TL;DR
Experimental results demonstrate that, compared with mainstream semantic segmentation models, MDeepLab significantly reduces the number of parameters while maintaining high segmentation accuracy, exhibiting promising engineering application value for automated road crack detection.
Abstract. Pavement crack segmentation is important for road infrastructure inspection, but practical deployment remains challenging because many high-performance deep learning models require substantial computational resources. This issue is particularly critical in large-scale Mobile Mapping System (MMS)-based workflows, where large volumes of road surface images must be processed efficiently. To address this problem, this study proposes a lightweight CNN-Mamba hybrid architecture for crack segmentation as a deployment-oriented redesign of CT-CrackSeg. The proposed method replaces the original MobileViT-based global modelling modules with EfficientViM-inspired blocks based on hidden-state mixer-based state space duality (HSM-SSD), while preserving the overall encoder-decoder structure and refining the boundary enhancement branch with DCNv2-based deformable convolution. Experiments on the publicly available GAPs384 and CamCrack789 datasets show that the proposed model maintains competitive topology-aware segmentation performance while substantially improving computational efficiency. Compared with CT-CrackSeg, the proposed model increases inference speed from 1.49 to 4.44 FPS on GAPs384 and from 1.32 to 3.92 FPS on CamCrack789, while reducing peak memory consumption from 2827 MB to 355 MB on both datasets. At the same time, the clDice score remains comparable, changing from 0.760 to 0.758 on GAPs384 and from 0.921 to 0.922 on CamCrack789. These results indicate that the proposed architecture provides a favourable balance between crack segmentation quality and deployment efficiency, making it a practical option for large-scale pavement inspection and photogrammetric infrastructure monitoring.
Highlights What are the main findings? YOLO26s, a lightweight deep learning model, achieves 89.0% mAP@0.5 for multi-class pavement crack detection (longitudinal, transverse, pothole, and alligator cracks) while reducing 14.3% parameters and 7.7% FLOPs compared to YOLOv8s, enabling efficient real-time edge deployment. Evaluated on a multinational dataset (RDD2022) spanning four countries (USA, Norway, Japan, China), YOLO26s demonstrates robust cross-domain generalization across diverse environmental conditions, road textures, and imaging perspectives. What are the implications of the main findings? The lightweight architecture supports sustainable infrastructure management by facilitating early, accurate crack detection on resource-constrained edge devices, reducing material waste, energy consumption, and carbon emissions from delayed or repeated road maintenance. The model’s real-time capability and computational efficiency make it scalable for large-scale road network monitoring, offering a practical AI-driven solution for municipalities and maintenance authorities in both developed and emerging regions. Abstract Maintaining durable road infrastructure is crucial for reducing resource consumption, minimizing repair costs, and supporting sustainable urban mobility. However, accurately detecting small and morphologically diverse pavement cracks remains challenging due to variations in lighting, road textures, and crack shapes across different geographic regions. YOLO (You Only Look Once) is one of the most widely adopted deep learning (DL) frameworks for object detection. Traditional inspection methods are labor-intensive and often inconsistent, while existing DL models can be computationally heavy or limited to single crack types, restricting real-time deployment and scalability. To address these challenges, this study presents YOLO26s, a lightweight DL model for multi-class pavement crack detection across diverse environmental and geographic conditions. Using a curated subset of 6972 annotated images from the Road Damage Dataset 2022, YOLO26s identifies four crack types: longitudinal, transverse, pothole, and alligator cracks. Compared to baseline models (YOLOv8s, YOLOv8n, YOLO26n), YOLO26s achieves higher detection accuracy (mAP@0.5 = 89.0%) while reducing computational complexity by 14.3% in parameters and 7.7% in FLOPs, enabling real-time deployment on edge devices. By facilitating early and accurate crack detection, the proposed approach supports proactive maintenance, extends pavement lifespan, and reduces material and energy usage, contributing to more sustainable road network management. These findings highlight the potential of efficient AI-driven inspection systems to enhance environmental and economic sustainability in civil infrastructure.
Pavement crack detection is a critical prerequisite for road maintenance. However, traditional manual inspection and existing deep learning methods struggle to balance inference efficiency with the segmentation accuracy of fine-grained cracks, often leading to topological fractures in slender cracks and missed detections in low-contrast environments. To address these challenges, this paper proposes CBAM-LKC-YOLOv8-seg, a lightweight end-to-end framework designed for high-precision segmentation and automated quantification. The core innovations include (1) a synergistic feature enhancement mechanism that integrates a lightweight convolutional block attention module with 7 × 7 large kernel convolution. This design effectively recalibrates feature weights to suppress background noise while expanding the effective receptive field to preserve the topological continuity of slender cracks. (2) An adaptive training strategy tailored for extreme class imbalance, utilizing optimized loss weights to enhance boundary sensitivity. (3) A robust end-to-end quantification pipeline that fuses morphological processing with skeletonization algorithms to translate segmentation masks into precise geometric parameters (length and width). Experiments on the CRACK500 dataset and cross-scenario test sets demonstrate that the proposed model achieves an mIoU of 77.0% (8.7% higher than the native YOLOv8-seg) and a recall rate of 82.5% for small cracks ( < 5 pixels ). Furthermore, the framework controls measurement errors within 3% while maintaining a lightweight footprint of 13.1M parameters and an inference speed of 28.9 FPS. This method effectively resolves the trade-off between accuracy and efficiency, providing a reliable automated tool for large-scale road maintenance.
Accurate road crack detection is essential for intelligent pavement inspection, yet thin crack morphology, cluttered backgrounds, and deployment constraints still challenge lightweight detectors. This paper presents an improved YOLOv11s-based detector for road distress recognition. Three coordinated modules are introduced: a C3k2- SHSA-CGLU backbone block for stronger contextual perception and dynamic crack-feature filtering, a GLSABiFPN neck for bidirectional multi-scale fusion with enhanced fine-detail retention, and a lightweight shared-convolution detection head for compact prediction. Experiments on the China subset of RDD2022 show that the proposed method improves mAP@0.5 from 87.2% to 89.4% and reduces parameters from 9.41 M to 7.32 M compared with YOLOv11s. Additional cross-dataset results on GRDDC2020 indicate acceptable generalization, while the reduced parameter count and compact model size suggest good deployment potential. Overall, the method provides a practical balance between detection accuracy and model compactness for automated pavement inspection.
Accurate pavement crack segmentation is essential for structural health monitoring, yet existing methods often face a trade-off between segmentation accuracy and computational efficiency. To address this issue, a novel teacher–student framework, termed RTCS-T and RTCS-S, is proposed. The teacher network RTCS-T is constructed based on the Swin Transformer to capture long-range dependencies and multi-scale contextual information. To further enhance crack representation, a strip refinement module is introduced to model directional structural features, while a cascaded atrous spatial pyramid pooling module is employed to improve multi-scale feature aggregation. Based on the teacher network, a lightweight student model RTCS-S is developed by using depthwise separable convolutions to achieve efficient inference. In addition, a foreground-aware and boundary-aware knowledge distillation strategy is introduced to guide the transfer of structural and contextual information from the teacher to the student. Experiments on the Crack500, DeepCrack, and CFD datasets demonstrated competitive performance against representative segmentation models. On CFD, RTCS-S achieved an F1 Score of 0.7514 and an mIoU of 0.7962. Notably, RTCS-S required only 1.82 M parameters and 1.13 GFLOPs and achieved a model inference speed of 680 FPS on an RTX 4090 GPU. When deployed on an RDK X5 edge-computing platform, the complete pipeline achieved an end-to-end throughput of 34 FPS, with an average latency of approximately 29.4 ms and peak memory consumption of 1.8 GB. These results demonstrate that the proposed framework provides an efficient solution for automated pavement crack detection and shows strong potential for practical road inspection applications.