⚡ electrical engineering

Infrared Transparency as a Discriminative Signal and Task-Specific Spectral Ranges for Ink and Substrate Analysis in Historical Manuscripts

This study demonstrates that infrared transparency serves as a key discriminative signal for historical manuscript analysis, revealing that short-wave infrared ranges are far more effective than visible light for ink discrimination while substrate characterization remains robust across all spectral bands, thereby supporting task-specific selection of spectral coverage.

Original authors: Osama Abdelaziz, Younes Akbari, Somaya Al-Maadeed, Kishor Kumar Sadasivuni, Sherine El-Menshawy

Published 2026-08-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Osama Abdelaziz, Younes Akbari, Somaya Al-Maadeed, Kishor Kumar Sadasivuni, Sherine El-Menshawy

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

History is often written in ink that has faded, stained, or been obscured by the passage of centuries. For conservators and historians, knowing exactly what kind of ink was used on a fragile manuscript is not just a matter of curiosity; it is a key to understanding when a document was made, who wrote it, and how it should be preserved. Different inks age in different ways, and some, like iron gall ink, can actually eat away at the paper or parchment they sit on. To study these materials without damaging them, scientists use a technique called hyperspectral imaging. Instead of seeing an image with just three colors like a standard camera, this technology captures a full spectrum of light at every single point on the page. It is like taking a photograph where every pixel contains a detailed chemical fingerprint of the material beneath it. The challenge has always been deciding which parts of the light spectrum are actually useful. Cameras can see visible light, near-infrared light, and short-wave infrared light, but no single camera sees all of them at once. Researchers have long suspected that different types of light reveal different secrets, but until now, there was no clear proof of which specific range of light is best for identifying historical inks versus identifying the paper or parchment they are written on.

A team of researchers at Qatar University set out to solve this puzzle using a public collection of historical documents and test samples known as the HYPERDOC dataset. They wanted to find out if they could strip away large portions of the light spectrum and still get accurate results, or if they needed the full range of data. To do this, they built a computer system capable of reading these light spectra and trained it to distinguish between three common types of historical ink: iron gall, carbon-based, and sepia. They also asked the system to identify the background material, such as parchment or cotton-linen paper. The researchers did not just guess; they ran the exact same computer model nine different times, each time feeding it a different slice of the light spectrum. Some runs used only the visible colors we see with our eyes, others used only the near-infrared, and some used the short-wave infrared or a combination of them. They ensured their test was fair by splitting the data so that the computer was tested on documents it had never seen before, preventing it from simply memorizing the answers.

The results were surprisingly clear and overturned a common assumption in the field. The researchers found that the visible part of the spectrum, the colors we can see, provided almost no extra help in telling the inks apart. When they removed the visible light from their analysis, the computer's ability to identify the ink barely changed. In fact, the near-infrared and short-wave infrared ranges were far more powerful. The short-wave infrared alone was able to identify the inks with a high degree of accuracy, far outperforming the visible light. The reason for this lies in how these inks interact with light. As the light waves get longer and move into the infrared, iron gall ink begins to become transparent, allowing the sensor to see the paper underneath it. Carbon and sepia inks, however, remain dark and opaque, continuing to block the light. This difference in transparency is the key signal the computer uses to tell them apart. The visible light, by contrast, sees all three inks as dark and similar, offering no way to distinguish them.

The study also revealed that the best way to analyze these documents is not to use a single camera that sees everything, but to combine two specific ranges of light. By merging the near-infrared and short-wave infrared data, the researchers achieved the highest accuracy, correctly identifying the ink and the background material in nearly every case. This combination allowed the system to see the full picture: where the ink was opaque and where it had become transparent. Interestingly, the researchers discovered that the light needed to identify the ink is not the same as the light needed to identify the paper. While the ink classification relied heavily on the specific behavior of the ink in the infrared, the paper and parchment were easy to tell apart across the entire range of light they tested. This suggests that the tools used to study these documents should be chosen based on the specific question being asked, rather than just using whatever equipment is available.

This work provides a practical guide for future conservation efforts. It shows that for identifying historical inks, expensive equipment that captures visible light is largely unnecessary if the goal is simply to classify the ink type. Instead, focusing on the infrared spectrum offers a clearer, more reliable signal. The researchers confirmed that their method works well on the specific types of dark inks found in their dataset, achieving a level of accuracy that matches or exceeds previous methods that used more complex or different approaches. However, they noted that their findings are specific to these dark inks and the light range they measured. They did not test ultraviolet or mid-infrared light, nor did they test colored inks or pigments. The study concludes that while the visible spectrum is redundant for this specific task, the infrared spectrum holds the critical information needed to read the chemical history of a manuscript without ever touching it.

Technical Summary: Infrared Transparency as a Discriminative Signal and Task-Specific Spectral Ranges for Ink and Substrate Analysis in Historical Manuscripts

Problem Statement
The identification of ink types in historical manuscripts is critical for dating, attribution, and conservation, as different inks (e.g., iron gall vs. carbon-based) deteriorate differently. While hyperspectral imaging (HSI) offers non-destructive analysis, the optimal spectral range for distinguishing between spectrally similar dark inks (iron gall, carbon, and sepia) remains an open question. Furthermore, it is unclear whether the spectral ranges best suited for ink classification are the same as those required for characterizing the substrate (parchment vs. paper). Conventional practice often splits acquisition into Visible/Near-Infrared (VNIR) and Short-Wave Infrared (SWIR) for different purposes, but this division has not been quantitatively validated for historical ink discrimination, where materials are often indistinguishable in the visible spectrum.

Methodology
The study utilizes the public HYPERDOC dataset, comprising spatially aligned VNIR (400–1000 nm, 121 bands) and SWIR (900–1700 nm, 161 bands) images from 61 historical documents and mock-up samples. The experimental design involves:

Spectral Configuration Comparison: The authors trained an identical Residual U-Net segmentation model on nine distinct spectral configurations spanning 400–1700 nm. These configurations isolated specific ranges (VIS, NIR, SWIR-Low, SWIR-High) and combinations (fused inputs) to determine which bands contribute discriminative information.
Data Partitioning: To ensure realistic evaluation, data was split at the document level (72% train, 15% validation, 13% test), ensuring no pixels from the same document appeared in both training and testing sets. Each configuration was run five times with different random seeds to report mean performance and standard deviation.
Physical Analysis: Beyond model training, the authors measured ink–substrate contrast ( $C_c(\lambda)$ ) directly as a function of wavelength to understand the physical mechanisms driving classification. They also evaluated substrate separability (parchment vs. cotton-linen) using model-free metrics (Cohen's d and AUC).
Task Definitions: The primary task was a four-class segmentation (Iron Gall, Carbon, Sepia, and Other/Non-ink). A secondary experiment reproduced the three-class benchmark (Metallo-gallate, Carbon-containing, Non-carbon-containing) from previous literature to allow direct comparison.

Key Results

Redundancy of the Visible Range: The visible spectrum (400–700 nm) provided negligible additional information for ink discrimination. Removing it from a fused input changed the macro F1 score by only 0.009. In contrast, SWIR data alone achieved a macro F1 of 0.920, significantly outperforming visible data (0.660) and NIR data (0.791).
SWIR Superiority and Band Count: The advantage of SWIR was attributed to spectral content rather than the number of bands. A subsampled SWIR configuration (SWIR-61, matching the band count of NIR) achieved a macro F1 of 0.914, retaining 95% of the full SWIR advantage.
Optimal Configuration: The best performance was achieved by fusing NIR and SWIR (700–1700 nm), yielding 0.977 accuracy and 0.941 macro F1. This outperformed both individual spectral models and their weighted soft-voting combination, suggesting that learning joint spectral-spatial features is more effective than post-hoc fusion.
Physical Mechanism (Differential Transparency): The study identified differential infrared transparency as the primary discriminative signal. Iron gall ink progressively loses contrast and becomes transparent to the sensor as wavelength increases, reaching the noise floor at approximately 1695–1700 nm. Conversely, carbon and sepia inks remain measurably opaque throughout the 400–1700 nm range. Consequently, the classification of iron gall relies on detecting its "disappearance" rather than its color.
Substrate vs. Ink Discrimination: Unlike ink classification, which relies on specific infrared extinction behaviors, substrate characterization (parchment vs. cotton-linen) showed strong separability across the entire measured range (median Cohen's d = 2.04). This indicates that ink and substrate analysis rely on different, task-specific spectral information.
Benchmark Comparison: Under a matched document-level split, the proposed model outperformed four of five traditional machine learning baselines (KNN, LDA, RF, PLS-DA) and matched SVM on the three-class benchmark, though it remained below the performance of DeepLabV3 reported in the original HYPERDOC study.

Significance and Claims
The paper claims that spectral coverage for heritage imaging should be selected based on the analytical objective rather than acquired uniformly. The results demonstrate that:

Task-Specific Ranges: Ink identification in historical manuscripts is driven by SWIR data and the physical phenomenon of differential transparency, rendering the visible range redundant for this specific task.
Physical Interpretability: The classification success is directly linked to the physical extinction behavior of iron gall ink, providing a measurable optical basis for the "disappearance" signal used by the model.
Methodological Rigor: By holding the model architecture fixed and varying only spectral inputs under strict document-level splits, the study isolates the contribution of spectral bands, offering a clear guide for instrument design and data acquisition strategies in heritage science.

The authors note limitations, including the use of cropped minicubes rather than full pages, the exclusion of colored inks, and the lack of ultraviolet or mid-infrared data, leaving these areas for future investigation.

More like this