Pixel Checks Stop Processed Images From Altering Mandatory Text

By Nathan Brooks

At reduced resolution, the downloaded product images showed a mean absolute greyscale difference of well under one grey level from the team’s processed files.

On 15 September 2026, Gleb Tsipursky argued in Metrology News that quality leaders should “measure disagreements between automated inspection and qualified human judgment before reducing human review.”

Flick’s small case sat outside factory inspection but tested the same principle: an automated result gains trust only through an independent reference. Justin Flick writes for SOMUVA UG (haftungsbeschränkt) in Lüdenscheid, Germany, which runs the German-language vapechamp.de shop.

What must the image pipeline preserve?

Phone photographs of carton back panels needed to become clean, straightened images on white backgrounds for online product pages. Legally required manufacturer and importer names, addresses and warning texts sat alongside phone numbers and other contact details. Flick defined the measurand as the printed text, not the image’s appearance, making visual polish secondary to character fidelity.

Why did the generative output fail?

The first attempt used a generative image model. Given a straightened carton image and an instruction to preserve every printed character, number and symbol, it produced sharp print, even light and a clean background.

Comparison with the source found that “2801” had become “2001,” a phone number contained incorrect digits, and one comma became a full stop while another full stop became a comma. The plausible errors were difficult to notice at a glance, but they altered mandatory information.

The team discarded every generated image; none reached publication. A generative model can produce text that plausibly resembles the source without preserving what the source printed, and visual inspection cannot establish equivalence when both typography and errors look credible.

How did deterministic processing preserve the text?

A deterministic chain replaced the model. A perspective transform rectified the carton face from four corner points. The software estimated background illumination and divided it out to flatten uneven lighting, then white-balanced the image to a target tone.

Dividing by an estimated illumination field corrects uneven-light shading, but no numerical illumination target or tolerance appears in the case record; the settings and photographed light determine the result.

Gentle, edge-preserving denoising followed. Text blocks became protected regions in which pixels passed through background cleanup unchanged, and the processed face moved onto a white background with a soft shadow. The system redrew and re-typeset nothing; every final character came from photographed pixels.

What served as the independent reference?

For each image, the team kept a reference crop containing rectification, illumination correction and white balance, but no denoising or cleanup. They compared each processed face with that reference before adding the white background.

One principle was an OCR diff: an open-source engine with German and English language data read both images after upscaling, aligned the word sequences and listed every deviation.

The other was an ink check independent of reading. Pixels clearly darker than the carton background in the reference counted as print; if they became lighter in the output and most of a connected group disappeared, the check flagged a lost character. Together, the checks separated lost ink from OCR disagreement rather than reducing both to one similarity score.

Which processing steps failed the gate?

The gate showed that classical filters do not guarantee safety. A speckle filter turned “130” into “30” and removed colons; a threshold cleanup stage erased the dot in a web address.

The team dropped both stages because any operation that decides what counts as noise can classify punctuation as noise.

How much uncertainty remains in OCR?

The OCR diff had its own error. On small print, the engine misread the reference, including a “5” as an “S.” A raw diff could therefore blame processing for a reference error, producing false rejects; ignoring all differences could hide real changes and produce false accepts.

The team treated the work like a gauge study, separating measurement-system variation before assigning it to the image. The ink check and a separate review provided evidence independent of OCR.

Following Tsipursky’s third step, the team identified the source of each disagreement rather than assuming the model or person was wrong. Every disagreement became a stop for resolution, not a score to average away.

What checks preceded and followed publication?

A second automated review, independent of the processing chain, checked images in two rounds. In the first, a composite failed because one inserted line could not be verified reliably; the team removed it and used the carton’s genuine information block.

Before replacing an image, the team backed up the previous state for restoration. After publication, it downloaded the delivered files and compared them with its own versions. At reduced resolution, the mean absolute greyscale difference was well under one grey level.

The figure covers the reported comparison, not a general accuracy specification. Neither the reduced pixel dimensions nor the file count appears in the case record, so teams must establish their own tolerance and test conditions.

What should quality teams require?

The case supports nine release controls:

• Define the measurand: characters matter more than appearance in mandatory text.
• Keep a separate reference from the earliest reliable stage; never compare an output only with itself.
• Use two principles, such as an OCR diff and a pixel-level ink check.
• Characterise OCR errors on the reference before trusting its verdicts.
• Treat every disagreement as a stop, not a score to average away.
• Test classical filters as rigorously as AI tools.
• Add an independent second review; remove anything that cannot be verified.
• Back up before every change and pixel-compare delivered files.
• Use frontal, unwrapped, frame-filling photographs in daylight.

Generative models remain useful for images with no text requiring reading. For mandatory information, Flick’s team now uses classical processing plus a text comparison against the original.

Tsipursky’s closing test applies: “when the machine and a qualified person disagree, do we know who was right, why, and what we changed afterward?” For teams adopting automated image cleanup, the release question is the same: can every disputed glyph be traced to an independent reference, and does the delivered file still pass the final comparison?