Pavement crack segmentation network automatic generation method based on neural architecture search

Abstract

This application provides an automated method for generating road crack segmentation networks based on neural architecture search. The method first constructs a customized neural architecture search space for crack segmentation tasks. Then, it employs an efficient search strategy to automatically generate a diverse and efficient set of candidate network architectures by comprehensively evaluating the differences in structural distance, computational cost, and operator sequences. Next, a composite loss function is used during training to address the sample imbalance problem, and a multi-objective function considering both accuracy and efficiency is used for evaluation. Finally, the optimal architecture is automatically selected, and training and testing are completed. This method significantly reduces the computational cost of neural architecture search and can automatically and efficiently generate lightweight crack segmentation networks with both high segmentation accuracy and low computational complexity. It effectively solves the technical challenges of existing manual design methods that rely on experience, and the high cost and poor task adaptability of general automated methods.

Description

Technical Field

[0001] This application relates to the fields of computer vision and artificial intelligence, specifically to an automatic detection and segmentation method for road surface crack images, and more particularly to an automated generation method for network architecture based on Neural Architecture Search (NAS). Background Technology

[0002] The automatic and accurate detection and segmentation of pavement cracks is of great significance for road maintenance, traffic safety assessment, and digital management of infrastructure. With the development of deep learning technology, semantic segmentation methods based on convolutional neural networks have become the mainstream technical approach in this field.

[0003] Currently, the relevant technical solutions can be mainly divided into two categories.

[0004] The first category is hand-designed networks based on expert experience, such as the widely used U-Net network and its various improved variants. These methods typically employ a predefined encoder-decoder architecture. The encoding part extracts multi-level features through convolution and pooling operations, while the decoding part restores spatial resolution and outputs the segmentation result through upsampling operations. Skip connections are often used to fuse shallow detail features with deep semantic information. Although these methods have achieved initial application results, their performance is highly dependent on the researcher's domain knowledge and requires extensive repetitive experiments. They necessitate manual adjustments to numerous hyperparameters such as network depth, width, and module types, exhibiting significant manual design bottlenecks. Specifically, their main drawbacks are twofold: first, the design process is inefficient and has high marginal costs; second, it is difficult to systematically and automatically find the optimal trade-off between model segmentation accuracy and model computational efficiency. Model segmentation accuracy is typically measured by metrics such as intersection-union ratio (IU) and recall, while model computational efficiency focuses on the number of parameters and floating-point operations. This difficulty in finding a balance often results in models that cannot directly meet the stringent lightweight requirements of practical edge device deployments.

[0005] To reduce reliance on manual design, a second approach attempts to introduce general neural architecture search techniques into crack segmentation tasks. Examples of such methods include DARTS and ENAS, whose core idea is to automatically explore network structures. However, directly applying these techniques to crack segmentation introduces new problems. First, the search process is extremely computationally expensive, typically requiring tens or even hundreds of GPU days of computing power, making it difficult to promote and apply in practical engineering projects. Second, the pre-defined search space and optimization objectives are usually designed for tasks such as general image classification, failing to fully consider the specific needs of road crack segmentation. These specificities include high sensitivity to elongated, irregular, and low-contrast targets in images, and varying degrees of dependence on multi-scale contextual information. Therefore, the network structures obtained by general neural architecture search methods are often not optimal solutions for crack segmentation tasks, exhibiting poor task adaptability.

[0006] In summary, existing technologies suffer from inherent limitations such as low automation and difficulty in balancing accuracy and efficiency when using manual network design. Meanwhile, methods that directly employ general neural architecture search suffer from excessively high search costs and insufficient task-specificity in generating networks. Therefore, there is an urgent need in this field for a technical solution that can address the specific characteristics of crack segmentation tasks, achieve efficient automated search, and directly generate specialized neural network architectures that achieve the optimal balance between recognition accuracy and computational efficiency. Summary of the Invention

[0007] In view of this, this application provides an automated generation method and circuit for road surface crack segmentation networks based on neural architecture search, in order to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial alternative.

[0008] The technical solution of this application embodiment is an automated generation method for road surface crack segmentation networks based on neural architecture search, comprising the following steps: An automated method for generating pavement crack segmentation networks based on neural architecture search includes the following steps: S1: Data preprocessing and search space construction: Obtain the road surface crack image dataset and perform preprocessing, construct a neural architecture search space for crack segmentation tasks. The search space is based on a directed acyclic graph representation and includes various basic operation sequences and inter-node connection relationships. S2: Automated generation of candidate networks based on efficient strategies: An efficient neural architecture search strategy is adopted to automatically generate multiple candidate neural network architectures in the search space, and a set of candidate architectures with diverse structures and high computational efficiency is selected based on the comprehensive difference evaluation criterion. The comprehensive difference includes at least structural distance difference, computational difference, and operator sequence difference. S3: Training and evaluation of candidate networks: Train the networks in the candidate architecture set on the training set, supervised by a composite loss function, and evaluate them on the validation set using a multi-objective evaluation function that considers both segmentation accuracy and model efficiency metrics. S4: Selection and Derivation of Optimal Network Architecture: Based on the evaluation results, select the network architecture that is optimal in terms of both accuracy and efficiency from the candidate networks and derive it; S5: Final training and testing of the optimal network: The derived optimal network architecture is retrained and tested to obtain the final deployable road crack segmentation model.

[0009] Further preferred: In step S1, each searchable module in the search space includes a start node, an end node, and at least one intermediate node. Each node is associated with an operation sequence selected from a predefined operation sequence library, which includes at least one of convolution operations, activation functions, normalization operations, and attention mechanisms.

[0010] Further preferred: In step S2, the structural distance difference is calculated by comparing the differences in the shortest path length, longest path length, and average path length from the start node to the end node in the directed acyclic graphs corresponding to the two candidate architectures; the computational complexity difference is calculated by comparing the differences in the number of floating-point operations required for the two candidate architectures to complete a single forward propagation; and the operator sequence difference is calculated by comparing the differences in the operation sequences corresponding to each node in the two candidate architectures.

[0011] Further preferred: In S2, the efficient neural architecture search strategy includes: after generating a new candidate architecture, calculating its comprehensive difference degree with each architecture in the selected architecture set, and only including the candidate architecture in the candidate architecture set when the comprehensive difference degree between the candidate architecture and all existing architectures reaches or exceeds a preset threshold.

[0012] In a further preferred embodiment, in step S3, the composite loss function is a linear combination of weighted binary cross-entropy loss and Dice loss, wherein the weight coefficient of the positive sample in the weighted binary cross-entropy loss is determined according to the ratio of the number of crack pixels to the number of background pixels.

[0013] Further preferred: In step S3, the segmentation accuracy metrics considered by the multi-objective evaluation function simultaneously include the intersection-union ratio and the F1 score, and the model efficiency metrics include the number of network parameters and the number of floating-point operations.

[0014] Further preferred: In S5, the final deployable road surface crack segmentation model is an encoder-decoder structure, wherein the encoder includes a downsampling layer and a searchable module generated by the search, the decoder includes an upsampling layer, and the encoder and decoder perform feature fusion through skip connections.

[0015] Further preferred: The searchable module includes a lightweight attention mechanism module.

[0016] Based on the same concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for automatically generating road surface crack segmentation networks based on neural architecture search as described above.

[0017] Based on the same concept, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for automatically generating road surface crack segmentation networks based on neural architecture search as described above.

[0018] Compared with the prior art, this application has the following significant advantages: By introducing a comprehensive difference-guided, efficient search strategy, redundant architecture evaluations are significantly reduced, shortening the complete neural architecture search cycle from hundreds of GPU hours to tens of GPU hours, making automated network design feasible. At the same time, by constructing a task-oriented search space and combining multi-objective evaluation, the network architecture that achieves the best balance between crack segmentation accuracy and model computation efficiency can be automatically explored and screened, directly outputting a lightweight, high-performance model that meets the needs of actual edge deployment, eliminating the tedious manual tuning and lightweight redesign process.

[0019] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of this application will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1The overall flowchart of the method for automatically generating road surface crack segmentation networks based on neural architecture search provided in the first embodiment.

[0022] Figure 2 A detailed flowchart illustrating the automated generation method for road surface crack segmentation networks based on neural architecture search provided in the second embodiment.

[0023] Figure 3 This is a schematic diagram of three representative searchable module structures automatically generated through neural architecture search in the second embodiment.

[0024] Figure 4 This is a schematic diagram of the lightweight pavement crack segmentation network architecture finally generated in the second embodiment.

[0025] Figure 5 The image shows a road surface crack image collected in the third embodiment.

[0026] Figure 6 A complete automated process diagram is provided for the third embodiment. Detailed Implementation

[0027] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of this application. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0028] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0029] First embodiment: like Figure 1 As shown, an automated method for generating road surface crack segmentation networks based on neural architecture search includes the following steps: S1: Data preprocessing and search space construction: Obtain the road surface crack image dataset and preprocess it. Construct a neural architecture search space for crack segmentation tasks. The search space is based on a directed acyclic graph representation and contains various basic operation sequences and inter-node connection relationships. Each searchable module in the search space includes a start node, an end node, and at least one intermediate node. Each node is associated with an operation sequence selected from a predefined operation sequence library, which includes at least one of convolution operations, activation functions, normalization operations, and attention mechanisms.

[0030] S2: Automated generation of candidate networks based on efficient strategies: An efficient neural architecture search strategy is adopted to automatically generate multiple candidate neural network architectures in the search space, and a set of candidate architectures with diverse structures and high computational efficiency is selected based on the comprehensive difference evaluation criterion. The comprehensive difference includes at least structural distance difference, computational difference, and operator sequence difference. Among them, the structural distance difference is calculated by comparing the differences in the shortest path length, longest path length, and average path length from the start node to the end node in the directed acyclic graphs corresponding to the two candidate architectures; the computational difference is calculated by comparing the differences in the number of floating-point operations required for the two candidate architectures to complete a single forward propagation; and the operator sequence difference is calculated by comparing the differences in the operation sequences corresponding to each node in the two candidate architectures.

[0031] An efficient neural architecture search strategy includes: after generating a new candidate architecture, calculating its overall difference with each architecture in the selected architecture set, and only including the candidate architecture in the candidate architecture set when the overall difference with all existing architectures reaches or exceeds a preset threshold.

[0032] S3: Training and evaluation of candidate networks: Train the networks in the candidate architecture set on the training set, supervised by a composite loss function, and evaluate them on the validation set using a multi-objective evaluation function that considers both segmentation accuracy and model efficiency metrics. The composite loss function is a linear combination of the weighted binary cross-entropy loss and the Dice loss. The weight coefficient of the positive sample in the weighted binary cross-entropy loss is determined according to the ratio of the number of crack pixels to the number of background pixels.

[0033] The multi-objective evaluation function considers segmentation accuracy metrics including intersection-over-union ratio (IoU) and F1 score, while model efficiency metrics include the number of network parameters and the number of floating-point operations.

[0034] S4: Selection and Derivation of Optimal Network Architecture: Based on the evaluation results, select the network architecture that is optimal in terms of both accuracy and efficiency from the candidate networks and derive it; S5: Final training and testing of the optimal network: The derived optimal network architecture is retrained and tested to obtain the final deployable road crack segmentation model.

[0035] The final deployable road crack segmentation model is an encoder-decoder structure, in which the encoder contains a downsampling layer and a searchable module generated by the search, the decoder contains an upsampling layer, and the encoder and decoder perform feature fusion through skip connections. The searchable module contains a lightweight attention mechanism module.

[0036] Second embodiment: like Figure 2The diagram shows the complete process of an automated method for generating road surface crack segmentation networks based on neural architecture search.

[0037] S1: Data Preprocessing and Search Space Construction. A dataset containing road surface images and their corresponding crack annotations is obtained, and preprocessed including standardization and augmentation. The dataset is then divided into training and validation sets. Simultaneously, a neural architecture search space is constructed for the crack segmentation task, integrating various basic operations such as convolution, pooling, activation functions, and attention mechanisms. The set Represented as basic operators, It is represented as a sequence of operators.

[0038] It employs a directed acyclic graph (DAG) and binary encoding to represent and define a flexible set of network architectures that can be explored by search algorithms. in Represented as a directed acyclic graph, These represent directed acyclic graphs respectively. The vertices and edges, .

[0039] like Figure 3 During the neural architecture search process, the system automatically explores and generates a variety of searchable modules with different structures. The three modules shown in the figure are just representative examples. These modules, as important components of the U-shaped segmentation network, are all automatically generated by the neural architecture search and play a key role in the crack segmentation task: their diverse connection methods and unit structures adapt to the feature extraction requirements of cracks at different scales and morphologies, jointly supporting the construction of a complete encoder-decoder architecture that balances expressive power, stability, and task adaptability.

[0040] The search space designed in this method directly addresses the core challenges of pavement crack segmentation: the elongation, irregularity, and multi-scale characteristics of cracks. This design ensures that the automatically generated network architecture inherently possesses a strong ability to represent these key crack properties.

[0041] To address the coexistence of detailed and global information in crack images, a multi-granularity feature channel mechanism was designed for the search space. This mechanism correlates the channel number configuration with the feature scale: small cardinality channel options (e.g., 8, 16) are provided in the detail-sensitive layer to focus on local texture; while large cardinality channel options (e.g., 128, 256) are provided in the context aggregation layer to enhance semantic integration. By decoupling the channel number search from the network layer functions, this mechanism enables the model to autonomously adapt the optimal channel topology according to the feature requirements of different levels in the crack segmentation task.

[0042] S2: Automated Generation of Candidate Networks Based on an Efficient Strategy. An efficient neural architecture search strategy is employed to automatically generate N different candidate neural network architectures within the search space constructed in step S1. The comprehensive difference between this candidate architecture and each existing architecture in the current set of selected architectures is calculated. This comprehensive difference is obtained by a weighted sum of three quantitative indicators: in, , , and These represent the overall difference, distance difference, computational difference, and operator sequence difference, respectively.

[0043] The core of the search strategy lies in introducing a comprehensive evaluation criterion that integrates distance difference, computational cost difference, and operator sequence difference.

[0044] One is the distance difference, which is calculated by comparing the shortest, longest, and average path lengths from the start node to the end node in the directed graphs within the two architectures: in, Let S represent the shortest path length, longest path length, and average path length in graph G from the starting node S to the ending node T, respectively. This represents the normalization function.

[0045] The second factor is the difference in computational complexity, calculated by comparing the number of floating-point operations required for the two architectures to complete a single round of forward propagation. in, This indicates the number of floating-point operations required for the model to complete one training cycle.

[0046] The third is the operator sequence dissimilarity, which is quantified by comparing the operation sequences corresponding to all nodes in the two architectures and calculating the weight score of their longest matching substring: in, , and Representing convolutional layers Activation function and normalization The weights; Representing operator sequence and The longest common contiguous subsequence between them This represents the weight score of the longest matching substring. The sum of the differences in the operator sequence is finally passed through Normalization is performed.

[0047] The search direction is guided by a comprehensive evaluation criterion of distance dissimilarity, computational dissimilarity, and operator sequence dissimilarity, ensuring that the generated candidate networks achieve a balance between structural diversity and computational efficiency, thereby avoiding blind and time-consuming global enumeration in a huge search space. This step aims to quickly and intelligently screen out a subset of potential architectures, rather than training all possible architectures, which is key to achieving efficiency in the search process.

[0048] S3: Training and Evaluation of Candidate Networks. The N candidate neural network architectures generated in step S2 are trained on the training set to optimize their weight parameters, and then evaluated on the validation set. To effectively address the problem of extreme imbalance between crack pixels and background pixels in crack images, a composite loss function is used for supervision during training. This function combines weighted binary cross-entropy loss and Dice loss. The former assigns higher weights to sparse crack pixels to focus on the target region, while the latter directly optimizes the overlap between the predicted and true segmented regions to improve the model's robustness to imbalanced data. The function combining weighted binary cross-entropy loss and Dice loss can be expressed as: in, This represents the total number of pixels in the image. For real labels, It is a pixel The predicted probability, This is the positive sample weighting coefficient, which is the ratio of the number of background pixels to the number of crack pixels. Let be a very small constant, denoted as Used to avoid division by zero errors. and They represent and The weights are determined. In the evaluation phase, a multi-objective function is used to quantify and score the trained candidate networks. This function considers both segmentation accuracy metrics (such as F1-score and intersection-over-union ratio) and model efficiency metrics (such as the number of network parameters and the number of floating-point operations). This step, through the targeted training and comprehensive evaluation described above, provides each candidate architecture with a comprehensive score reflecting its task performance and computational efficiency.

[0049] S4: Selection and Derivation of the Optimal Network Architecture. Based on the evaluation results obtained in step S3, the network architecture that best performs in terms of both accuracy and efficiency is automatically selected from N candidate networks. The selected optimal architecture is then finalized and derived to form a complete and specific encoder-decoder segmentation network structure.

[0050] S5: Final Training and Testing of the Optimal Network. The optimal network architecture derived in step S4, as follows... Figure 4 This structure is the final model automatically generated by the neural architecture search system described in this application. This network structure achieves an extremely lightweight design while ensuring high-precision feature extraction, fully integrating standard convolution, normalization, lightweight attention mechanisms, and multi-scale encoding / decoding paths. By intelligently combining basic operators and adaptively optimizing channel dimensions, the network constructs an efficient processing path from shallow detail perception to deep semantic fusion.

[0051] In its implementation, the network employs 3×3 convolutions combined with batch normalization and ReLU activation as stable and reliable feature extraction units. A lightweight CBAM attention module is introduced at key layers to enhance the modeling ability of crack structure continuity and spatial context. Max pooling is used for feature downsampling during the encoding stage, while transposed convolutions are used to gradually restore spatial resolution during the decoding stage, achieving multi-scale feature fusion. Furthermore, the number of channels in each layer is adaptively determined through the search process, resulting in a channel configuration of "narrow channels in shallow layers and wide channels in deep layers," significantly reducing parameter redundancy.

[0052] After generating the architecture, the network is retrained on the merged training and validation sets to optimize the weight parameters, resulting in a final deployable crack segmentation model. Subsequently, the model's performance is evaluated on an independent test set to verify its generalization ability, and the predicted segmentation results of road cracks are actually output, completing the entire closed loop from automatic architecture generation to practical model application.

[0053] Third embodiment: This embodiment provides a complete automated process, such as Figure 6 The following steps are used to generate a lightweight, high-performance pavement crack segmentation network: (1) Data Preparation and Preprocessing. To comprehensively evaluate the performance of the proposed method, this embodiment trains and tests on multiple publicly available pavement crack datasets. For each dataset, the training and testing sets are randomly divided in an 8:2 ratio. All input images are uniformly scaled to a fixed size and standardized. In addition, image data augmentation techniques such as random horizontal flipping and brightness adjustment are used during the training phase to improve the generalization ability and robustness of the generative model.

[0054] (2) Customized Neural Architecture Search Space Construction. The search space of this application is constructed based on a U-shaped encoder-decoder backbone network. This backbone network starts with a standard convolutional layer for initial feature extraction, and its main body consists of alternating downsampling layers, upsampling layers, and multiple searchable modules with undetermined internal structures. Finally, a segmentation prediction map is output through a convolutional layer. The searchable modules are the core units for automated design in this application. Each module contains a start node, a stop node, and four intermediate nodes. Each node is associated with an operation sequence selected from a predefined operation sequence library. This library provides a variety of candidate sequences, including basic convolutional units, ReLU activation function, normalized BN, and SE attention mechanism or CBAM attention mechanism. The connection relationship between nodes within a module is defined by a connection matrix and follows the forward propagation constraint, that is, a node can only be connected to its subsequent nodes. When multiple edges converge to the same node, feature fusion is performed by summing pixel values. The module structure allows for cross-layer residual connections.

[0055] (3) Efficient Architecture Search Based on Diversity Measurement. This step aims to automatically generate a set of candidate neural network architectures that are significantly different from each other from the search space described above. The core process is as follows: First, a candidate architecture defined by an operation sequence and a connection matrix is ​​randomly generated. Then, the comprehensive difference between the candidate architecture and each existing architecture in the current selected architecture set is calculated. If the comprehensive difference between the candidate architecture and any architecture in the existing architecture set is lower than a preset threshold, it is determined that its similarity is too high and is discarded, and a new candidate architecture is regenerated; if its comprehensive difference with all existing architectures reaches or exceeds the threshold, it is included in the set. This generation and screening process is repeated until a predetermined number of candidate architectures are obtained. This diversity-aware strategy effectively avoids redundant evaluation of similar architectures and is the key to achieving high-efficiency search in this method.

[0056] (4) Candidate Architecture Training, Evaluation, and Optimal Model Selection. After obtaining the candidate architecture set, each architecture is trained. To address the severe imbalance in the number of foreground and background pixels in the crack segmentation task, a hybrid loss function is used for supervision during training. This function is a linear combination of weighted binary cross-entropy loss and Dice loss. After training, the segmentation accuracy and model efficiency of each architecture are evaluated on the validation set. Finally, based on the principle of multi-objective optimization, the architecture that performs best in terms of accuracy metrics such as intersection-union ratio and F1 score, and efficiency metrics such as parameter count and computational complexity, is selected as the final generated network, denoted as LiteCrackNet.

[0057] The LiteCrackNet network generated through the automated process described above achieved leading performance in public benchmark tests. On the CrackTree206 dataset, it achieved an intersection-over-union (IoU) ratio of 94.21% and an F1 score of 96.92%; on the CFD dataset, its IoU was 86.88% and its F1 score was 92.81%. Simultaneously, the model exhibits excellent lightweight characteristics, with only 0.58 million parameters and only 3.00 FLOPs required for a single forward propagation. Furthermore, the complete neural architecture search process consumes only 17.3 GPU hours, significantly lower than similar automated methods. This demonstrates that the method in this embodiment can efficiently and automatically generate dedicated crack segmentation networks that combine high accuracy with low resource consumption.

[0058] like Figure 5 As shown in the image, the actual collected road surface crack images demonstrate that, compared to existing methods, the crack segmentation network generated in this application exhibits superior performance in various scenarios: In areas with significant interference, the method in this application can effectively suppress erroneous responses to non-crack features and accurately focus on the real crack structure; in interference-free scenarios, it can also segment crack regions more completely and coherently, especially demonstrating more refined capture capabilities and higher recall rates for small and discontinuous crack morphologies, significantly improving the completeness, accuracy, and robustness of the segmentation results.

[0059] Based on this, the automated method provided in this application, embodied in the generated LiteCrackNet model, has achieved breakthrough improvements in segmentation accuracy, model efficiency, and search cost, specifically in the following three aspects: 1. It achieves the optimal balance between accuracy and lightweight design, directly generating high-performance, deployable models.

[0060] This application fundamentally solves the problem that manual design and existing automated methods struggle to achieve the optimal balance between accuracy and efficiency. As shown in the comprehensive comparison in Table 1, the generated LiteCrackNet model achieves the highest IoU (94.21%) and F1 score (96.92%) among all compared methods on the CrackTree206 dataset. Simultaneously, this model is extremely lightweight, with only 0.58 M parameters and a computational cost of only 3.00 GFLOPs. Especially when compared with similar NAS models, LiteCrackNet significantly outperforms models with fewer parameters in accuracy (e.g., its IoU on CrackTree206 exceeds AutoCrackNet by approximately 2.47 percentage points), demonstrating that this application can automatically discover the optimal architecture that balances top-level accuracy and ultimate efficiency, eliminating tedious manual trade-offs and lightweighting steps.

[0061] 2. Improve the efficiency of neural architecture search to a practical level and break through the bottleneck of computational cost.

[0062] This application revolutionizes the time-consuming nature of automated design through an innovative and efficient search strategy. As shown in Table 1, existing NAS methods (such as AutoCrackNet) require over 120 GPU hours for searching. In contrast, this application reduces the entire search cycle to 17.3 GPU hours, achieving an efficiency improvement of approximately 7 times. This order-of-magnitude improvement enables the automated design of task-specific networks to be completed within a day, removing a key obstacle to the practical application and continuous iteration of the technology.

[0063] 3. It provides an efficient, high-performance end-to-end automation solution, significantly reducing the technical threshold.

[0064] In summary, this application integrates the advantages of both aspects, forming a complete process from data input to optimal model output. Users do not need extensive experience in architecture design, nor do they need to manually fine-tune multiple objectives such as accuracy, speed, and size. They can obtain a model as accurate and lightweight as LiteCrackNet in a very short time through this application. This greatly reduces the R&D cycle, manpower costs, and technical barriers in the field of intelligent pavement crack detection, and promotes the standardization and large-scale application of advanced deep learning technology in this field.

[0065] Table 1. Quantitative analysis of the method in this embodiment and existing algorithms on the CrackTree206 dataset (- for hand-designed neural networks, no search time).

[0066] Based on the same concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it realizes the above-mentioned automated generation of a road surface crack segmentation network based on neural architecture search. become method.

[0067] Based on the same concept, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for automatically generating road surface crack segmentation networks based on neural architecture search.

[0068] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An automated method for generating pavement crack segmentation networks based on neural architecture search, characterized in that, Includes the following steps: S1: Data preprocessing and search space construction: Obtain the road surface crack image dataset and perform preprocessing, construct a neural architecture search space for crack segmentation tasks. The search space is based on a directed acyclic graph representation and includes various basic operation sequences and inter-node connection relationships. S2: Automated generation of candidate networks based on efficient strategies: An efficient neural architecture search strategy is adopted to automatically generate multiple candidate neural network architectures in the search space, and a set of candidate architectures with diverse structures and high computational efficiency is selected based on the comprehensive difference evaluation criterion. The comprehensive difference includes at least structural distance difference, computational difference, and operator sequence difference. S3: Training and evaluation of candidate networks: Train the networks in the candidate architecture set on the training set, supervised by a composite loss function, and evaluate them on the validation set using a multi-objective evaluation function that considers both segmentation accuracy and model efficiency metrics. S4: Selection and Derivation of Optimal Network Architecture: Based on the evaluation results, select the network architecture that is optimal in terms of both accuracy and efficiency from the candidate networks and derive it; S5: Final training and testing of the optimal network: The derived optimal network architecture is retrained and tested to obtain the final deployable road crack segmentation model.

2. The method for automatically generating road surface crack segmentation networks based on neural architecture search according to claim 1, characterized in that, In step S1, each searchable module in the search space includes a start node, an end node, and at least one intermediate node. Each node is associated with an operation sequence selected from a predefined operation sequence library, which includes at least one of convolution operations, activation functions, normalization operations, and attention mechanisms.

3. The method for automatically generating road surface crack segmentation networks based on neural architecture search according to claim 2, characterized in that, In S2, the structural distance difference is calculated by comparing the differences in the shortest path length, longest path length, and average path length from the start node to the end node in the directed acyclic graphs corresponding to the two candidate architectures; the computational complexity difference is calculated by comparing the differences in the number of floating-point operations required for the two candidate architectures to complete a single forward propagation; and the operator sequence difference is calculated by comparing the differences in the operation sequences corresponding to each node in the two candidate architectures.

4. The method for automatically generating road surface crack segmentation networks based on neural architecture search according to claim 1, characterized in that, In S2, the efficient neural architecture search strategy includes: after generating a new candidate architecture, calculating its comprehensive difference degree with each architecture in the selected architecture set, and only including the candidate architecture in the candidate architecture set when the comprehensive difference degree between the candidate architecture and all existing architectures reaches or exceeds a preset threshold.

5. The method for automatically generating road surface crack segmentation networks based on neural architecture search according to claim 1, characterized in that, In S3, the composite loss function is a linear combination of weighted binary cross-entropy loss and Dice loss, wherein the weight coefficient of positive samples in the weighted binary cross-entropy loss is determined according to the ratio of the number of crack pixels to the number of background pixels.

6. The method for automatically generating road surface crack segmentation networks based on neural architecture search according to claim 1, characterized in that, In S3, the segmentation accuracy metrics considered by the multi-objective evaluation function simultaneously include the intersection-union ratio and the F1 score, and the model efficiency metrics include the number of network parameters and the number of floating-point operations.

7. The method for automatically generating road surface crack segmentation networks based on neural architecture search according to claim 1, characterized in that, In S5, the final deployable road crack segmentation model is an encoder-decoder structure, wherein the encoder includes a downsampling layer and a searchable module generated by the search, the decoder includes an upsampling layer, and the encoder and decoder perform feature fusion through skip connections.

8. The method for automatically generating road surface crack segmentation networks based on neural architecture search according to claim 7, characterized in that, The searchable module includes a lightweight attention mechanism module.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for automatically generating road surface crack segmentation networks based on neural architecture search as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for automatically generating road surface crack segmentation networks based on neural architecture search as described in any one of claims 1 to 8.

Images