Data-Efficient and Context-Aware Deep Learning Models for Crack Analysis: From Robust Detection to Precise Segmentation
Abstract: Cracks are among the earliest and most critical indicators of structural deterioration in civil infrastructure, including road pavements, concrete bridges, and buildings, and their timely detection and precise delineation are essential for ensuring safety, prioritizing maintenance, and reducing long-term repair costs. Yet, automated crack analysis remains a fundamentally challenging computer vision (CV) problem. Cracks are typically thin, irregular, and fragmented, often embedded within complex backgrounds and affected by illumination changes, surface texture variations, shadows, and noise. Moreover, acquiring dense pixel-level annotations for segmentation is labor-intensive and costly, limiting the scalability of fully supervised approaches. While machine learning (ML) and, more recently, deep learning (DL) methods have significantly advanced automated inspection, many existing models still struggle to handle complex real-world scenes, achieve robust cross-dataset generalization, and maintain data efficiency.
This thesis develops a unified, data-efficient, and context-aware DL-based framework for crack analysis, spanning robust object detection to precise semantic segmentation. The thesis begins with crack detection under real-world variability, proposing enhanced ensemble strategies and attention-driven architectures that improve localization accuracy and robustness across heterogeneous datasets.
Building on this detection stage, the thesis advances to pixel-level crack segmentation using convolutional neural networks (CNNs). It first introduces a cascaded segmentation framework that stacks complementary models in a progressive refinement scheme, enabling logit-guided boundary correction and improving the recovery of thin crack structures. This principle is further extended through the design of an efficient multi-decoder architecture with enhanced feature fusion, allowing more effective exploitation of multi-scale representations while maintaining computational practicality. Based on these foundations, context-aware and refinement-based segmentation strategies are developed, integrating Context-Aware Module (CAM), Adaptive Feature Refinement (AFR), and Selective Feature Aggregation (SFA). Together, these architectural advances enhance the integration of global context and fine structural detail, leading to sharper boundary delineation, improved robustness, and stable performance under complex real-world inspection conditions.
To further improve generalization and adaptability, a Mixture-of-Experts (MoE) framework is proposed for crack segmentation. Rather than relying on a single monolithic model, the MoE strategy dynamically selects complementary expert models through hard, soft, and Top-$K$ gating mechanisms. This dynamic expert selection consistently surpasses individual models across multiple benchmark datasets, demonstrating improved robustness to domain shifts and heterogeneous surface characteristics.
Recognizing the high cost of dense annotation, the thesis further explores learning with reduced annotation requirements by leveraging a bounding box-guided segmentation paradigm. Instead of treating bounding boxes merely as weak labels, they are leveraged as spatial guidance to modulate intermediate feature representations during training. This approach reduces dependency on strong pixel-level supervision while preserving competitive segmentation accuracy, offering a scalable and practical solution for large-scale infrastructure monitoring.
Across detection, segmentation, ensemble modeling, and annotation-efficient strategies, this thesis advances crack analysis toward models that are not only accurate but also robust, adaptable, and data-efficient. Extensive evaluation on multiple public and in-house datasets demonstrates competitive and state-of-the-art performance under realistic conditions, narrowing the gap between controlled benchmark evaluation and reliable real-world deployment.