How to read faded or erased text in manuscript photographs

An ordinary color photograph may preserve tonal or color differences that make faint writing easier to inspect after processing. It may also omit information present under ultraviolet, infrared, transmitted, raking, or multispectral illumination. Processing an RGB photograph cannot establish that erased text, palimpsest undertext, or underdrawing survives, and it cannot prove that missing information is lost.

Start with channel isolation

An RGB photograph contains red, green, and blue channel data. A mark with weak contrast in the composite image may have stronger contrast in one channel. The best channel depends on the ink, support, illumination, camera response, and color management. Channel inspection can show evidence captured by that photograph; it cannot show whether useful evidence exists outside the captured spectral bands.

Levels and gamma: stretch what's there

Faded text usually occupies a narrow band of tones: pale brown on cream. Levels clips the histogram to that band and stretches it across the full range; gamma bends the midtones where the ink lives. This is the workhorse for text that is faint but present. If levels on the best single channel doesn't bring it up, move to the stronger tools.

Decorrelation stretch: exaggerate color differences

Decorrelation stretch transforms correlated color channels so small color differences become easier to see. It can also magnify noise and create color boundaries that do not correspond to writing. Its assumptions may fit some photographs poorly. Research on image processing for papyri stresses that scholarly derivatives must be checked for artifacts and interpreted against the source image.

CLAHE: local contrast for uneven pages

Contrast-limited adaptive histogram equalization, or CLAHE, adjusts local contrast in small regions. It may help with uneven illumination or low local contrast, but it may also magnify noise, halos, stains, compression artifacts, and texture. A stronger pattern in a processed view is a lead for inspection, not proof of a letter or underdrawing.

Doing this in practice

All four techniques are available piecemeal: channel work and levels in any photo editor, decorrelation stretch in ImageJ with the DStretch plugin, CLAHE in ImageJ or Python's OpenCV. Recto/Verso puts all of them in the cataloging app itself: open the folio, apply channel isolation, levels, gamma, decorrelation stretch, or CLAHE with live preview, and save the processed version as a separate copy attached to that folio without replacing the source photograph's pixel data. The adjusted image stays with the folio's record rather than in a loose Downloads file.

Failure in an RGB derivative does not show that the underlying signal is gone. Library of Congress spectral imaging work uses different wavelengths and illumination geometries because they can record information absent from an ordinary color photograph. Consult the repository's imaging or conservation staff before requesting special-light examination or drawing conclusions from processed images.