Abstract The preservation of paper-based cultural heritage materials is a persistent challenge, especially in tropical and resource-limited contexts where environmental and biological stressors accelerate deterioration. This study investigates the application of deep learning to detect early-stage biodeterioration – specifically insect and mold damage – on digitized paper artefacts. A curated dataset of 274 conservation-grade images was annotated into two classes and evaluated using stratified three-fold cross-validation. Four ImageNet-pretrained convolutional neural networks (CNNs) – VGG-16, ResNet-50, ResNet-101, and MobileNet V2 – were benchmarked. While all models achieved comparable mean accuracy (∼0.719), MobileNet V2 delivered a higher weighted F 1 -score (0.681) and significantly lower training latency outperforming deeper architectures in both efficiency and minority-class detection. This result is particularly relevant for preventive conservation workflows that require rapid identification of rare deterioration events. MobileNet V2’s lightweight architecture also enables potential deployment on edge devices, supporting real-time diagnostics in collection environments with limited computing infrastructure. The findings affirm the viability of compact CNNs for accurate, class-balanced performance under practical constraints. Limitations include a relatively small dataset and controlled imaging conditions, suggesting the need for expanded, multi-institutional corpora and more diverse acquisition scenarios. Future work will incorporate model explainability techniques (e.g., Grad-CAM) and explore transformer-based alternatives. This research contributes a scalable, reproducible AI-based framework for integrating machine learning into digital heritage diagnostics, advancing the intersection of conservation science and library technology.
Ali et al. (Fri,) studied this question.