AGMAMBA: A LIGHTWEIGHT ADAPTIVE GUIDED STATE SPACE MODEL FOR PIXEL-LEVEL CRACK SEGMENTATION IN CIVIL INFRASTRUCTURE

Accurate pixel-level segmentation of crack regions is essential for the inspection and maintenance of civil infrastructure such as bridges, pavements, and buildings. Early detection of structural cracks enables timely maintenance actions and reduces the risk of structural failure. However, existing segmentation approaches often struggle to simultaneously capture fine crack textures and suppress complex background noise, which can significantly degrade detection accuracy. Additionally, many high-performing models require large computational resources, limiting their applicability in real-time infrastructure monitoring systems.

Challenges in Existing Crack Segmentation Methods

Current deep learning-based crack detection methods frequently encounter two primary challenges. First, cracks typically exhibit thin, irregular, and discontinuous patterns, making it difficult for models to dynamically capture their texture and morphology. Second, background elements such as shadows, stains, and surface roughness can easily be misidentified as cracks. These challenges often lead to reduced segmentation accuracy or excessive computational complexity when attempting to improve feature extraction.

AGMamba Architecture and Lightweight Design

To overcome these limitations, this study introduces AGMamba, a lightweight adaptive guided state space model designed for efficient crack segmentation. The architecture focuses on capturing crack texture features while minimizing redundant background information. With only 3.40 million parameters and 20.99 GFLOPs, AGMamba achieves a strong balance between computational efficiency and segmentation performance, making it suitable for practical civil infrastructure inspection applications.

Crack Perception Module (CPM)

A key component of the proposed model is the Crack Perception Module (CPM), which integrates two complementary mechanisms:

• Adaptive Guided Scanning Strategy (AGSS2D) – prioritizes crack regions during feature scanning to improve the efficiency of texture extraction.

• Selective Key Clue Modeling (SKCM) – selectively aggregates information from critical crack edges and structural features.

Together, these modules allow the network to focus on meaningful crack patterns while reducing the influence of irrelevant background features.

Frequency-Domain Feature Perception

To further enhance segmentation performance, the model incorporates a High-Low Frequency Feature Perception (HLFP) strategy and a Frequency-Domain Segmentation Head (FDSH). These components analyze differences between high-frequency crack textures and low-frequency background patterns. By leveraging frequency-domain information, the framework effectively suppresses background interference and improves crack boundary detection accuracy.

Experimental Results and Performance Evaluation

Extensive experiments were conducted on four public crack datasets, demonstrating that AGMamba consistently outperforms existing state-of-the-art (SOTA) segmentation models. On the Crack500 dataset, the proposed model achieved an F1 score of 0.7622 and an mIoU of 0.7808, representing improvements of 1.41% and 1.21%, respectively, over previous SOTA methods. These results confirm the model’s ability to achieve high segmentation accuracy while maintaining low computational cost, making it highly suitable for automated infrastructure inspection systems.

Global Civil Engineering Awards

🔗 Nominate now! 👉 https://civilengineeringawards.com/award-nomination/?ecategory=Awards&rcategory=Awardee

🌐 Visit: civilengineeringawards.com

📩 Contact: contact@civilengineeringawards.com

#BridgeInspection
#PavementMonitoring
#AIinCivilEngineering
#StructuralSafety
#AutomatedInspection
#DigitalInfrastructure
#LightweightAI
#EngineeringVision
#CivilEngineeringResearch
#SmartMaintenance

 

Transformer-Based Intelligent Defect Detection in Civil Infrastructure

Detecting structural and functional defects in large-scale civil infrastructure during operational stages is critical for ensuring safety, serviceability, and sustainability. Traditional inspection methods are often labor-intensive, subjective, and inefficient for large-scale systems. With the rapid evolution of artificial intelligence, deep learning has emerged as a powerful tool for intelligent defect detection. Recently, Transformer-based self-attention models have gained prominence as effective alternatives to convolutional neural networks (CNNs), offering superior capability in modeling long-range dependencies and parallel computation. This research-oriented survey systematically explores the integration of Transformer architectures into civil engineering defect detection applications.

Evolution from CNNs to Transformer Models

Conventional CNN-based approaches have demonstrated strong performance in localized feature extraction for defect detection tasks; however, their limited receptive fields restrict the modeling of global structural relationships. Transformers overcome these limitations through self-attention mechanisms that capture global contextual information across large-scale datasets. This paradigm shift has motivated civil engineering researchers to adopt Transformer models for infrastructure monitoring, enabling more comprehensive understanding of spatial and temporal defect patterns.

Transformer Architectures for Engineering Defect Detection

This survey reviews more than 40 Transformer-based engineering defect detection algorithms, highlighting key architectural variants such as Vision Transformers (ViT), hybrid CNN-Transformer models, and hierarchical attention frameworks. These architectures are tailored to handle high-resolution images, sensor data, and multimodal inputs commonly encountered in civil infrastructure monitoring. The adaptability of Transformer models allows effective feature learning across complex structural geometries and diverse defect manifestations.

Application Scenarios in Civil Infrastructure

Transformer-based defect detection methods have been extensively applied to critical civil infrastructure systems, particularly roadways, tunnels, and bridges. In roadway monitoring, Transformers enable accurate detection of cracks, potholes, and surface degradation. Tunnel inspection applications benefit from long-range dependency modeling in low-light and complex environments, while bridge monitoring leverages attention mechanisms to identify structural anomalies across large spans and interconnected components.

Challenges and Limitations

Despite their advantages, Transformer-based approaches face several challenges in civil engineering applications. These include high computational costs, large data requirements, limited labeled datasets, and difficulties in real-time deployment. Additionally, domain adaptation across different infrastructure types and environmental conditions remains a significant research challenge. Addressing these issues is essential for practical, scalable adoption in real-world infrastructure systems.

Future Research Directions

Future development of Transformer-based intelligent detection systems should focus on lightweight architectures, self-supervised learning, multimodal data fusion, and domain-specific model optimization. Integrating Transformers with digital twins, edge computing, and real-time monitoring systems holds promise for next-generation smart infrastructure management. This survey provides foundational insights and reference strategies to guide researchers in advancing intelligent defect detection within civil engineering.

🏗️ Civil Engineering Awards  

👉 Visit our Website: civilengineeringawards.com

#RoadwayEngineering

#VisionTransformer

#AIinEngineering

#SelfAttention

#IntelligentSystems

#DigitalInfrastructure

#EngineeringResearch

#StructuralSafety

#AutomationInCivil

#FutureEngineering

#SustainableInfrastructure