Structural health monitoring (SHM) is a significant contributor to the safety and longevity of civil infrastructure. Out of several indicators, surface cracks are among the most apparent and striking indicators of degradation. Conventional crack detection techniques, although helpful, tend to involve manual intervention and lack accuracy within challenging conditions. To overcome these limitations, this paper proposes a deep learning-driven model, fusing the feature extraction ability of Residual Networks (ResNet) with the interpretability of Gradient-weighted Class Activation Mapping (Grad-CAM++). The resulting hybrid architecture not only supports accurate crack detection but also furnishes clear visual understanding of the decision-making process of the model. The proposed model has been tested on a publicly downloadable dataset with the help of advanced preprocessing techniques along with information obtained with the help of data augmentation techniques. The test results prove the model to outperform existing ones, including UNet, UNet++, and ResUNet, among others. The proposed solution has obtained an accuracy value of 89.20% along with IoU of 82.2%, with high values for Dice coefficient, Jaccard index, recall, precision, and F1-score, thereby verifying robustness and reliability across various measures. By being able to offer both accuracy and interpretability, this particular model emerges as a viable solution for the purpose of conducting real-time, automated structural inspections. The potential to operate efficiently enables it to serve as an ideal contributor to the support for long-term maintenance policies as well as for the purpose of ensuring the structural integrity of critical infrastructure systems.