Unmixing ATR-{\mu}FTIR spectroscopic images of cross-sections of historical oil paintings

2026-06-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Heritage Science & Art Conservation · Depth: Expert, long

Summary

An unsupervised CNN autoencoder for blind unmixing of ATR-μFTIR hyperspectral images (HSIs) of historical oil painting cross-sections has been developed. This method, termed FTIR-unmixer, addresses the challenges of interpreting complex, heterogeneous, and degraded heritage science data. It introduces a weighted spectral angle distance (WSAD) loss function to mitigate sensitivity to atmospheric and acquisition artifacts across more than 1500 bands. WSAD employs automatic band-reliability weights derived from robust measures of spatial flatness, neighbor agreement, and spectral roughness. The FTIR-unmixer was successfully demonstrated on an ATR-μFTIR cross-section from the Ghent Altarpiece, recovering components like proteins, metal soaps, and calcium oxalates with improved spatial coherence and suppressed CO2-related residuals compared to standard spectral angle distance (SAD) loss.

Key takeaway

For heritage scientists analyzing complex ATR-μFTIR cross-sections, this method offers a robust, automated alternative to manual spectral interpretation. You can achieve more accurate material characterization and spatial mapping by implementing the FTIR-unmixer with its WSAD loss. This approach reduces bias from atmospheric artifacts and noise, providing clearer insights into material composition and degradation processes in historical artworks. Consider integrating this CNN autoencoder for improved efficiency and objectivity in your analytical workflow.

Key insights

An unsupervised CNN autoencoder with a weighted spectral angle distance loss effectively unmixes complex ATR-μFTIR data from historical paintings.

Principles

• Hyperspectral unmixing benefits from exploiting local spatial structure.
• Band reliability weighting improves interpretability in contamination-prone spectral regions.
• Data-driven weights can automatically suppress noise and acquisition artifacts.

Method

FTIR-unmixer uses a CNN autoencoder with a patch-wise linear mixing model. It employs a WSAD loss, calculating band-reliability weights from spatial flatness, neighbor agreement, and spectral roughness.

In practice

• Apply WSAD loss to hyperspectral data with unreliable or noisy bands.
• Use patch-based CNN autoencoders for spectral-spatial unmixing tasks.
• Evaluate endmember count by inspecting for merged or duplicate abundance maps.

Topics

• ATR-μFTIR Spectroscopy
• Hyperspectral Unmixing
• CNN Autoencoders
• Heritage Science
• Ghent Altarpiece
• Weighted Spectral Angle Distance

Code references

• ShivamP1993/ftir-unmixing

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Unregistered Spectral Image Fusion: Unmixing, Adversarial Learning, and Recoverability — An unsupervised framework simultaneously super-resolves unregistered hyperspectral and multispectral images, integrating unmixing and adversarial learning with theoretical recoverability guarantees.
• Case Western Reserve Researchers Using AI at Hair-Width Scale to Reveal Renaissance Master’s Hidden Hand | Newswise - Newswise — AI and physics reveal microscopic brushstroke patterns to attribute and authenticate Renaissance art.
• Diffusion Autoencoder for Unsupervised Artifact Restoration in Handheld Fundus Images — An unsupervised diffusion autoencoder restores fundus image artifacts by learning semantic representations from high-quality data.
• $${\bf{Micro}}{{\mathbb{S}}}{\bf{plit}}$$ Micro S plit : semantic unmixing of fluorescent microscopy data — MicroSplit enables simultaneous multi-structure imaging in one channel, providing denoised, unmixed outputs with uncertainty quantification.
• Optical Music Recognition for Real-World Manuscripts with Synthetic Data — Synthetic data, combined with minimal in-domain real data, significantly improves Optical Music Recognition for diverse real-world musical manuscripts.
• ATT-CR: Adaptive Triangular Transformer for Cloud Removal — Adaptive Triangular Transformer (ATT-CR) efficiently removes clouds from remote sensing images by combining linear triangular attention and feature-selective gating.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.