With the rapid expansion of global road networks, pavement maintenance is increasingly challenged by ifne and irregular cracks caused by trafic loads and environmental conditions. Traditional inspection methods, including manual patrols and classical image processing, are often ineficient and lack sensitivity to subtle crack patterns. To address these limitations, we propose a novel road crack recognition framework that integrates object detection with semantic segmentation. The detection module enhances YOLOv11 by incorporating a Diverse Branch Block and a Triplet Attention Module to improve multiscale feature extraction with low computational cost. The segmentation module extends TransUNet by replacing the standard Transformer Encoder with BiFormer Block and embedding a parallel Swin Transformer Block, enabling efective global-local context fusion. Experimental results demonstrate that the improved detection model achieves 83.9% mAP@50 and 65.8% mAP@[50:95] on the Ultralytics CrackSeg dataset. Meanwhile, the enhanced segmentation model attains 78.46% mean Intersection-overUnion on the CRACK500 dataset. These findings confirm the efectiveness of the proposed multi-attention architecture for accurate and scalable road crack analysis.
The rapid expansion and aging of road infrastructure have intensified the demand for accurate and scalable pavement crack detection methods to ensure timely maintenance and structural safety. However, cracks are often fine, irregular, and low-contrast, making them dificult to detect with traditional manual or rule-based methods.
Deep learning-based object detection models[
To address these precision limitations from both tasks, we propose two improved models for road crack recognition: an enhanced object detector based on YOLOv11 incorporating a Diverse Branch Block and Triplet Attention Module, and a segmentation network that upgrades TransUNet by replacing its Transformer Encoder with BiFormer Block and adding a parallel Swin Transformer Block.
Overall, our main contributions can be summarized below: • We embed the Diverse Branch Block into the YOLOv11 backbone and insert the Triplet Attention Module in the neck to enrich multi-path features and joint spatial–channel attention. • We replace the Transformer layer of TransUNet Encoder with BiFormer Block and add a parallel Swin Transformer Block path to create a dual-encoder that captures complementary global-local context.
The remainder of this article is organized as follows. Section 2 reviews the previous road crack detection and the related YOLO models and semantic segmentation models. Details of our proposed methods are introduced in Section 3. Section 4 presents the experiments and the datasets. Finally, Section 5 concludes the paper.
Object detection and semantic segmentation are two mainstream vision approaches used in automated pavement crack analysis. Object detection locates crack regions with bounding boxes, ofering real-time inference and localization capabilities. Semantic segmentation, in contrast, provides pixel-level classification, enabling precise extraction of crack shapes and boundaries. Due to their respective strengths—eficiency in detection and accuracy in structural delineation—both methods have been widely applied in pavement inspection scenarios. They help overcome challenges posed by the fine, irregular, and low-contrast nature of cracks, especially under varying illumination and surface textures.
YOLO models are widely applied in crack detection due to their real-time speed and
localization capabilities[
Semantic segmentation enables pixel-level crack extraction and is efective for detecting
ifne or irregular patterns. Yoon et al. [
In this study, we propose two improved models for road crack recognition: an object detection network based on YOLOv11 and a segmentation network built upon TransUNet. In the detection branch, the original convolutional backbone is replaced with a Diverse Branch Block to enhance multi-scale feature extraction, while a Triplet Attention Module is integrated into the neck to strengthen spatial and channel-wise attention. In the segmentation branch, the Transformer encoder in TransUNet is replaced with BiFormer Block for improved global context modeling, and a parallel Swin Transformer Block is introduced to capture hierarchical local features.
YOLOv11 is a recent advancement in the YOLO series that aims to improve detection
performance through architectural refinement and enhanced feature representation. It maintains
the fast, single-stage design characteristic of previous YOLO models while introducing better
optimization for small objects, high-density scenes, and complex backgrounds. These
improvements position YOLOv11 as an efective baseline for real-time object detection tasks such as
road crack detection [
YOLOv11 introduces architectural improvements to enhance detection accuracy and maintain real-time performance. Its backbone employs the SPPF and C2PSA modules to expand receptive ifelds and strengthen spatial attention. The neck adopts a feature pyramid structure with C3k2 and upsampling layers for efective multi-scale feature fusion, while the head generates predictions at diferent scales using lightweight CBS Blocks. These enhancements improve robustness to small and dense objects, making YOLOv11 well-suited for fine-grained crack detection tasks.
As shown in Figure 2, we propose an improved YOLOv11 model based on Diverse Branch Block and Triplet Attention.
YOLOv11 demonstrates strong real-time performance and competitive detection accuracy,
benefiting from its eficient single-stage design and optimized feature processing pipeline.
However, it still encounters challenges in detecting fine-grained or low-contrast targets, particularly
in cluttered or textured scenes, where the standard backbone may lack suficient feature diversity
and spatial focus. Inspired by Fan et al.[
Diverse Branch Block is a reparameterizable convolution module designed to enrich the representational capacity of convolutional layers by introducing multi-branch structures during training. As illustrated in Figure 3, Diverse Branch Block consists of four parallel branches: a standard k×k convolution, a 1×1 convolution followed by k×k, a 1×1 convolution with average pooling, and a standalone average pooling path. All branches are followed by batch normalization and aggregated before a nonlinear activation.
During inference, these branches are mathematically merged into a single equivalent
convolution kernel, allowing Diverse Branch Block to maintain inference eficiency. This design
enables richer gradient flow and feature diversity during training without incurring runtime
cost. Inspired by Li et al.[
Triplet Attention Module (TAT) is an attention mechanism designed to capture spatial and
channel-wise dependencies more efectively by employing a triplet-branch structure. As
illustrated in Figure 4, it comprises three parallel branches, each computing attention along diferent
axis pairs: (height × channel), (width × channel), and (height × width). These branches apply
convolutional transformations followed by sigmoid activation and inter-branch aggregation
to capture richer feature interactions. According to Misra et al.[
In our model, we insert TAT into the neck of YOLOv11 to enhance spatial–channel attention before final prediction, improving its ability to detect fine and context-sensitive cracks.
TransUNet is a hybrid segmentation network that integrates CNN-based encoders with
Transformer-based global attention, efectively combining local detail extraction and
longrange context modeling. It outperforms traditional U-Net models, especially in scenarios with
irregular shapes or low-contrast boundaries. These capabilities make TransUNet a strong
baseline for pixel-wise segmentation tasks such as road crack detection, where both fine-grained
localization and structural context are essential [
The architecture adopts an encoder-decoder design, where a CNN backbone is used for hierarchical feature extraction and a Transformer bottleneck enhances global feature representation. The decoder integrates skip connections to refine spatial detail and recover semantic resolution. This synergy between convolutional and self-attention mechanisms improves boundary delineation and semantic coherence, making TransUNet particularly efective for dense prediction tasks like road crack segmentation.
As shown in Figure 6, we propose an improved TransUNet model based on BiFormer Block and Swin Transformer Block.
While TransUNet efectively combines CNN encoders and Transformer-based bottlenecks,
it still faces challenges in preserving fine-grained spatial features, particularly in road crack
segmentation where boundary precision and structural continuity are vital. To address these
issues, we propose an enhanced TransUNet framework that incorporates Swin Transformer
Blocks[
BiFormer is a lightweight vision transformer that introduces a Bi-level Routing Attention (BRA)
mechanism to balance global representation and computational eficiency. As shown in Figure
7, the BiFormer Block incorporates depthwise convolution (DWConv), layer normalization (LN),
and a multilayer perceptron (MLP) alongside the BRA module. The BRA adaptively selects
query-key pairs to reduce redundant attention computation, enabling both global context
modeling and eficient feature routing. Residual connections are used throughout to preserve
gradient flow and feature continuity. Inspired by Wang et al.[
The Swin Transformer Block represents a hierarchical vision transformer architecture that eficiently models both local and global dependencies through a window-based multi-head self-attention mechanism. As illustrated in Figure 8, the block sequentially employs regular window-based multi-head self-attention (W-MSA) and shifted window-based multi-head selfattention (SW-MSA), enabling cross-window contextual interaction while preserving linear computational complexity with respect to image size. Each attention layer is followed by a multi-layer perceptron (MLP), with both modules encapsulated within residual connections and layer normalization to enhance optimization stability. By incorporating shifted windows, the architecture introduces inductive biases such as locality and translational equivariance, which are absent in standard Transformer designs. Within our framework, the integration of Swin Transformer Blocks significantly strengthens the encoder’s representational capacity, particularly in modeling fine-grained structures and preserving boundary continuity critical to tasks like road crack segmentation.
The proposed Merge Method is a hierarchical attention-based fusion mechanism designed to
integrate multi-scale features from parallel encoder branches. It comprises two sequential
attention modules: a coordinate attention block [
Initially, feature maps from diferent encoding stages are concatenated along the channel dimension to preserve spatial alignment. The coordinate attention block encodes directional information via separate height- and width-wise pooling, followed by channel transformation, thereby enhancing position-aware channel attention for improved localization of crack boundaries.
Subsequently, the SaE module models inter-channel dependencies through multi-path convolutions, capturing contextual diversity across varying receptive fields. This dual-attention fusion boosts feature selectivity and spatial consistency, ultimately enhancing the efectiveness of hierarchical feature aggregation for downstream segmentation tasks.
In this section, we evaluate the efectiveness of the proposed improved YOLO and TransUNet
models using two benchmark datasets: the oficial Ultralytics Crack Segmentation Dataset
and a reduced version of the Crack500 Dataset[
The improved YOLO model is evaluated on the oficial Ultralytics Crack Segmentation Dataset, which contains a single crack category and is divided into 3,717 training images, 200 validation images, and 112 testing images. For the improved TransUNet model, we use the reduced version of the Crack500 dataset due to the high computational cost and memory requirements of the full dataset, which includes two semantic classes—crack and background—with 2,413 images for training and 603 for testing.
All experiments were conducted on a computing platform equipped with an NVIDIA GeForce RTX 4070 GPU, utilizing PyTorch 2.0.1 and CUDA 11.8 as the deep learning framework. The improved YOLO model was implemented based on the Ultralytics 8.3.0 framework, with training performed over 100 epochs using a batch size of 16, while maintaining default hyperparameter settings. For the improved TransUNet model, the same hardware and software environment was employed. Training followed an iteration-based strategy with 20,000 iterations, using a batch size of 4. The Adam optimizer was adopted with an initial learning rate set to 0.0001.
Typical metrics used for object detection tasks have been used to evaluate models in this study, including Precision, Recall, mAP50, and mAP50-95. The definitions of these metrics are as follows: = =
= =
+
+ ∑︀ (+1 − ) (+1) =1
1 ∑︀ (1) (2) (3) (4) (5)
To further assess the crack detection capability of the improved TransUNet model, the mean Intersection over Union (mIoU) was incorporated as an additional evaluation metric alongside Precision and Recall. It’s definition is shown below: = 1 ∑︁
=0 + +
The Average Precision (AP) of all classes is the area of the region below the precision-recall curve. represents the recall of the th value, and (+1) represents the highest precision value in the range to +1. The mAP is calculated by averaging the AP of each class in the dataset. mAP50 is obtained by averaging the AP (IoU = 0.5) of all classes, and mAP50-95 is obtained by averaging the mAPs at diferent IoUs between 0.5 and 0.95. mIoU measures the average overlap between the predicted and ground truth regions across all classes.
To validate the efectiveness of the proposed model in crack detection, YOLOv5 and YOLOv8 are introduced as baseline comparison methods. The performance of each model on the oficial Ultralytics Crack Segmentation Dataset is summarized in Table 1.
Compared to YOLOv5, our model improves mAP50 by 4.35% and mAP50–95 by 6.2%, indicating enhanced detection accuracy. It also outperforms YOLOv8 and YOLOv11 in mAP50–95, demonstrating better localization. These results confirm the efectiveness of our architectural improvements for crack detection.
To validate the efectiveness of the proposed enhancements, ablation study results for each individual improvement are presented in Table 2.
The introduction of the Diverse Branch Block leads to a 0.6% increase in mAP50 and a 0.1% gain in mAP50–95, highlighting its contribution to multi-scale feature enhancement. Incorporating the TAT module further improves mAP50 by 0.8% and mAP50–95 by 3.3%, demonstrating its efectiveness in spatial and channel attention. Their combination yields the highest performance, validating the synergy of both modules.
To intuitively demonstrate the efectiveness of the proposed improvements, a visual comparison is conducted between the baseline YOLOv11 and the enhanced model. Representative results are presented in the figure 9.
We observe that the proposed method achieves more accurate crack localization and higher confidence scores compared to YOLOv11. It produces fewer false detections and better captures complete crack structures, particularly in fine or low-contrast regions, demonstrating improved detection robustness and spatial precision.
All comparative models in this study were reproduced using the MMSegmentation 0.29.1 framework. To validate the efectiveness of the proposed model in crack segmentation, UNet and UNet++ are introduced as baseline comparison methods. The performance of each model on the reduced Crack500 Dataset is summarized in Table 3. All comparative models in this study were reproduced using the MMSegmentation 0.29.1 framework.
Compared to TransUnet, our model improves mRecall by 17.36% and mIoU by 9.34%, indicating enhanced segmentation completeness and region overlap. It also outperforms Unet and Unet++ in all metrics, demonstrating superior precision–recall balance. These results confirm the efectiveness of our architectural improvements for semantic segmentation.
To validate the efectiveness of the proposed enhancements, ablation study results for each individual improvement are presented in Table 4 .
The integration of the Biformer module leads to an increase of 15.74% in mRecall and 8.11% in mIoU compared to the baseline TransUNet, highlighting its efectiveness in improving contextual understanding. Incorporating the Swin Transformer further enhances mIoU by 0.45%, demonstrating its strength in capturing long-range dependencies. The combination of our architectural refinements achieves the highest mRecall and mIoU, validating the complementary benefits of both modules for segmentation accuracy.
To intuitively demonstrate the efectiveness of the proposed improvements, a visual comparison is conducted between the UNet++ and the enhanced model. Representative results are presented in the figure10 .
We observe that the proposed method delivers more precise crack segmentation compared to UNet++. It better preserves the continuity and topology of fine cracks, especially in noisy or complex backgrounds. The results exhibit fewer broken or fragmented regions and reduced false positives, indicating enhanced segmentation accuracy and structural consistency.
In this study, we proposed improved network architectures for both road crack detection and segmentation tasks. For object detection, we enhanced the backbone of YOLOv11 by introducing the Diverse Branch Block and integrated the Triplet Attention Module into the neck to improve spatial and channel attention. For semantic segmentation, we modified the original TransUNet by incorporating Swin Transformer Blocks and a BiFormer Block, and designed a hierarchical Merge mechanism based on coordinate and excitation attention to strengthen multi-scale feature fusion.
Experiments conducted on the Ultralytics Crack Segmentation Dataset and the reduced Crack500 dataset demonstrate that our improved YOLOv11 model achieves superior mAP50 and mAP50–95 performance compared to YOLOv5, YOLOv8, and the original YOLOv11. Similarly, the improved TransUNet outperforms baseline segmentation networks in terms of mIoU, particularly in capturing fine-grained crack boundaries.
In summary, our enhanced detection model significantly boosts sensitivity and robustness for detecting small or subtle cracks, while the improved segmentation model ofers superior spatial accuracy and contextual understanding, proving the efectiveness of our design in real-world road crack analysis tasks.
The author(s) have not employed any Generative Al tools.