Abstract
Automated crack detection plays a vital role in the structural health monitoring of civil infrastructure, yet existing methods often remain limited to binary crack identification and are computationally demanding for real-time or edge deployment. This study presents a lightweight convolutional neural network, developed through the CNN-Block Development Mechanism (CNN-BDM), for multi-class crack and surface-type classification across six categories: cracked and uncracked concrete, plaster, and wall surfaces. The proposed framework integrates domain-driven data augmentation, balanced label design, and systematic regularization to achieve a compact yet high-performing model. Through iterative refinement, the final Lite-V2 architecture achieves a macro-F1 score of 0.928 and a test accuracy of 0.957 on the SDNET2018 dataset using only 0.28 million parameters. Cross-domain evaluations further validate the model's generalization, attaining F1-scores of 0.975 on CrackForest (CFD) and 0.96 on DeepCrack. Grad-CAM visualizations confirm interpretable feature localization, while perturbation experiments under brightness and blur variations demonstrate robust resilience to real-world distortions. Comparative analysis against MobileNetV2, EfficientNet-B0, and ResNet-18 reveals that Lite-V2 delivers the highest accuracy and efficiency with up to 40× fewer parameters and significantly reduced inference latency (11 ms) on a Raspberry Pi 4. These results establish Lite-V2 as an efficient, explainable, and deployment-ready framework for practical crack classification and condition monitoring in resource-constrained environments.