Efficient crack and surface-type recognition via CNN-block development mechanism and edge profiling.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Automated crack detection is crucial for infrastructure health monitoring.
• Existing methods often lack multi-class capability and are computationally intensive for real-time applications.
• There is a need for efficient, deployable models for edge devices.

Purpose of the Study:

• To develop a lightweight convolutional neural network (CNN) for multi-class crack and surface-type classification.
• To create a compact yet high-performing model suitable for real-time and edge deployment.
• To ensure the model is explainable and robust to real-world distortions.

Main Methods:

• Developed a lightweight CNN using the CNN-Block Development Mechanism (CNN-BDM).
• Integrated domain-driven data augmentation, balanced label design, and systematic regularization.
• Iteratively refined the architecture to create the Lite-V2 model.
• Validated performance on SDNET2018, CrackForest (CFD), and DeepCrack datasets.
• Employed Grad-CAM for explainability and perturbation experiments for robustness.

Main Results:

• The Lite-V2 architecture achieved a macro-F1 score of 0.928 and 95.7% accuracy on SDNET2018 with only 0.28 million parameters.
• Demonstrated strong generalization with F1-scores of 0.975 on CFD and 0.96 on DeepCrack.
• Achieved significantly reduced inference latency (11 ms) on a Raspberry Pi 4, outperforming MobileNetV2, EfficientNet-B0, and ResNet-18.
• Showed robust resilience to brightness and blur variations.

Conclusions:

• Lite-V2 is an efficient, explainable, and deployment-ready framework for crack classification.
• The model is highly suitable for practical condition monitoring in resource-constrained environments.
• The CNN-BDM approach facilitates the development of compact and effective deep learning models.