Real-time road crack segmentation method and system based on three-branch collaborative architecture

Abstract

The application discloses a real-time pavement crack segmentation method and system based on a three-branch collaborative architecture, and relates to the technical field of computer vision. A to-be-segmented pavement image is acquired and input into a TriCrackNet segmentation model, multi-scale semantic features are extracted through an EASPP module of a semantic branch, spatial detail features and edge texture features are respectively extracted by a spatial branch and a boundary branch; semantic global information is used to supplement and locally adaptively enhance the spatial and boundary features by means of an EFEM module; adaptive weight fusion is performed on the three types of features by means of an EFFM module, and comprehensive fusion features are obtained; and after the fusion features are processed by upsampling, a pavement crack segmentation result is output by a segmentation head. The system corresponds to an image acquisition module, a TriCrackNet segmentation model, a loss calculation module and a result output module. The application considers the integrity of crack segmentation, edge sharpness and real-time inference efficiency by means of three-branch collaboration and lightweight module design, and is suitable for engineering actual deployment requirements.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a real-time road surface crack segmentation method and system based on a three-branch collaborative architecture. Background Technology

[0002] As an early and typical manifestation of road structural deterioration, accurate detection and quantitative segmentation of road cracks are core technical prerequisites for ensuring highway operational safety and optimizing maintenance resource allocation. With the advancement of intelligent transportation and digital infrastructure construction, traditional methods such as manual inspection, threshold segmentation, and edge detection are no longer adequate for the large-scale, high-frequency, and high-precision road defect detection needs. Semantic segmentation technology based on deep learning, with its end-to-end feature extraction and mask output capabilities, is gradually becoming the mainstream technical approach for road crack detection. This type of method directly learns the feature differences between cracks and the background from the original road surface image through deep neural networks, enabling automatic crack identification and pixel-level segmentation in complex scenarios, demonstrating significant advantages in both detection efficiency and accuracy. However, practical engineering applications pose a dual constraint on pavement crack segmentation technology: on the one hand, crack morphology exhibits characteristics such as multi-scale, weak texture, and slender extension, and is easily affected by noise such as uneven lighting, dirt interference, and pavement wear, requiring the model to have strong feature representation and anti-interference capabilities; on the other hand, deployment scenarios such as vehicle-mounted detection equipment and edge computing terminals have strict limitations on inference speed and computing power consumption, requiring the model to achieve real-time inference while ensuring segmentation accuracy. Current mainstream segmentation methods often struggle to achieve the optimal balance between accuracy and speed, becoming a key bottleneck restricting the implementation of the technology.

[0003] To resolve the conflict between accuracy and real-time performance, researchers have gradually shifted from single-path encoder-decoder architectures to multi-branch collaborative network architectures, improving efficiency through feature decoupling and hierarchical processing. Multi-branch architectures typically assign features from different levels to independent branches for targeted extraction. Higher-level branches capture global semantic information to distinguish cracks from the background, while lower-level branches focus on preserving spatial details and edge textures. Finally, a feature fusion module integrates the outputs from multiple branches, compressing computation while retaining key features. This architecture alleviates the efficiency problem of single-path models to some extent, but it still has significant limitations for specific targets like road cracks: First, the multi-scale feature extraction module design is redundant, and traditional structures such as Spatial Pyramid Pooling (ASPP) have high computational costs, making them difficult to adapt to lightweight requirements. Second, the feature enhancement and fusion mechanisms do not fully incorporate the morphological characteristics of cracks, resulting in inefficient interaction between low-level spatial features and high-level semantic features, easily leading to missed detections of small cracks or blurred edges. Third, the loss function design lacks constraints on crack edges and spatial continuity, resulting in segmentation results that are prone to breakage and incompleteness, failing to meet the requirements of engineering quantitative analysis.

[0004] Chinese patent (publication number: CN116052105A) proposes a method for identifying, classifying, and calculating the area of ​​road cracks. This method uses crack images captured by a road inspection vehicle as a dataset. First, it trains an object detection algorithm to complete crack region cropping. Then, it performs semantic segmentation and annotation on the cropped images and trains the model. Finally, it uses a region growing algorithm to denoise the segmentation results and synthesize a binary image of the original size, achieving accurate calculation of the crack area. This effectively avoids interference from non-crack areas in the denoising process and improves the accuracy of crack quantification analysis. However, this method relies on a two-stage process of object detection and semantic segmentation, which significantly increases inference latency. Furthermore, the cropping operation may lose contextual information about the cracks, making it difficult to meet the needs of real-time scenarios such as vehicle-mounted inspection. Chinese patent (publication number: CN117315261A) proposes a road crack defect identification and segmentation method based on a UNet network. By constructing a symmetric encoder-decoder structure and introducing residual dilated convolutional blocks with different dilation rates and MobileOne efficient convolutional blocks to optimize feature extraction, it achieves high-precision crack identification and defect quantification calculation, solving the problems of excessive manual intervention and low efficiency in traditional methods. However, this method still uses a single-path encoding-decoding architecture, which has limited ability to capture crack edges and spatial details, and the model is not lightweight enough. It is difficult to achieve real-time inference on low-computing embedded devices and cannot meet the deployment requirements of edge computing scenarios.

[0005] Therefore, there is an urgent need for a three-branch collaborative architecture method for real-time pavement crack segmentation. By decoupling the extraction processes of semantic, spatial, and edge features, a lightweight multi-scale feature extraction, feature enhancement, and fusion module is designed. While ensuring the integrity of crack details and edge clarity, the computational load and parameter count of the model are significantly reduced. This constructs a technical solution that balances high-precision segmentation and real-time inference, addressing the core shortcomings of existing methods in terms of accuracy-speed balance, robustness in complex scenarios, and adaptability to edge devices. This will provide efficient and reliable technical support for smart highway maintenance. Summary of the Invention

[0006] To address the aforementioned technical problems, this application discloses a real-time pavement crack segmentation method and system based on a three-branch collaborative architecture; the real-time pavement crack segmentation method based on the three-branch collaborative architecture includes:

[0007] Obtain the road surface image to be segmented, and input the road surface image into the TriCrackNet segmentation model built based on a three-branch collaborative architecture;

[0008] The deep semantic features of the road surface image are extracted at multiple scales through the Efficient Hollow Spatial Pyramid Pooling (EASPP) module in the semantic branch of the TriCrackNet segmentation model, resulting in multi-scale semantic features.

[0009] Spatial detail features of the road surface image are extracted through the spatial branch of the TriCrackNet segmentation model, and crack edge texture features of the road surface image are extracted through the boundary branch of the TriCrackNet segmentation model, thus obtaining spatial detail features and edge texture features.

[0010] The efficient feature enhancement module EFEM utilizes the global information of the multi-scale semantic features to perform semantic supplementation and local feature adaptive enhancement on the spatial detail features and edge texture features, respectively, to obtain enhanced spatial features and enhanced boundary features.

[0011] The multi-scale semantic features, enhanced spatial features, and enhanced boundary features are adaptively weighted and fused using the efficient feature fusion module EFFM to obtain fused features.

[0012] Based on the fusion features and after upsampling, the final segmentation result of the road surface cracks is output through the segmentation head;

[0013] The loss function of the TriCrackNet segmentation model is a multi-loss weighted fusion function, including boundary branch loss, spatial branch loss, semantic branch multi-scale loss, and final segmentation mask loss.

[0014] Preferably, the TriCrackNet segmentation model is trained by performing multiple types of data augmentation operations on the road surface image, including geometric transformations and pixel-level transformations.

[0015] The enhanced data is divided into training, validation and test sets according to a preset ratio, and the image pixel values ​​are normalized to a preset range to improve the model’s generalization ability and training stability.

[0016] Preferably, the efficient void space pyramid pooling (EASPP) module extracts multi-scale semantic features, including:

[0017] Input features after the last downsampling of the semantic branch Features are divided into multiple groups along the channel dimension;

[0018] Progressive dilated convolution and global average pooling are performed sequentially on the multiple sets of features. After each set of features is extracted, a residual connection is made with the next set of features to serve as the input for the next set of features.

[0019] The convolutional output features of each group are concatenated with the features upsampled by global average pooling, and then fused with the residuals of the original input features via a 1×1 convolution to output multi-scale semantic features. The specific calculation formula is as follows:

[0020]

[0021] in, The dilation rate of the dilated convolution is... Indicates global average pooling. Indicates an upsampling operation. Indicates the kernel size as The convolution operation.

[0022] Preferably, the spatial branching extracts spatial detail features, including:

[0023] The TriCrackNet segmentation model uses the output features of the Stem layer as input and performs multiple convolution operations to maintain the feature dimension unchanged.

[0024] The efficient feature enhancement module EFEM is inserted after the preset convolutional block to guide spatial feature extraction using global information from semantic branches;

[0025] During the spatial feature upsampling process, the Stem layer features and spatial features are connected by addition to supplement spatial detail information and obtain spatial detail features.

[0026] Preferably, the boundary branch extracts edge texture features, including:

[0027] The TriCrackNet segmentation model uses the output features of the Stem layer as input and performs multiple convolution operations to maintain the feature dimension unchanged.

[0028] The efficient feature enhancement module EFEM is inserted to enhance feature representation. A boundary header is set after EFEM to output an edge map. Edge texture features are extracted by constraining the boundary loss function.

[0029] Preferably, the efficient feature enhancement module EFEM performs feature enhancement, including:

[0030] Global average pooling is performed on the low-level features and high-level features respectively to obtain the corresponding global features; the low-level features include spatial branch or boundary branch features, and the high-level features are semantic branch features;

[0031] Semantic weights are obtained by performing 1×1 convolution and Sigmoid activation on high-level global features;

[0032] The lower-level global features are supplemented by the semantic weights to obtain enhanced global features.

[0033] The difference between low-level features and corresponding global features is used to obtain enhanced local features through 1×1 convolution and ReLU² activation;

[0034] The enhanced global features are added to the enhanced local features to output the enhanced features. The calculation formula is:

[0035]

[0036] in, To enhance the global features, Low-level features These are low-level global features.

[0037] Preferably, the efficient feature fusion module EFFM performs feature fusion, including:

[0038] The enhanced boundary features are convolved with 3×3 to obtain boundary weights, which are then broadcast to the spatial and semantic branches to obtain enhanced spatial and semantic features.

[0039] The enhanced spatial features and enhanced semantic features are subjected to mean pooling and max pooling along the channel dimension, respectively, and then concatenated to obtain the corresponding representative features.

[0040] Matrix operations are performed on the enhanced spatial features, enhanced semantic features, and boundary weights, and the feature relation matrix is ​​obtained by convolution and sigmoid activation.

[0041] The fusion features are obtained through adaptive weighted fusion. The calculation formula is:

[0042]

[0043] in, To enhance post-semantic features, To enhance the post-spatial features, This is the characteristic relation matrix.

[0044] Preferably, the calculation formula for the multi-loss weighted fusion function is as follows:

[0045]

[0046] in, The binary cross-entropy loss is for the boundary branch. The cross-entropy loss of the spatial branch, , For the multi-scale cross-entropy loss of semantic branches, , The cross-entropy loss is the final segmentation mask; each loss term corresponds to a preset weight coefficient.

[0047] Preferably, in the calculation of the final segmentation mask loss term, the edge map output by the boundary branch is used as a weighting factor and multiplied element-wise with the segmentation mask. The enhanced mask result and the real label are then used to calculate the loss.

[0048] A real-time pavement crack segmentation system based on a three-branch collaborative architecture includes:

[0049] The image acquisition module is used to acquire the road surface image to be segmented and to preprocess the road surface image;

[0050] The TriCrackNet segmentation model includes semantic branches, spatial branches, boundary branches, the efficient feature enhancement module EFEM, the efficient feature fusion module EFFM, and a segmentation head;

[0051] The semantic branch integrates the efficient void space pyramid pooling (EASPP) module to extract multi-scale semantic features from road surface images.

[0052] The spatial branch extracts spatial detail features from the road surface image, and the boundary branch is used to extract crack edge texture features from the road surface image.

[0053] The efficient feature enhancement module EFEM utilizes global information from multi-scale semantic features to semantically supplement and adaptively enhance spatial detail features and edge texture features.

[0054] The efficient feature fusion module EFFM performs adaptive weighted fusion of multi-scale semantic features, enhanced spatial features, and enhanced boundary features.

[0055] The segmentation head upsamples the fused features and outputs the segmentation results of the road surface cracks.

[0056] The loss calculation module calculates the loss value during model training using a multi-loss weighted fusion function to guide model parameter optimization.

[0057] The results output module outputs and displays the final segmentation results of road surface cracks.

[0058] Compared with the prior art, the technical solution of this application has the following technical effects:

[0059] This invention achieves decoupled extraction and fusion of pavement crack features through a three-branch collaborative architecture, which can accurately capture multi-scale semantic information, spatial details and edge textures of cracks, ensuring the integrity and clarity of the segmentation results, and providing a reliable basis for the quantitative analysis of pavement defects and maintenance decisions.

[0060] This invention employs a lightweight multi-scale feature extraction, feature enhancement, and fusion module, which significantly reduces the amount of computation and parameters while ensuring feature representation capabilities. This enables the model to adapt to low-computing-power edge devices and vehicle-mounted detection scenarios, achieving efficient and real-time road crack segmentation and inference.

[0061] This invention constrains model training through a multi-loss weighted fusion function, strengthens the learning of crack edges and spatial details, improves the model's robustness to complex lighting and noise interference scenarios, reduces the problem of missing small cracks and edge blurring, and stably outputs high-quality segmentation masks.

[0062] The end-to-end segmentation process of this invention simplifies the operation steps of crack detection, eliminating the need for additional preprocessing or post-processing. It can directly output segmentation results from the original pavement image, improving the automation level and engineering implementation efficiency of pavement crack detection, and facilitating large-scale promotion and application.

[0063] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0064] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0065] 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0066] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0067] Figure 1 Flowchart of the steps in a real-time pavement crack segmentation method based on a three-branch collaborative architecture;

[0068] Figure 2 : A schematic diagram of the overall architecture of the TriCrackNet segmentation model and the connections between its branches and modules;

[0069] Figure 3Schematic diagram of the internal structure and feature processing of the Efficient Void Space Pyramid Pooling (EASPP) module;

[0070] Figure 4 A schematic diagram of the dual-path feature enhancement process of the High-Efficiency Feature Enhancement Module (EFEM);

[0071] Figure 5 : Schematic diagram of the adaptive weight fusion structure of the efficient feature fusion module (EFFM);

[0072] Figure 6 : Schematic diagram of the module composition of a real-time pavement crack segmentation system based on a three-branch collaborative architecture;

[0073] Figure 7 Visual comparison of TriCrackNet and existing state-of-the-art methods on the Crack500 dataset;

[0074] Figure 8 Visual comparison of TriCrackNet and existing state-of-the-art methods on the DeepCrack dataset;

[0075] Figure 9 Visual comparison of TriCrackNet and existing state-of-the-art methods on the CrackForest dataset;

[0076] Figure 10 Visual comparison of TriCrackNet's generalization experiments on the DeepCrack dataset. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0078] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0079] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0080] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0081] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0082] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0083] Example 1

[0084] This embodiment mainly describes a real-time pavement crack segmentation method based on a three-branch collaborative architecture, such as... Figure 1 As shown, it specifically includes:

[0085] Obtain the road surface image to be segmented, and input the road surface image into the TriCrackNet segmentation model built based on a three-branch collaborative architecture;

[0086] The deep semantic features of the road surface image are extracted at multiple scales through the Efficient Hollow Spatial Pyramid Pooling (EASPP) module in the semantic branch of the TriCrackNet segmentation model, resulting in multi-scale semantic features.

[0087] Spatial detail features of the road surface image are extracted through the spatial branch of the TriCrackNet segmentation model, and crack edge texture features of the road surface image are extracted through the boundary branch of the TriCrackNet segmentation model, thus obtaining spatial detail features and edge texture features.

[0088] The efficient feature enhancement module EFEM utilizes the global information of the multi-scale semantic features to perform semantic supplementation and local feature adaptive enhancement on the spatial detail features and edge texture features, respectively, to obtain enhanced spatial features and enhanced boundary features.

[0089] The multi-scale semantic features, enhanced spatial features, and enhanced boundary features are adaptively weighted and fused using the efficient feature fusion module EFFM to obtain fused features.

[0090] Based on the fusion features and after upsampling, the final segmentation result of the road surface cracks is output through the segmentation head;

[0091] The loss function of the TriCrackNet segmentation model is a multi-loss weighted fusion function, including boundary branch loss, spatial branch loss, semantic branch multi-scale loss, and final segmentation mask loss.

[0092] Furthermore, the TriCrackNet segmentation model adopts an end-to-end three-branch collaborative architecture, which follows a complete process of Stem layer - branch feature extraction - feature fusion - upsampling - segmentation output. Its architecture is as follows: Figure 2 As shown, after the road surface image to be segmented is input into the model, a standardization preprocessing is first performed to map the image pixel values ​​to a preset range. Then, it is adaptively scaled proportionally according to the model's preset input size to maintain the aspect ratio of the crack target. Zero-value padding is performed on edge regions with insufficient size to avoid feature extraction deviations caused by image distortion. The preprocessed image is input to the Stem layer, which consists of continuous convolutional blocks and pooling layers. It is used to complete the initial feature extraction and output a basic feature map with uniform dimensions. This feature map will serve as the common input for the semantic branch, spatial branch, and boundary branch. During the training phase, the three branches retain auxiliary training links to enhance feature learning. During the inference phase, all auxiliary heads are removed, and only the core feature extraction and fusion links are retained to reduce computation and improve inference speed.

[0093] Furthermore, the semantic branch takes the base feature map output from the Stem layer as input and cascades it with residual convolutional blocks (CBs) through multiple downsampling operations to progressively compress the spatial size of the feature map and increase the number of feature channels, thereby extracting deep semantic information about road surface cracks. During downsampling, a convolutional operation with a stride of 2 is used to achieve spatial dimension compression, while residual connections ensure effective gradient propagation, avoiding the gradient vanishing problem common in deep networks. After the final downsampling, an efficient hollow spatial pyramid pooling (EASPP) module is embedded, the structure of which is as follows: Figure 3 As shown, the core is used to capture multi-scale features of cracks. Specifically, the EASPP module first processes the input feature map... Features are divided into four groups along the channel dimension. ,in The number of input feature channels, and These represent the height and width of the feature maps, respectively. Progressive dilated convolutions are performed sequentially on the four feature maps, with the dilation rate of each convolution being... The receptive fields are set to 6, 12, and 18 respectively to achieve multi-scale feature capture by progressively expanding the receptive field. Each set of extracted features is then residually concatenated with the next set, allowing the next set of features to be fused with preceding feature information, thus improving feature utilization. Simultaneously, a global average pooling (GAP) operation is performed on the feature map. The pooling result is upsampled to the same size as the convolutional output features and then concatenated with each set of convolutional output features along the channel dimension. A 1×1 convolution is then performed to compress the number of channels, and finally, residual fusion is performed with the original input feature map to output multi-scale semantic features. The calculation formula is as follows:

[0094]

[0095] in, Indicates the kernel size as Convolution operation, This represents a bilinear interpolation upsampling operation. The semantic branch constructs a lightweight U-shaped structure during the upsampling process, enabling the reuse of features at different levels through residual connections. At the same time, auxiliary segmentation heads are set at the last two residual connections, and the branch's sensitivity to crack features at different scales is enhanced through multi-scale loss constraints.

[0096] Furthermore, the spatial branch also uses the base feature map output by the Stem layer as input, maintaining the spatial dimension of the feature map throughout the feature extraction process. It refines the feature representation through multiple convolutional block (CB) operations, avoiding the loss of spatial details caused by downsampling. For example... Figure 2 As shown, an efficient feature enhancement module (EFEM) is inserted after the second convolutional block of the spatial branch. This module receives high-level semantic features from the semantic branch as guidance signals to enhance the intermediate features of the spatial branch, strengthening spatial detail information related to cracks and suppressing background noise interference. During the spatial feature upsampling process, element-wise addition is used to achieve skip connections between the basic features of the Stem layer and the features of the spatial branch, directly supplementing the underlying spatial information. This ensures that the fine location and morphological features of cracks are not weakened, and the final output accurately reflects the spatial detail features of crack spatial distribution.

[0097] Furthermore, the boundary branch and the spatial branch adopt the same basic structure, using the basic feature map output from the Stem layer as input, and maintaining the feature dimension unchanged through multiple convolutional block (CB) operations, focusing on the extraction of crack edge texture information. For example... Figure 2As shown, an EFEM module is also inserted into the boundary branch. This module supplements the edge features with global semantic features from the semantic branch, improving edge feature discriminability and reducing interference from background noise such as road stains and wear on edge extraction. A boundary header is then set after the EFEM module. This header consists of a 1×1 convolution and a sigmoid activation function, with an output dimension of [missing information]. The edge map, the edge Figure 1 On the one hand, it is used to calculate the boundary loss to constrain branch training; on the other hand, it provides boundary weight information for subsequent feature fusion and final loss calculation. Finally, through the constraint of the boundary loss function, clear and complete crack edge texture features are output.

[0098] Furthermore, the structure of the efficient feature enhancement module (EFEM) is as follows: Figure 4 As shown, the core is used to enhance the interaction between low-level features (spatial branch or boundary branch features) and high-level features (semantic branch features), avoiding computational redundancy caused by complex attention mechanisms. For the input low-level features... and high-level features First, perform global average pooling (GAP) on both to obtain the corresponding global features. and ,Right now , ,in This represents a global average pooling operation, used to compress feature maps to a channel-level global statistic. For high-level global features... Perform a 1×1 convolution to compress the number of channels, and then use a Sigmoid activation function to generate semantic weights. The calculation formula is: This weight is used to capture the correlation between high-level semantics and low-level features, highlighting semantic information related to cracks. Semantic weights are then applied. With low-level global features Element-by-element multiplication, then summation Perform residual connections to obtain enhanced global features. This allows semantic information to supplement global features. Simultaneously, it computes low-level features. and The difference is used to separate local features. A 1×1 convolution and ReLU² activation operation are performed on this difference feature to enhance local details related to the crack and suppress noise, resulting in enhanced local features. Finally, the enhanced global features are... The enhanced local features are added element-wise to output the enhanced features. The calculation formula is: ,in, To enhance the global features, Low-level features These are low-level global features.

[0099] Furthermore, the structure of the efficient feature fusion module (EFFM) is as follows: Figure 5 As shown, this method enables efficient adaptive fusion of multi-scale semantic features, enhanced spatial features, and enhanced boundary features, avoiding feature redundancy and information loss caused by traditional concatenation or addition operations. Let the enhanced spatial features be... Enhanced boundary features are Multi-scale semantic features are First, the boundary features Perform a 3×3 convolution operation, with an output dimension of Boundary weights Through broadcasting mechanism By adding element-wise features to both spatial and semantic branches, edge texture information is supplemented, resulting in enhanced spatial features. and enhanced semantic features .right and Perform mean pooling and max pooling operations along the channel dimension respectively, and concatenate the pooling results along the channel dimension to obtain their respective representative features. and ,in , , This indicates a channel-dimensional concatenation operation. , and Perform matrix multiplication, followed by 3×3 convolution and sigmoid activation function to generate a three-branch feature relation matrix. The calculation formula is: ,in This represents the matrix multiplication operation. Finally, it ends with... and As adaptive weights, respectively with and The elements are multiplied one by one and then added together to output the final fused feature. The calculation formula is: ,in, To enhance post-semantic features, To enhance the post-spatial features, Characteristic relation matrix

[0100] Furthermore, fusion features The image is upsampled 8 times to restore the same spatial dimensions as the original road surface image. The upsampling process uses bilinear interpolation to ensure a smooth transition of the feature map and avoid the checkerboard effect. The upsampled feature map is then input into a segmentation head, which consists of a 1×1 convolution and a Softmax activation function. The convolution operation maps the number of feature channels to a binary classification dimension (cracks and background), outputting a pixel-level segmentation mask. In the mask, regions with a value of 1 correspond to crack targets, and regions with a value of 0 correspond to background regions. This is used as the final segmentation result for the road surface cracks.

[0101] Furthermore, the loss function of the TriCrackNet segmentation model consists of a weighted average of six loss terms, used to comprehensively constrain the accuracy of feature extraction in each branch and the final segmentation result. The calculation formula is as follows: ;in, The binary cross-entropy loss for the boundary branch is used to constrain the difference between the edge map and the true edge labels extracted by the Canny operator; The cross-entropy loss of the spatial branch is used to optimize the extraction accuracy of spatial detail features; , The multi-scale cross-entropy loss of the semantic branch corresponds to the output of the auxiliary segmentation head at the last two residual connections of the semantic branch, and the multi-scale constraint improves the multi-scale adaptability of semantic features. , The cross-entropy loss for the final segmentation mask is used to directly optimize the consistency between the segmentation result and the ground truth (GT). The weights of each loss term are set as follows: , , , ,in During the calculation process, the edge map output by the boundary branch needs to be multiplied element by element with the segmentation mask as a weight factor. The loss calculation is then performed on the enhanced mask result and the real label to further enhance the learning of crack boundary information.

[0102] This implementation details how the decoupled extraction and collaborative fusion of semantic, spatial, and boundary branches accurately captures multi-scale semantic, spatial details, and edge texture features of cracks. Combined with three lightweight modules—EASPP, EFEM, and EFFM—it enhances feature representation while avoiding computational redundancy, significantly improving the completeness and edge clarity of crack segmentation in complex scenarios. The multi-loss weighted fusion function further optimizes the model's learning direction through hierarchical constraints, enabling the method to maintain high segmentation accuracy while keeping the inference speed efficient and real-time, thus adapting to the deployment needs of low-computing-power scenarios such as vehicle detection and edge computing.

[0103] Example 2 describes in detail a real-time pavement crack segmentation system based on a three-branch collaborative architecture, used to implement the aforementioned real-time pavement crack segmentation method based on a three-branch collaborative architecture, such as... Figure 6 As shown, it includes:

[0104] The image acquisition module is used to acquire the road surface image to be segmented and to preprocess the road surface image;

[0105] The TriCrackNet segmentation model includes semantic branches, spatial branches, boundary branches, the efficient feature enhancement module EFEM, the efficient feature fusion module EFFM, and a segmentation head;

[0106] The semantic branch integrates the efficient void space pyramid pooling (EASPP) module to extract multi-scale semantic features from road surface images.

[0107] The spatial branch extracts spatial detail features from the road surface image, and the boundary branch is used to extract crack edge texture features from the road surface image.

[0108] The efficient feature enhancement module EFEM utilizes global information from multi-scale semantic features to semantically supplement and adaptively enhance spatial detail features and edge texture features.

[0109] The efficient feature fusion module EFFM performs adaptive weighted fusion of multi-scale semantic features, enhanced spatial features, and enhanced boundary features.

[0110] The segmentation head upsamples the fused features and outputs the segmentation results of the road surface cracks.

[0111] The loss calculation module calculates the loss value during model training using a multi-loss weighted fusion function to guide model parameter optimization.

[0112] The results output module outputs and displays the final segmentation results of road surface cracks.

[0113] Furthermore, the image acquisition module supports multiple data source inputs, including real-time road surface video streams captured by vehicle-mounted cameras, drone aerial images, and stored road surface image files. The module has a built-in image format parsing unit, compatible with mainstream image formats such as JPG, PNG, and TIFF. The preprocessing process includes pixel value standardization (mapping pixel values ​​to the [0,1] interval), adaptive size adjustment (scaling proportionally to the TriCrackNet model's preset input size and filling zero values ​​at the edges), and preliminary noise suppression (using Gaussian filtering to remove high-frequency noise), ensuring that the input image meets the model's feature extraction requirements. The result output module has multiple result display and storage functions, and can output pixel-level segmentation masks and crack area annotation maps in real time (overlaying crack outlines on the original image). It also supports the output of quantitative data of the segmentation results (crack length, area, etc.), with storage formats including image files (PNG / TIFF) and structured data files (JSON / CSV), adapting to the data interface requirements of highway maintenance management systems.

[0114] Furthermore, the TriCrackNet segmentation model is the core computational unit of the system, achieving accurate feature extraction and efficient fusion through a three-branch collaborative design.

[0115] The semantic branch, as the core of high-level feature extraction, consists of multi-level downsampling convolutional blocks (CBs), residual connections, and the EASPP module. The downsampling process compresses the feature map spatial dimension through 3×3 convolutions with a stride of 2, simultaneously increasing the number of feature channels to enhance semantic expressiveness. Residual connections run through each downsampling level, ensuring the effectiveness of gradient backpropagation. The EASPP module is embedded after the last downsampling step. Its core function is to efficiently capture multi-scale crack features: the module first divides the input feature map into four groups along the channel dimension, and then performs progressively dilated convolutions with dilation rates of 6, 12, and 18 on each group sequentially. The output of each convolution is fused with the feature residual of the next group. Simultaneously, it combines global average pooling (GAP) upsampling features, compresses them through 1×1 convolution channels, and then connects them with the original input feature residuals, ultimately outputting multi-scale semantic features. The calculation formula strictly follows the definition in the claims, avoiding computational redundancy through feature grouping and hierarchical residual fusion.

[0116] The spatial and boundary branches employ a lightweight parallel structure, both taking the base feature map output from the Stem layer as input and maintaining the feature dimension unchanged throughout to preserve detailed information. The spatial branch extracts initial spatial features through two convolutional blocks (CBs) and then embeds an EFEM module to receive high-level semantic guidance from the semantic branch. During the upsampling stage, additive skip connections are used to fuse the base features from the Stem layer, supplementing the underlying spatial details. The boundary branch has the same structure as the spatial branch. The features enhanced by the EFEM module are output as an edge map via a boundary header (1×1 convolution + sigmoid activation), which serves both as a loss constraint and as boundary weights for subsequent feature fusion. Both branches interact with the semantic branch through the EFEM module, focusing on the spatial distribution of cracks and edge texture extraction, respectively.

[0117] Furthermore, the efficient feature enhancement module EFEM is a key unit connecting high-level semantic features and low-level detail features. Its structure employs a dual-path design of global semantic supplementation and local detail enhancement. The module input includes low-level features (spatial branch or boundary branch features) and high-level features (semantic branch features), and global average pooling is used to extract the global features of both. and ;right Semantic weights are generated through 1×1 convolution and Sigmoid activation. ,and Weighted fusion followed by residual connections yields enhanced global features. Simultaneously calculate low-level features and The difference is processed by 1×1 convolution and ReLU² activation to enhance local details and suppress noise; finally, the features of the two paths are added to output the enhanced features. The complex attention mechanism is replaced by simple operation, which improves the feature expressiveness while controlling the amount of computation.

[0118] Furthermore, the efficient feature fusion module (EFFM) is responsible for the adaptive fusion of the three-branch features, achieving accurate fusion through boundary feature broadcasting and dynamic weight allocation. Boundary weights are generated by performing a 3×3 convolution on the edge maps output from the boundary branches. The data is broadcast to both the spatial and semantic branches and added element-wise to enhance edge texture information. Mean pooling and max pooling are then performed on the enhanced spatial and semantic features respectively, and the resulting features are concatenated to obtain representative features. and ;Will , and After matrix multiplication, a feature relation matrix is ​​generated by 3×3 convolution and sigmoid activation. ;by and To achieve adaptive weighting, spatial and semantic features are weighted separately and then summed to output fused features, ensuring efficient integration of semantic, spatial, and edge features of the crack.

[0119] Furthermore, the segmentation head employs a lightweight design, consisting of 1×1 convolutions and Softmax activations. It maps the fused features output by EFFM to a binary classification dimension (crack / background). The preceding stage embeds 8x bilinear interpolation upsampling to restore the original image size before outputting a segmentation mask. The loss calculation module is the core constraint unit for model training, employing a multi-loss weighted fusion function, including a boundary branch binary cross-entropy loss. Spatial branch cross-entropy loss Semantic branch multi-scale cross-entropy loss and Final segmentation mask cross-entropy loss and The weighting coefficients for each loss term strictly follow the definitions in the claims, where The edge map output by the boundary branch needs to be multiplied element-wise with the segmentation mask as a weight factor to strengthen the learning of crack boundaries. The loss calculation results are backpropagated to each branch and module to guide the iterative optimization of model parameters.

[0120] This embodiment details how a pavement crack segmentation system achieves fully automated processing from image input to result output through modular design. The multi-source adaptation and preprocessing functions of the image acquisition module ensure the standardization of input data. The core architecture of the TriCrackNet segmentation model endows the system with powerful feature extraction and fusion capabilities. The loss calculation module provides precise guidance for model optimization. The multi-format display and storage functions of the result output module enhance the system's engineering practicality. The system significantly reduces the cost and risk of manual inspection and provides reliable data support for maintenance decisions through efficient and accurate segmentation results, significantly improving the automation level of highway defect detection and the efficiency of engineering implementation.

[0121] Based on Embodiment 1 or 2, this embodiment details the core objective of accurately reflecting the needs of real-time segmentation scenarios. It selects three mainstream road crack datasets—CrackForest, Crack500, and DeepCrack—as test benchmarks, covering real road scenarios with different resolutions and crack morphologies to ensure the comprehensiveness and representativeness of the verification results. All methods involved in the comparison run in a unified hardware and software environment, strictly adhering to the same testing protocol: input images use a uniform resolution, and inference time only counts the model's forward propagation process, excluding interference from non-core aspects such as data loading, preprocessing, and postprocessing; each model undergoes 10 warm-ups before testing, and the average FPS is calculated based on 1000 test images to ensure the accuracy of the speed metric. The evaluation system uses accuracy (… The method is evaluated from three dimensions: accuracy, speed (FPS), and weighted score. The weighted score is normalized and then merges accuracy and speed to objectively reflect the overall performance of the method in real-time segmentation tasks.

[0122] The Crack500 dataset contains a large number of high-resolution (2560×1440) images of road surface cracks, with diverse crack morphologies and complex background interference, making it an important benchmark for validating the accuracy and efficiency of segmentation methods. The comparison results between our method and other state-of-the-art methods are shown in Table 1 below:

[0123] Table 1. Comparison and verification results of the Crack500 dataset.

[0124] The results show that on the Crack500 dataset, TriCrackNet's accuracy metric ( TriCrackNet achieved an FPS of 65.23%, ranking first among all compared methods. This represents a 0.15% improvement over DSNet-head64 (ranked second) and a 0.74% improvement over LPS-Net, which focuses on speed optimization. This fully demonstrates the advantages of the three-branch collaborative architecture in capturing complex crack features. In terms of speed, TriCrackNet achieved 181.58 FPS, slightly lower than LPS-Net (210.56) and DDRNet-23-S (200.29), but significantly higher than most high-precision methods (such as DSNet-head64 with only 47.28 FPS). The WeightedScore, a core metric for evaluating the practicality of real-time segmentation methods, saw TriCrackNet achieve a score of 91.74%, far surpassing other methods and representing a 3.6 percentage point improvement over LPS-Net (88.14%), validating its optimal balance between accuracy and speed.

[0125] like Figure 7 As shown, from top to bottom, the images are: the original image, the ground truth, and TriCrackNet, DSNet, CarNet, PIDNet, LPS-Net, ECSNet, DDRNet-Slim, and BiseNetV2. The inference results of BiseNetV1 demonstrate the visual comparison between TriCrackNet and existing state-of-the-art methods on the Crack500 dataset, characterized by high resolution (2560×1440), complex crack morphology, and strong background interference. As shown in the figures, the original images contain various crack morphologies, including mesh-like, elongated, and intersecting cracks, with exposed aggregate and widespread stains, placing extremely high demands on the detail capture and anti-interference capabilities of the segmentation methods. The Ground truth row clearly marks the true distribution and edge contours of the cracks, providing a benchmark for subsequent comparisons. TriCrackNet's segmentation results are closest to the Ground truth among all methods, not only fully preserving the connectivity of mesh-like cracks and the continuity of elongated cracks, but also accurately delineating subtle changes in crack edges, without obvious breaks, missed detections, or blurred edges. Compared to other methods, while DSNet and CarNet can capture major cracks, they exhibit obvious breaks at small branches and intersections. Methods like PIDNet and LPS-Net suffer from varying degrees of edge blurring and noise interference, misclassifying background as cracks in some areas. Lightweight methods such as the BiseNet series and DABNet suffer from serious missed detection problems, failing to identify a large number of small cracks. TriCrackNet's three-branch collaborative architecture, combined with the EASPP, EFEM, and EFFM modules, enables it to efficiently extract multi-scale semantic, spatial details, and edge texture features even in complex scenes, achieving high-precision crack segmentation.

[0126] The DeepCrack dataset contains a large number of images of road surface and concrete cracks with a resolution of 544×384. The crack textures are complex and vary greatly in scale, placing higher demands on the model's ability to capture multi-scale features. The comparison results are shown in Table 2 below:

[0127] Table 2. Comparison and verification results of the DeepCrack dataset.

[0128] On the DeepCrack dataset, TriCrackNet continues to maintain its high accuracy advantage. TriCrackNet achieved an FPS of 87.69%, ranking first among all compared methods. This represents a 0.05% improvement over LPS-Net (87.64%), which ranked second, and a 1.53% improvement over DDRNet-23-S (86.16%), demonstrating strong adaptability to different types of cracks. In terms of speed, TriCrackNet achieved an FPS of 258.19, second only to BiseNetV1 (293.49). However, BiseNetV1's accuracy was 4.41 percentage points lower than TriCrackNet, failing to meet the requirements for high-precision segmentation. In terms of the comprehensive balanced metric, WeightedScore, TriCrackNet far surpassed CarNet with a score of 88.62%, a 6.68 percentage point improvement over CarNet (81.94%), which ranked second, further proving its core advantage of balancing accuracy and real-time performance in complex crack scenarios.

[0129] like Figure 8 As shown, from top to bottom, the images are: the original image, the ground truth, and TriCrackNet, DSNet, CarNet, PIDNet, LPS-Net, ECSNet, DDRNet-Slim, and BiseNetV2. The inference results of BiseNetV1 are presented, showcasing a visual comparison between TriCrackNet and existing state-of-the-art methods on the DeepCrack dataset. This dataset contains a large number of road and concrete cracks with complex textures and large scale differences, and the cracks often exhibit bifurcated and intersecting patterns, placing higher demands on the model's multi-scale feature capture capabilities. The crack morphologies in the original images are diverse, ranging from large dendritic cracks to fine hairline cracks, and the ground truth row accurately annotates the true contours of various crack types. TriCrackNet's segmentation results are the best among all methods, not only fully preserving the connectivity of large dendritic cracks but also accurately identifying fine hairline cracks, while clearly outlining the edge details at crack bifurcations, without any obvious breaks or missed detections. Compared to other methods, high-precision methods such as DSNet and CarNet exhibit varying degrees of blurring and breakage at subtle cracks; methods like PIDNet and LPS-Net show false connections or missed detections at crack bifurcation points; while lightweight methods such as ECSNet and DABNet suffer from severe loss of detail, with many subtle cracks going undetected, and even background noise being misidentified as cracks in some areas. TriCrackNet effectively solves the challenge of multi-scale crack recognition through multi-scale feature extraction using the EASPP module and feature enhancement using the EFEM module. Furthermore, the adaptive fusion of the EFFM module further enhances the representation of crack edges and spatial details, enabling it to maintain high-precision segmentation even in complex crack scenarios.

[0130] The CrackForest dataset, with a resolution of 480×320, contains crack images under various lighting conditions and road surface conditions, making it a classic dataset for validating model robustness. The comparison results are shown in Table 3 below:

[0131] Table 3 Comparison and verification results of the CrackForest dataset

[0132] On the CrackForest dataset, the accuracy metric of TriCrackNet ( With an FPS of 73.45%, TriCrackNet ranked second, only slightly lower than DSNet-head64 (73.51%), but the difference was only 0.06 percentage points, almost equal. In terms of speed, TriCrackNet achieved an FPS of 274.39, far exceeding DSNet-head64 (219.85), an improvement of 24.8%. In terms of the overall balanced metric, WeightedScore, TriCrackNet significantly outperformed all compared methods with a score of 92.13%, an improvement of 10.02 percentage points compared to CarNet (82.11%), which ranked second, demonstrating a strong overall performance advantage. In addition, TriCrackNet's parameter count (6.44M) and computational cost (5.64 GFLOPs) are within a reasonable range, and it has a lightweight advantage compared to some complex models (such as PIDNet-S-Simple with 7.62M parameters), laying the foundation for subsequent deployment on embedded devices.

[0133] like Figure 9 As shown, from top to bottom, the images are: the original image, the ground truth, and TriCrackNet, DSNet, CarNet, PIDNet, LPS-Net, ECSNet, DDRNet-Slim, and BiseNetV2. The inference results of BiseNetV1 demonstrate the visual comparison between TriCrackNet and existing state-of-the-art methods on the CrackForest dataset. This dataset, with a resolution of 480×320, contains crack images under various lighting conditions and road surface conditions, serving as a classic benchmark for validating model robustness. Cracks in the original images exhibit brightness differences under different lighting conditions, and some areas contain road wear, stains, and other interference, posing challenges to the model's anti-interference ability and detail capture capabilities. The Ground truth row clearly labels the true distribution of various crack types. TriCrackNet's segmentation results are closest to the Ground truth among all methods, not only stably identifying cracks in both strong and weak light regions but also accurately preserving the continuity and edge details of the cracks, without exhibiting significant lighting sensitivity issues. Compared to other methods, DSNet and CarNet suffer from edge blurring and breaks in areas of uneven illumination; PIDNet and LPS-Net exhibit background misclassification in worn areas; while lightweight methods such as the BiseNet series and DABNet suffer from serious missed detection problems, failing to identify numerous fine cracks, and even misclassifying background noise as cracks in some areas. TriCrackNet's three-branch collaborative architecture captures globally illumination-invariant features through semantic branches, preserves details through spatial branches, and strengthens edges through boundary branches. Combined with optimizations in the EFEM and EFFM modules, it achieves stable and accurate crack segmentation even under complex illumination and background interference, demonstrating strong robustness.

[0134] like Figure 10 As shown, from top to bottom, the images are: the original image, the ground truth, and TriCrackNet, DSNet, CarNet, PIDNet, LPS-Net, ECSNet, DDRNet-Slim, and BiseNetV2. The inference results of BiseNetV1 are presented, showcasing a visual comparison of the generalization experiments of TriCrackNet and existing state-of-the-art methods on the DeepCrack dataset. This experiment, trained on a mixed dataset and tested on an unknown dataset, verifies the model's cross-dataset adaptability. The crack morphology in the original images differs significantly from that in the training set, containing more Y-shaped, parallel, and other special crack morphologies, and the background interference is more complex, placing extremely high demands on the model's generalization ability. The ground truth row accurately annotates the true contours of various special cracks. TriCrackNet's segmentation results are the best among all methods, not only completely identifying Y-shaped, parallel, and other special crack morphologies, but also accurately delineating subtle changes in crack edges, without obvious breakage, missed detection, or edge blurring issues. Compared to other methods, high-precision methods such as DSNet and CarNet exhibit significant false connections or missed detections at special crack morphologies; methods like PIDNet and LPS-Net suffer from blurring and breakage at edge details; while lightweight methods such as ECSNet and DABNet suffer from severe missed detection problems, failing to identify a large number of special crack morphologies, and even misclassifying background noise as cracks in some areas. TriCrackNet's three-branch collaborative architecture and lightweight module design allow it to learn general crack features while retaining adaptability to special morphologies. Combined with the constraints of a multi-loss weighted fusion function, it effectively improves the model's generalization ability, achieving high-precision crack segmentation even in cross-dataset scenarios.

[0135] Through comprehensive comparative validation on three major pavement crack datasets, TriCrackNet demonstrates significant advantages in accuracy, speed, and overall balanced performance: In terms of accuracy, it outperforms other datasets by 100% on all three datasets. All three datasets rank first or second, with the Crack500 and DeepCrack datasets ranking first. In terms of speed, the FPS is among the highest, significantly higher than other methods at the same accuracy level. On the weighted score, a comprehensive metric, all three datasets rank first, with a clear lead. These results fully demonstrate that the proposed three-branch collaborative framework, EASPP multi-scale feature extraction module, EFEM feature enhancement module, and EFFM feature fusion module are scientifically sound and effective. They efficiently solve the core problem of balancing accuracy and speed in real-time pavement crack segmentation, providing a high-performance and practical technical solution for real-world engineering scenarios such as highway maintenance and vehicle-mounted inspection.

[0136] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A real-time pavement crack segmentation method based on a three-branch collaborative architecture, characterized in that, include: Obtain the road surface image to be segmented, and input the road surface image into the TriCrackNet segmentation model built based on a three-branch collaborative architecture; The deep semantic features of the road surface image are extracted at multiple scales through the Efficient Hollow Spatial Pyramid Pooling (EASPP) module in the semantic branch of the TriCrackNet segmentation model, resulting in multi-scale semantic features. Spatial detail features of the road surface image are extracted through the spatial branch of the TriCrackNet segmentation model, and crack edge texture features of the road surface image are extracted through the boundary branch of the TriCrackNet segmentation model, thus obtaining spatial detail features and edge texture features. The efficient feature enhancement module EFEM utilizes the global information of the multi-scale semantic features to perform semantic supplementation and local feature adaptive enhancement on the spatial detail features and edge texture features, respectively, to obtain enhanced spatial features and enhanced boundary features. The multi-scale semantic features, enhanced spatial features, and enhanced boundary features are adaptively weighted and fused using the efficient feature fusion module EFFM to obtain fused features. Based on the fusion features and after upsampling, the final segmentation result of the road surface cracks is output through the segmentation head; The loss function of the TriCrackNet segmentation model is a multi-loss weighted fusion function, including boundary branch loss, spatial branch loss, semantic branch multi-scale loss, and final segmentation mask loss.

2. The real-time pavement crack segmentation method based on a three-branch collaborative architecture according to claim 1, characterized in that, The TriCrackNet segmentation model is trained by performing various data augmentation operations on road images, including geometric transformations and pixel-level transformations. The enhanced data is divided into training, validation and test sets according to a preset ratio, and the image pixel values ​​are normalized to a preset range to improve the model's generalization ability and training stability.

3. The real-time pavement crack segmentation method based on a three-branch collaborative architecture according to claim 1, characterized in that, The High Efficiency Hollow Space Pyramid Pooling (EASPP) module extracts multi-scale semantic features, including: Input features after the last downsampling of the semantic branch Features are divided into multiple groups along the channel dimension; Progressive dilated convolution and global average pooling are performed sequentially on the multiple sets of features. After each set of features is extracted, a residual connection is made with the next set of features to serve as the input for the next set of features. The convolutional output features of each group are concatenated with the features upsampled by global average pooling, and then fused with the residuals of the original input features via a 1×1 convolution to output multi-scale semantic features. The specific calculation formula is as follows: in, The dilation rate of the dilated convolution is... Indicates global average pooling. Indicates an upsampling operation. Indicates the kernel size as The convolution operation.

4. The real-time pavement crack segmentation method based on a three-branch collaborative architecture according to claim 3, characterized in that, The spatial branching extracts spatial detail features, including: The TriCrackNet segmentation model uses the output features of the Stem layer as input and performs multiple convolution operations to maintain the feature dimension unchanged. The efficient feature enhancement module EFEM is inserted after the preset convolutional block to guide spatial feature extraction using global information from semantic branches; During the spatial feature upsampling process, the Stem layer features and spatial features are connected by addition to supplement spatial detail information and obtain spatial detail features.

5. The real-time pavement crack segmentation method based on a three-branch collaborative architecture according to claim 4, characterized in that, The boundary branch extracts edge texture features, including: The TriCrackNet segmentation model uses the output features of the Stem layer as input and performs multiple convolution operations to maintain the feature dimension unchanged. The efficient feature enhancement module EFEM is inserted to enhance feature representation. A boundary header is set after EFEM to output an edge map. Edge texture features are extracted by constraining the boundary loss function.

6. The real-time pavement crack segmentation method based on a three-branch collaborative architecture according to claim 5, characterized in that, The efficient feature enhancement module EFEM performs feature enhancement, including: Global average pooling is performed on the low-level features and high-level features respectively to obtain the corresponding global features; the low-level features include spatial branch or boundary branch features, and the high-level features are semantic branch features; Semantic weights are obtained by performing 1×1 convolution and Sigmoid activation on high-level global features; The lower-level global features are supplemented by the semantic weights to obtain enhanced global features. The difference between low-level features and corresponding global features is used to obtain enhanced local features through 1×1 convolution and ReLU² activation; The enhanced global features are added to the enhanced local features to output the enhanced features. The calculation formula is: in, To enhance the global features, Low-level features These are low-level global features.

7. The real-time pavement crack segmentation method based on a three-branch collaborative architecture according to claim 6, characterized in that, The efficient feature fusion module EFFM performs feature fusion, including: The enhanced boundary features are convolved with 3×3 to obtain boundary weights, which are then broadcast to the spatial and semantic branches to obtain enhanced spatial and semantic features. The enhanced spatial features and enhanced semantic features are subjected to mean pooling and max pooling along the channel dimension, respectively, and then concatenated to obtain the corresponding representative features. Matrix operations are performed on the enhanced spatial features, enhanced semantic features, and boundary weights, and the feature relation matrix is ​​obtained by convolution and sigmoid activation. The fusion features are obtained through adaptive weighted fusion. The calculation formula is: in, To enhance post-semantic features, To enhance the post-spatial features, This is the characteristic relation matrix.

8. The real-time pavement crack segmentation method based on a three-branch collaborative architecture according to claim 1, characterized in that, The calculation formula for the multi-loss weighted fusion function is as follows: in, The binary cross-entropy loss is for the boundary branch. The cross-entropy loss is the result of spatial branching. , For the multi-scale cross-entropy loss of semantic branches, , The cross-entropy loss is the final segmentation mask; each loss term corresponds to a preset weight coefficient.

9. The real-time pavement crack segmentation method based on a three-branch collaborative architecture according to claim 8, characterized in that, In the calculation of the final segmentation mask loss term, the edge map output by the boundary branch is used as a weight factor and multiplied element-wise with the segmentation mask. The enhanced mask result and the real label are then used to calculate the loss.

10. A real-time pavement crack segmentation system based on a three-branch collaborative architecture, characterized in that, include: The image acquisition module is used to acquire the road surface image to be segmented and to preprocess the road surface image; The TriCrackNet segmentation model includes semantic branches, spatial branches, boundary branches, the efficient feature enhancement module EFEM, the efficient feature fusion module EFFM, and a segmentation head; The semantic branch integrates the efficient void space pyramid pooling (EASPP) module to extract multi-scale semantic features from road surface images. The spatial branch extracts spatial detail features from the road surface image, and the boundary branch is used to extract crack edge texture features from the road surface image. The efficient feature enhancement module EFEM utilizes global information from multi-scale semantic features to semantically supplement and adaptively enhance spatial detail features and edge texture features. The efficient feature fusion module EFFM performs adaptive weighted fusion of multi-scale semantic features, enhanced spatial features, and enhanced boundary features. The segmentation head upsamples the fused features and outputs the segmentation results of the road surface cracks. The loss calculation module calculates the loss value during model training using a multi-loss weighted fusion function to guide model parameter optimization. The results output module outputs and displays the final segmentation results of road surface cracks.

Images