Towards AI-Assisted Preventive Conservation in Libraries: Deep Learning for the Detection of Insect and Mold Damage in Ancient Manuscripts

1 Introduction

Paper-based documentary heritage is intrinsically vulnerable to biological deterioration, a vulnerability that is magnified in tropical and subtropical climates where elevated temperature and relative humidity accelerate the growth of biodeteriogenic agents. Empirical surveys of library stacks and archival repositories consistently identify filamentous fungi – principally Aspergillus, Cladosporium, Fusarium and Scopulariopsis – as the dominant colonizers of paper fibres (Bankole 2010). These taxa produce organic acids, extracellular enzymes and colored metabolites that weaken cellulose, disfigure surfaces, and present allergenic or toxigenic hazards to staff and patrons (Mishra 2017). Parallel investigations have documented mechanical damage inflicted by insects such as cockroaches (Blattodea) and termites (Isoptera), whose feeding, tunnelling, and frass deposition reduce paper integrity and accelerate subsequent microbial invasion (Mosneagu 2012). Current mitigation relies on environmental regulation, scheduled visual inspection, and, where warranted, selective application of biocides. While indispensable, these approaches are labor-intensive, reactive, and subject to human observational thresholds that often miss incipient biodeterioration.

Advances in artificial intelligence (AI) and machine learning (ML) offer a transformative alternative by enabling non-invasive, rapid, and repeatable diagnostics. Convolutional neural networks (CNNs) have recently achieved high classification accuracy for mold species directly on digitised manuscript images (Liu et al. 2024). Computer-vision pipelines based on deep learning detect early insect infestation in forestry and museum objects, demonstrating generalizability across biological taxa and substrates (Senthilkumar and Sumathi 2023). In parallel, supervised algorithms such as k-Nearest Neighbour (k-NN) have proven effective in predicting paper-quality parameters linked to environmental resilience (Kalavathi Devi et al. 2023). Despite these individual successes, a comprehensive evaluation of multiple ML architectures for simultaneous detection of fungal and insect deterioration in library collections remains absent from the conservation literature.

The present study therefore investigates which machine-learning architecture delivers the most accurate, efficient, and generalizable performance in recognizing and classifying paper deterioration attributable to fungi and insects. To answer this question, a domain-specific, expert-annotated image corpus of bio determinant for paper deterioration will be compiled and used to benchmark four supervised models – CNN, support-vector machine, random forest, and k-NN – under conservation-relevant evaluation metrics (precision, recall, F1, inference latency). The objective is to provide an evidence-based recommendation for a scalable, low-cost diagnostic workflow that can be integrated into preventive-conservation programmes. This research contributes novelty on three fronts. First, it creates the first open, high-resolution image dataset explicitly pairing fungal and insect damage typologies in library materials. Second, it offers the inaugural side-by-side comparison of heterogeneous ML classifiers applied to paper biodeterioration, thereby filling a methodological gap between single-organism studies and holistic collection management. Third, it positions the resulting workflow as an early-warning system suitable for climate-vulnerable and resource-constrained institutions, thus extending the protective reach of scientific conservation beyond laboratories with advanced imaging infrastructure. Collectively, these contributions advance the integration of data-driven decision support into heritage preservation practice and lay the groundwork for future implementations of autonomous monitoring in library stacks and archives.

2 Literature Review

2.1 Paper Deterioration

Paper has become a staple in storing knowledge and information for centuries. The library collection materials are dominated by paper which forms in various materials such as parchment, palm leaves, birch bark, leather, and adhesives used in bookbinding. However, time made those materials deteriorate (Pilette 2007). Many experts have discovered and classified the book damage in the library and provided complete information regarding the book damage. Currently, paper is made from cellulose from tree bark. It is organic and vulnerable to many biological attacks that may impact the substrates and graphic media and create stains and discoloration (Di Bella et al. 2015). There are two common types of paper deterioration in the library collection: insect and fungus. The insect attack and fungal growth by adverse environmental conditions such as extreme fluctuations of relative humidity cause dampness due to significant temperatures and light and atmospheric pollutants (Bankole 2010). The biological factor of paper deterioration could be a disaster for libraries.

The library storage environment is the main factor in deteriorating papers. Savoldelli et al. (2021) mention moisture is the main factor in inviting the presence of biological growths such as fungi, insects, and rodents, causing an infestation. Biological agents attack paper and other organic materials when temperature and humidity are uncontrolled. Uncontrolled humidities create fungus spores that could grow in the library collection and develop staining and deterioration. Fungus could rapidly lose strength in paper made from organic materials (cellulose).

The primary characteristic growth of mold and fungi is revealed by the stain in the paper, which made whitish patches on book covers and documents, which may become brownish or greenish (Sterflinger and Pinzari 2012). Mold and fungi grow in organic materials that are tightly packed and have stagnant pockets of moist air that favor mold growth. Pinheiro et al (2019) mentioned that several fungi or molds could excrete acidic compounds of pH in paper to their development. Fungal growth could increase a significant production of acidic metabolites in the paper, creating severe stains in the documents and deteriorating papers and their content (Figure 1).

Figure 1:

Fungal infection on paper deterioration.

Figure 2:

Insect infection in paper deterioration.

Another biological agent that greatly contributes to paper deterioration is insects. Insects usually attack paper and book covers (bindings) using leather, parchment, cardboard, wood, or wooden shelves and can be infested by a few insect pests. They will leave severe damage if it is not promptly found (Querner 2015). The hole usually recognizes the characteristic of insect damage in the paper (Figure 2). For instance, termites could be recognized in the library by the mud tunnels formed on the walls, bookcases and library furniture. Querner et al (2022) mentioned another critical group of pests on books and paper are the silverfish and book lice. Different species of Silverfish are found inside buildings (usually identified all as silverfish): Silverfish (Lepisma saccharina), paper fish (Ctenolepisma longicaudatum), firebrat (Thermobia domestica) and four-lined Silverfish (Ctenolepisma quadriseriata) all feed mainly on detritus, mold, human skin or hair (textiles, cotton, silk) but can also damage paper, bookbinding, wallpaper, papier-mâché́, starch glue, and cellulosic materials. Book lice (Psocoptera) can also be found in high numbers if humid conditions are present. Species like Liposcelis usually eat mold and starch but can damage paper, bookbinding, herbal specimens, wallpaper, and even stuffed animals.

The characteristic Insect damage usually left sharply-defined edges with curves, making it distinguishable from scuffing and mechanical damage. Death-watch and furniture beetles bore through the boards and text block, spider beetles and carpet beetles and moths only eat keratin, and silverfish usually graze on paper and book cloth, although if conditions are damp enough, they sometimes eat leather (Querner 2015). Silva et al. (2013) mentioned insects attack on the front cover and the edges of the pages, laying their eggs and eating the papers and moving to the book’s interior, chewing the corners near the spine when they hatched. The damages of insect could become more extensive and deeper when the larvae have become adult insects.

Every instance of paper deterioration has its pattern and characteristics. Currently, pattern recognition and computer problems have developed rapidly to help humans solve problems in recognizing objects. The Recognition system in computer vision is getting better to achieve near human levels of recognition. In library science, computer vision implementation discussion is still low. However, implementing deep learning in the library has changed the library environment. Searching for images is a radically evolving user experience. A photograph taken can be uploaded and matched to other images and an entire ecosystem of linked resources and descriptions (Coleman and Keller 2022). The connection between computer vision and the library could be a great discussion that leads to a new discovery in the libraries.

3 Methodology

4 Findings

The experimental pipeline processed 274 conservation-grade images that had been evenly stratified into three diagnostic categories – fungal damage, insect damage, and undamaged – a distribution confirmed by the console message “Found 274 images belonging to 3 classes.” For every backbone the corresponding pre-trained ImageNet weights were downloaded once from the Keras repository (≈56 MB for VGG-16, 95 MB for ResNet-50, 171 MB for ResNet-101, and 9 MB for MobileNet V2). This step added <60 s to the first fold but had no impact on subsequent computation because the weights were cached locally. To gauge the stability of each architecture under stratified resampling, the fold-level performance trajectory is examined, documenting the accuracy and weighted F₁ obtained on each of the three validation folds. Table 1 reproduces the raw fold-wise metrics.

Table 1 provides details of the per-fold accuracy and weighted F₁ for each backbone, followed by the cross-validated mean. The three folds exhibit minimal variance for VGG-16 and both residual networks, with accuracy confined to the 0.7140.725 band and F₁ clustered around 0.60–0.63. This consistency implies that, after freezing the convolutional tiers, these deeper architectures generalize uniformly across stratified partitions of the limited dataset. By contrast, MobileNet V2 displays a wider intra-fold spread – achieving its strongest performance in Fold 1 (0.750/0.691) and slightly weaker yet still competitive scores in Folds 2 and 3 (0.703/0.684 and 0.703/0.669, respectively). Despite this variability, the lightweight network attains the highest cross-validated F₁ (0.681) while matching the mean accuracy (0.719) of its deeper counterparts. The elevated F₁ indicates superior balance between precision and recall, attributable to improved detection of the minority insect-damage class. Consequently, MobileNet V2 delivers the most favorable trade-off between class-balanced discrimination and computational efficiency, reinforcing its suitability for real-time, resource-constrained preservation settings.

Figure 3 (below) condenses these results into a model-level comparison. Mean accuracies cluster tightly (0.7189–0.7190), indicating that – with frozen convolutional layers – the backbones converge on an almost identical decision boundary. The weighted F₁-score reveals the decisive difference: MobileNet V2 achieves 0.6811, outperforming VGG-16 (0.6189) and both residual networks (0.6015). Confusion-matrix inspection (not shown) attributes this superiority to higher recall for the minority insect-damage class – precisely the scenario in which early detection is mission-critical for preventive conservation.

Figure 3:

Average accuracy and F1 score.

Figure 1 above juxtaposes the cross-validated average accuracy (left panel) with the average weighted F₁-score (right panel) for the four transfer-learning backbones. The accuracy bars are virtually indistinguishable, clustering at ≈ 0.719 across VGG-16, ResNet-50, ResNet-101, and MobileNet V2. This visual parity corroborates the numerical finding in Table 1 that, once their convolutional filters are frozen, all networks capture an essentially shared decision boundary for the three-class problem.

A markedly different pattern emerges in the class-balanced metric. The right-hand panel shows that MobileNet V2 attains an F₁-score of 0.681, lifting it more than six percentage points above VGG-16 (0.619) and roughly eight percentage points above both residual networks (0.602). The divergence between the two plots highlights a critical insight: raw accuracy alone masks substantive disparities in minority-class recall. MobileNet V2’s superior F₁ reflects its heightened sensitivity to insect-damage instances, a property verified in the confusion-matrix analysis.

From an operational standpoint, this distinction is decisive. Heritage-monitoring systems must prioritize early detection of rare but consequential deterioration events, a model that balances precision and recall therefore confers greater preventive value than one optimized solely for overall hit rate. Accordingly, Figure 3 visually substantiates the textual conclusion that MobileNet V2 offers the most advantageous trade-off between class-balanced discrimination and computational efficiency, reinforcing its selection as the preferred backbone for real-time, resource-constrained conservation workflows. Taken together, the findings substantiate that a lightweight, interpretable convolutional architecture – trained on a domain-specific, biodeterioration-annotated image set – can equal or surpass deeper, more parameter-intensive networks in class-balanced performance while offering a markedly lower computational footprint. In practical terms, the hypothesis predicts that such a model will (i) achieve accuracy comparable to state-of-the-art backbones; (ii) deliver a higher weighted F₁-score by improving recall on minority deterioration classes (e.g., insect damage); and (iii) train and infer fast enough to be deployed in resource-constrained heritage institutions, thereby filling the operational gap identified in the literature review.

5 Discussion

The study demonstrates that the lightweight MobileNet V2 backbone attains accuracy parity with markedly deeper networks while securing the highest class-balanced performance and registering an order-of-magnitude reduction in computational latency. For preventive conservation, this result is consequential. Early-warning frameworks must maximize recall for low-incidence yet high-impact events – principally insect infestation and incipient mold – rather than marginal gains in global hit rate (Sokolova and Lapalme 2009). The superior weighted F₁ achieved by MobileNet V2 therefore yields a higher probability of flagging emergent biodeterioration before irreversible loss occurs, consonant with current guidelines that advocate rapid intervention, minimal object handling, and data-driven risk stratification (Ghaith and Hutson 2024).

From an operational standpoint, this substantially lowers the hardware barrier for small or resource-constrained repositories – an obstacle repeatedly highlighted in recent surveys across the Global South (Teel 2024). The models should enable edge deployment on low-power micro-servers or smartphones, thereby supporting in-situ shelf monitoring, roving condition audits, or crowd-sourced community science initiatives without breaching collection-handling protocols. In practice, there are three immediate gains to scalable stack surveillance. When coupled with inexpensive IoT cameras, the model can perform scheduled or event-triggered scans, automatically logging provenance-rich alerts into a collection-management system – extending the “digital-twin” paradigm beyond environmental telemetry into visual diagnostics (Barlindhaug 2022). Concerning triage and resource allocation, curators can prioritize manual inspection and laboratory testing for objects flagged with high activation maps, optimizing scarce conservation labor and chemical-testing budgets (Mishra 2017). Concerning longitudinal risk analytics, continuous inference streams permit time-series modelling of deterioration rates at the item, shelf, or room scale, informing HVAC tolerances and integrated pest-management thresholds in a statistically accountable manner (Gadd et al 2024).

Moreover, because MobileNet V2 retains full compatibility with on-device explainability methods such as Grad-CAM, saliency overlays can be presented directly in the user interface, fostering human-AI symbiosis and strengthening curatorial trust (Selvaraju et al. 2017). Finally, the low energy demand aligns with emerging sustainability mandates for cultural-heritage technology deployments, reducing both carbon footprint and lifecycle cost (UNESCO 2015). Collectively, these implications suggest that lightweight, class-balanced vision models constitute a pragmatic bridge between advanced AI research and day-to-day preservation practice, offering an immediately actionable pathway to data-driven collection stewardship. Several caveats temper the generalizability of these results. First, the dataset comprises only 274 images, a scale that, while representative of high-value heritage objects, constrains statistical power and limits the exploration of rare deterioration modes such as foxing or chemical scorch. Small-sample regimes are susceptible to variance inflation despite stratified cross-validation (Kohavi 1995). Second, all photographs were captured under controlled D65 illumination; performance may degrade under heterogeneous lighting typical of historic stacks. Third, the decision to freeze convolutional layers avoided over-fitting but may have capped the attainable recall on subtle fungal staining patterns. Fine-tuning the final residual block or employing regularized unfrozen training could further enhance sensitivity (Howard et al. 2017).

To strengthen external validity and operational readiness, future research should advance along three complementary axes. First, the training corpus should be expanded and augmented by assembling a multi-institutional dataset that spans diverse lighting conditions, capture devices, and deterioration phenotypes; synthetic transformations such as spectral jitter and elastic deformation can further mitigate sample scarcity without additional object handling (Shorten and Khoshgoftaar 2019). Second, model diversification and fine-tuning are needed such as systematically benchmarking self-supervised vision-transformer variants (Liu et al. 2024). Selectively fine-tuned convolutional networks, while applying Bayesian hyper-parameter optimization, would clarify trade-offs among network depth, latency, and recall. Third explainability should be embedded within a human-in-the-loop workflow by integrating pixel-level interpretability tools – such as Grad-CAM or Layer-wise Relevance Propagation – into curator dashboards and coupling them with an active-learning loop in which conservators validate or correct predictions, thereby iteratively refining model weights and generating provenance-rich audit trails (Selvaraju et al. 2017). Pursued together, these trajectories promise to elevate both the methodological rigor and the practical utility of AI-driven diagnostics for cultural-heritage preservation.

6 Conclusions

This study presents the first systematic, cross-validated comparison of four ImageNet-pre-trained convolutional backbones – VGG-16, ResNet-50, ResNet-101 and MobileNet V2 – applied to a newly curated, expert-annotated corpus of 274 paper-heritage images exhibiting fungal damage, insect damage and undamaged states. Under identical transfer-learning constraints, all architectures converged on an almost identical mean accuracy (≈0.719), confirming the broad transferability of high-level ImageNet features to cultural-heritage imagery. Crucially, however, MobileNet V2 achieved the highest weighted F₁-score (0.681 ± 0.012) and an order-of-magnitude reduction for deeper networks). These results substantiate the central hypothesis that a lightweight, interpretable network can equal or surpass deeper, parameter-intensive models in class-balanced performance while offering a markedly lower computational footprint.

The implications for preventive conservation are immediate. MobileNet V2’s superior recall for the minority insect-damage class enhances the probability of early intervention, while its modest hardware requirements enable deployment on low-cost edge devices, thereby extending advanced diagnostic capacity to resource-constrained institutions. By publicly releasing the biodeterioration dataset and reproducibility package, the study also furnishes the community with a benchmark for future methodological innovation. Several limitations merit acknowledgement: the modest sample size, controlled imaging conditions, and frozen-feature protocol may constrain generalisability; interpretability analyses remain to be integrated; and emerging vision-transformer architectures were not evaluated. Addressing these constraints through multi-institutional dataset expansion, model fine-tuning, explainable-AI overlays, and transformer benchmarking constitutes a clear agenda for subsequent research.

In sum, the findings advance the field toward scalable, data-driven preventive conservation, demonstrating that lightweight deep-learning models can deliver curator-relevant sensitivity without prohibitive computational overhead. By bridging the gap between laboratory-scale AI research and day-to-day collection stewardship, this work lays the groundwork for next-generation early-warning systems capable of safeguarding paper heritage in both technologically advanced and resource-limited settings.

Supplementary Materials

The source code supporting the findings of this study is openly available at GitHub: https://github.com/irhamniali/paper_deterioration_project.