AN EXPLAINABLE AI FRAMEWORK FOR AUTOMATED DAMAGE ASSESSMENT USING A SIAMESE RESNET50 NETWORK WITH GRAD-CAM AND OPENCV-BASED VISUAL LOCALIZATION

Abstract

Traditional damage assessment methods rely heavily on manual inspections, which are often time-consuming, labor-intensive, and prone to human error. While recent deep learning models have improved detection accuracy, many function as "black-box" systems, lacking the transparency required for critical industrial applications. In this regard, this research introduces an Explainable AI (XAI) framework for automated damage assessment using a Siamese Neural Network (SNN) with a ResNet50 backbone. Unlike conventional classification models, this architecture employs similarity learning to compare pre-damage and post-damage image pairs using Euclidean distance. To ensure interpretability, the framework incorporates OpenCV-based contour detection and coordinate mapping techniques to visually localize and highlight damaged areas with bounding boxes. Experimental results demonstrate that the proposed model achieves a testing accuracy of 95.60%, a precision of 94.80%, and an F1-Score of 94.35%, outperforming several existing CNN and YOLO-based benchmarks. This approach provides a reliable, accurate, and human-interpretable solution for structural and vehicle damage inspection.