Hyperparameter for each model.

Structural damage detection and health assessment are crucial for maintaining infrastructure safety and durability. This study presents a novel multi-scale vision-based framework that combines deep learning and machine learning for accurate and interpretable structural safety evaluation. Specifically, we integrate ResNet-50 and SegFormer models to jointly achieve coarse-level damage classification and fine-grained pixel-level segmentation. Seven key damage parameters are quantitatively extracted from high-resolution images—such as crack length, spalling area, and rebar exposure—and serve as interpretable features for safety assessment. A Random Forest (RF) model is developed to establish a nonlinear mapping from these visual features to structural safety levels. Experimental results demonstrate that the RF-based safety assessment model outperforms other traditional machine learning approaches, achieving an accuracy of 87.0%, F1-score of 0.76, and AUC of 0.83, highlighting its strong generalization and classification capabilities. This work offers a comprehensive and generalizable solution for automated structural damage detection and safety evaluation.