A Six-Stage Ablation-Driven Benchmarking Framework for Deep Learning-Based Deterioration Classification in Heritage Structures

This study presents two complementary contributions to the automated analysis of surface deterioration in historical buildings. First, we introduce a new large-scale dataset comprising 23,688 images across six deterioration categories, collected from diverse heritage materials and surface conditions. The dataset is designed to support reproducible benchmarking and will be publicly released. Second, we develop a structured, six-stage ablation and knowledge distillation framework that evaluates the incremental effect of widely used deep learning components—including ArcFace, class-balanced focal loss, MixUp/CutMix, Sharpness-Aware Minimization (SAM), Test-Time Augmentation (TTA), and standard distillation. Starting from an EfficientNet-B0 baseline and distilling into a ConvNeXtV2-Tiny student network, the framework achieves a Macro-F1 score of 0.9966 while reducing computational cost. The results demonstrate that meaningful accuracy gains can be achieved through carefully designed training strategies rather than architectural novelty. The framework also provides practical guidance for lightweight deployment in heritage monitoring applications.