced as an auxiliary dynamic decision-maker. Through the RLM policy generation algorithm, the changes in the misclassification patterns of aggregate-mortar-void are recorded and the improvement trend of material classification accuracy is monitored. Based on key indicators and expert experience constraints during the training process, the model classification weights are corrected and the training loss function is adjusted. By dynamically analyzing the training state, the model's material classification accuracy for the fine interface is corrected in real time, realizing the self-optimization training of the asphalt mixture tomographic scanning classification model.

[0071] The following is a specific example. The image training set used in this invention consists of 10 groups of asphalt specimens under different working conditions (Table 1). Ten images of each group of samples were selected at equal intervals for image annotation. Figure 1 The inference model used in this invention is the DeepSeek-R1 32b localized model.

[0072] Table 1 Test parameters of specimens under different working conditions

[0073]

[0074] The specific process includes:

[0075] Step 1: Construct a mixed dataset of CT tomographic images of asphalt mixtures;

[0076] Image annotation software was used to manually label asphalt mixture samples under 10 working conditions, and the image labels and ground value images were combined to establish a ground value data training set.

[0077] The true training set is substituted into the model for iterative training to obtain the parameters of the image segmentation model, and the unlabeled image is segmented to obtain pseudo-labels.

[0078] By using the model's reinforcement contrastive learning and consistency regularization criteria, the mixed dataset is cleaned and filtered, retaining data labels with a confidence level greater than 90% to form a mixed training dataset. The verification indicators are shown in Table 2.

[0079] Table 2 Dataset Validation Metrics

[0080]

[0081] Step 2: The causal feature correction module and the backbone feature extraction + edge detection decoder of the dual-stream architecture are built to construct a multi-scale dilated convolutional model CNN-S to achieve multi-scale classification of aggregate and asphalt mortar edges;

[0082] During model training, parallel convolutional layers are used to traverse the validation set data images. Confidence distribution images are obtained through visualization code, and the internal regions of the images are divided based on confidence thresholds. High-confidence and low-confidence regions are defined based on these thresholds, and macroscopic structural regions and microscopic interface distribution regions are defined, such as... Figure 2 ;

[0083] The backbone feature algorithm is used to quickly capture high-confidence regions mainly consisting of aggregate core area, void core area and mortar uniform area to obtain the spatial distribution of the structure.

[0084] By using an edge detection branch algorithm, the edge probability of pixels at the micro-interface is obtained, and an activation function is used for mapping to specifically identify the structural features of complex micro-interfaces.

[0085] Step 3: Deploy the DeepSeek-R1 inference model as an auxiliary dynamic decision-maker;

[0086] like Figure 3 As shown, a localized DeepSeek-R1 32b language large model is deployed and fine-tuned. A model connection port is established, and the data representation of evaluation metrics and model decision commands are defined during model training.

[0087] A historical trend analysis module is constructed to record the changes in the misjudgment patterns of aggregate-mortar-void in the segmentation model and the improvement trend of boundary recognition accuracy, and transmits this information to the auxiliary decision-making model to achieve self-correction and self-optimization of the model.

[0088] like Figure 4 As shown, embedded visualization verification code is used to monitor the model's segmentation effect and language model correction effect through image segmentation evaluation indicators. Based on the above model training process, a multi-source mixed training set of asphalt mixture industrial CT scans is input into the segmentation network for iterative optimization. After 35 training cycles, the final image segmentation model is generated. To verify the model's segmentation accuracy and generalization ability, three sets of asphalt mixture tomographic scan data are selected as sample training sets. Traditional convolutional neural networks (CNN), CNN-DNN dual-processing convolutional neural networks, and the DS-R1-CNN-S intelligent segmentation model are used for image segmentation, respectively, and model accuracy is verified. Specific data are shown below. Figure 4 :

[0089] After segmenting the test set samples using different models, it was found that the traditional convolutional neural network (CNN) model had the shortest training time. When combined with image segmentation results, it can be seen that this model can effectively segment the voids in the mixture, but its segmentation accuracy for asphalt mortar and aggregate components is poor, indicating overfitting. While the CNN-DNN dual-processing image segmentation model achieved segmentation of asphalt mortar and aggregate components in most areas of the image, it had the longest training cycle and still exhibited image segmentation errors in regions with similar gray levels (L1 region). This is mainly attributed to its fixed weight parameter mechanism failing to integrate the fine feature extraction capabilities of low-dilation-rate convolutional kernels. The force caused the high gray-scale similarity asphalt mortar to be misclassified as small-sized aggregate. In image (d), the model uses green, red+green, yellow and blue to label the four blocks of voids, boundaries, aggregate and asphalt mortar respectively. It can be seen that the DS-R1-CNN-S image segmentation model, based on the traditional segmentation model, has achieved accurate decoupling segmentation of fine aggregate and asphalt mortar in L1 gray-scale similarity area by introducing an adaptive multi-scale dilated convolution module and combining a dynamic parameter optimization strategy based on a large inference model. Furthermore, it has improved the segmentation accuracy of asphalt mixture CT images by optimizing the decision logic through cross-level feature transfer.

[0090] To quantify the segmentation accuracy of the image segmentation model on the asphalt mixture CT model, this study selected accuracy (Acc) and F1 score (see Formula 15) as image segmentation evaluation indicators. The training results are shown in [Figure 15]. Figure 5 and Figure 6The DS-R1-CNN-S model used in this study achieved a final accuracy of 75.6%, which is significantly higher than CNN-DNN (70.5%) and traditional CNN (67.1%). While traditional CNN has high training efficiency and a short training cycle, based on image data, it is known that the accuracy and F1 score of this model do not increase significantly in the later stages of image training and exhibit a plateau. This is because traditional convolutional neural network models have weak accuracy in recognizing gray-scale mixed regions, leading to overfitting and class bias during model training, severely affecting image segmentation accuracy. Although the CNN-DNN dual-processing model improves CT image segmentation accuracy to 70% through its dual-channel design, its fixed weight mechanism makes it prone to local optima during training, making it difficult to achieve fine-grained image segmentation. Figure 5 , Figure 6 As shown, the DS-R1-CNN-S segmentation model reaches an inflection point in stages D1, D2, and D3. This is because the model uses an interval update strategy to use the cross-entropy function loss rate, accuracy, center-IOU, edge-IOU, and the proportion of single-class blocks generated every ten training iterations as evaluation metrics for the large inference model. The model updates and validates the image segmentation model using the weights and loss function provided by the model, thus avoiding the model from getting trapped in local optima.

[0091] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A multi-scale classification method for small samples of asphalt mixture CT images based on deep reasoning decision-making, characterized in that, The method includes the following steps: S1. Construct a hybrid dataset of CT tomographic images of asphalt mixtures; S2. Construct a multi-scale dilated convolutional model CNN-S to achieve multi-scale classification of aggregate and asphalt mortar edges; The multi-scale dilated convolutional model CNN-S captures the macro- and micro-features of aggregates, voids, and mortar through parallel convolution with multiple dilation rates; it introduces a causal feature correction module to divide high and low confidence regions, with the backbone network processing the macro structure and the edge branch decoder identifying the micro interface. S3. Dynamic Inference Optimization: Deploy the DeepSeek-R1 inference model as an auxiliary dynamic decision-maker to record misjudgment patterns; dynamically adjust the weights and loss function of the CNN-S model based on training metrics and expert knowledge to achieve self-optimization training.

2. The method for small-sample multi-scale classification of asphalt mixture CT images based on deep reasoning decision-making according to claim 1, characterized in that, Step S1 involves constructing a mixed dataset of CT tomographic images of asphalt mixtures, which includes: S11. Select N asphalt specimens under various working conditions, and label each group with the true value labels of M CT images. S12. Use the initial CNN-S model to generate pseudo-labels for unlabeled images, and mix the real values ​​and pseudo-labels at a ratio of 1:9 to form an initial mixed training set. S13. Substitute the initial mixed training set into the model for secondary training. By strengthening contrastive learning and consistency regularization, labels with confidence > 90% are selected to construct a mixed dataset of asphalt mixture CT tomographic scan images.

3. The method for small-sample multi-scale classification of asphalt mixture CT images based on deep reasoning decision-making according to claim 2, characterized in that, The truth labels in step S11 are used to mark the voids, aggregates, asphalt mortar, and boundaries of the asphalt specimen.

4. The method for small-sample multi-scale classification of asphalt mixture CT images based on deep reasoning decision-making according to claim 1, characterized in that, Step S2 of the multi-scale dilated convolutional model CNN-S employs a two-stream feature decoder. The specific process of multi-scale classification is as follows: S21. Using the causal feature correction module, the test image is traversed through a lightweight parallel convolutional layer to obtain the confidence level at each position of the image. Based on the confidence level threshold, high and low confidence regions are divided, and the macroscopic structure region and microscopic interface distribution region of the image are defined. S22. The backbone feature extraction algorithm Encoder is used to capture the spatial distribution of each macroscopic structure in the scanned image through high dilation rate convolution of the aggregate core area, the void core area and the mortar uniform area. S23. An edge detection branch algorithm is adopted. A single-channel boundary probability distribution map is obtained through low dilation rate convolution. The Sigmoid activation function is used for mapping to obtain the edge probability of pixels at the fine interface. The complex fine interface structure features are identified through the edge probability map.

5. The method for small-sample multi-scale classification of asphalt mixture CT images based on deep reasoning decision-making according to claim 4, characterized in that, In step 23, pixels with an edge probability threshold ≥ 0.85 are determined to be valid edges.

6. The method for small-sample multi-scale classification of asphalt mixture CT images based on deep reasoning decision-making according to claim 4, characterized in that, High dilation rate convolution has a void ratio ≥ 6, while low dilation rate convolution has a void ratio ≤ 2.

7. The method for small-sample multi-scale classification of asphalt mixture CT images based on deep reasoning decision-making according to claim 1, characterized in that, Step S3, deploying the DeepSeek-R1 inference model as an auxiliary dynamic decision maker, includes: S31. Deploy the local runtime environment of the DeepSeek-R1 inference model, introduce domain knowledge and physical equation constraints to fine-tune the language model, and use the policy generation algorithm to limit the data representation of evaluation indicators and model decision commands during the training process. S32. Establish a model connection port, record the changes in the misjudgment pattern of aggregate-mortar-void in the segmentation model and the improvement trend of boundary recognition accuracy, and transmit them to the auxiliary decision model. Use linguistic logical reasoning and expert experience constraints to correct the model segmentation weights and adjust the training loss function, and correct the model's recognition accuracy of fine interfaces in real time to achieve self-optimization training of the image segmentation model.

8. The method for small-sample multi-scale classification of asphalt mixture CT images based on deep reasoning decision-making according to claim 1, characterized in that, Step S3 uses the cross-entropy function loss rate, accuracy, center-IOU, edge-IOU, and single-class block ratio generated every ten training iterations as evaluation indicators for parameter fine-tuning and loss function correction of the DeepSeek-R1 inference model through an interval update strategy.

Images