Generalising Hybrid OUR-Net -U-Net for Accurate Identification and Multi-Domain Structural Crack Segmentation Using Refinement Techniques

Technical Summary: Generalising Hybrid OUR-Net-U-Net for Accurate Identification and Multi-Domain Structural Crack Segmentation Using Refinement Techniques

Problem Statement
Structural cracks in civil infrastructure (bridges, buildings, pavements) are critical indicators of potential failure. While manual inspection is the traditional standard, it is labor-intensive, subjective, and prone to human error, particularly at scale. Deep learning-based segmentation models have improved detection but typically rely on fully supervised learning, which requires extensive, expensive, and time-consuming pixel-level annotations. Existing weakly supervised methods often produce coarse localization maps that lack the boundary precision required for accurate segmentation. Furthermore, many models struggle to generalize across diverse crack patterns (varying widths, orientations, and contrasts) and are sensitive to environmental noise, lighting variations, and surface textures.

Methodology
The paper proposes a weakly supervised framework that bridges the gap between image-level classification and dense pixel-level segmentation. The methodology consists of a six-stage pipeline:

1. Dataset Acquisition: The study utilizes a public concrete crack dataset containing 802 images with diverse crack patterns, surface textures, and lighting conditions. The data is split 80:20 for training and testing.
2. Preprocessing: Images undergo resizing, normalization, and data augmentation (rotation, flipping) to enhance feature consistency and robustness. Conventional techniques like Otsu thresholding and edge detection are investigated as baselines but are primarily used for preprocessing to enhance crack visibility.
3. Pseudo-Mask Generation (Hybrid U-Net): Instead of manual pixel labeling, the framework trains a CNN classifier using only image-level labels (crack vs. no-crack). It then employs Class Activation Mapping (CAM) and Gradient-weighted CAM (Grad-CAM) to generate coarse localization heatmaps. These maps are fused using a weighted binomial relation and refined through spatial consistency constraints and morphological operations to create high-quality "pseudo-masks."
4. Segmentation (OUR-Net): The generated pseudo-masks serve as supervision for training a hybrid OUR-Net architecture. Unlike standard U-Nets, OUR-Net utilizes Octave Convolutions to decompose feature maps into high-frequency (crack edges, thin details) and low-frequency (global structure, continuity) components. This allows the network to simultaneously capture fine hairline cracks and broader structural fissures. The decoder employs Octave Max Unpooling (OMU) to reconstruct spatial information accurately.
5. Refinement: The raw probability maps from OUR-Net are post-processed using thresholding, morphological operations (erosion, dilation, opening, closing) to remove noise and fill gaps, and skeletonization to reduce crack regions to single-pixel centerlines for better topological representation.
6. Visualization: The final refined mask is overlaid on the original image for structural assessment.

Key Contributions

• Weakly Supervised Framework: The integration of CAM/Grad-CAM with a Hybrid U-Net to generate pseudo-masks, reducing annotation dependency by approximately 39% compared to fully supervised methods.
• Hybrid Architecture: The combination of a classification backbone for localization and an OUR-Net segmentation model that leverages Octave Convolutions to handle multi-frequency crack features (both thin and wide cracks).
• Refinement Pipeline: The implementation of a comprehensive post-processing suite including thresholding, morphological operations, and skeletonization to improve crack continuity and remove false positives.
• Statistical Validation: A rigorous statistical analysis of model confidence using 95% confidence intervals across different crack severity levels (from "Very Mild" to "Heavy").

Results
The proposed framework was evaluated against state-of-the-art weakly supervised and fully supervised models (including OUR-Net-CAM, OUR-Net-GradCAM, DeepLabV3, and SegNet) using standard metrics: Accuracy, Precision, Recall, Dice Coefficient, and Intersection over Union (IoU).

• Performance: The proposed Hybrid OUR-Net-U-Net achieved an overall classification accuracy of 85% (specifically 0.8489 in the results section), a Dice coefficient of 0.72, and an IoU of 0.55.
• Comparison: The model outperformed other weakly supervised baselines. For instance, standard CAM-based localization yielded an IoU of 0.52, while the proposed method improved this to 0.55.
• Confidence Analysis: Statistical analysis revealed that prediction confidence increases with crack severity. "Heavy" cracks showed a mean confidence of 0.8423 with narrow confidence intervals, whereas "Very Mild" cracks had lower confidence (0.8199) and wider intervals, reflecting the inherent difficulty in detecting low-contrast, fine features.
• Annotation Efficiency: The study highlights a trade-off where the proposed method achieves competitive performance while significantly reducing the time and cost associated with dataset preparation.

Significance and Claims
The paper claims that this framework provides an efficient and scalable solution for automated structural health monitoring. By reducing the reliance on expensive pixel-level annotations, the method makes large-scale deployment more feasible for real-world infrastructure inspection. The author emphasizes that the integration of explainable AI techniques (CAM/Grad-CAM) with advanced segmentation (OUR-Net) effectively addresses the limitations of coarse localization in weakly supervised learning. The inclusion of morphological refinement and skeletonization further enhances the system's ability to identify and quantify fine cracks, which are often missed by standard segmentation models. The study concludes that while fully supervised models may offer marginally higher accuracy due to precise ground truth, the proposed weakly supervised approach offers a practical balance between detection reliability and annotation cost, making it suitable for automated, large-scale structural assessment.