When product images are cleaned for an online catalogue, the visual result can look better while the printed information becomes less accurate. A case study by Justin Flick shows why that risk deserves the same measurement discipline used in inspection: for carton back panels carrying legally required names, addresses, warnings and phone numbers, the target was not a prettier image but unchanged text. The team called the printed text the measurand and treated every character as data, not decoration.
The first attempt used a generative image model. It was given a straightened carton image and told to preserve every character, number and symbol. The output looked professional, with sharp print and even light, but comparison with the source found substantive changes: 2801 became 2001, a phone number contained wrong digits, and commas and full stops were swapped. These were plausible errors, easy to miss in visual review. Because the information was mandatory, the images were discarded and none were published. The lesson is not that generative models are uniquely careless; it is that a model optimised for plausible appearance has no inherent obligation to reproduce the printed glyph.
Why visual inspection is a weak reference
Human review is necessary but not sufficient. A reviewer looking at a clean image tends to judge whether it looks like a carton label, not whether each character matches the photographed original. Plausible substitutions survive that check. The case therefore separated two questions: does the image look clean, and does the text remain faithful? Only the second question could decide publication. This is the same logic Tsipursky applies to automated inspection: quality leaders should quantify where machine output and qualified human judgment disagree before reducing human review. In image processing, the qualified human judgment cannot be a glance at the final file; it must be a comparison against an independent reference.
A deterministic chain, with its own failure modes
The replacement process was deterministic. The carton face was rectified with a perspective transform from four corner points. Uneven lighting was estimated and divided out, then white-balanced to a target tone. Denoising was gentle and edge-preserving. Text blocks were marked as protected regions so their pixels passed through background clean-up unchanged. The result was placed on a white background with a soft shadow. Nothing was redrawn or re-typeset; every character came from photographed pixels.
That does not make classical processing automatically safe. The gate found that a speckle filter changed 130 to 30 and removed colons, while a threshold clean-up step erased the dot in a web address. Both stages were dropped. The operational rule is blunt: any step that decides what counts as noise can also decide that a full stop is noise. Classical filters need the same adversarial testing as AI tools.
Two measurement principles and a stop rule
For each image, the team kept a reference made from the same crop after rectification, illumination correction and white balance only, before denoising or clean-up. Each processed carton face was compared with that reference before it was placed on the background. The comparison used two principles. First, an OCR diff read both images with an open-source engine using German and English data after upscaling, aligned the word sequences and listed deviations. Second, an ink check ignored reading entirely: pixels clearly darker than the carton background in the reference counted as print, and a connected group that became light in the processed image was flagged as a lost character.
The OCR diff is itself a measurement system with error. On small print it misread the reference, for example taking a 5 for an S. A raw diff therefore reports disagreements that processing did not cause. If taken at face value, those become false rejects; if ignored wholesale, real changes become false accepts. The team used a gauge-study analogy: separate the measurement system's variation from the part's variation before attributing a difference to the process. The independent ink check and a separate review step helped identify the source of each disagreement. The governing principle was to treat every disagreement as a stop to be resolved, not as a similarity score to be averaged away.
Governance after the gate
The control chain did not end with image processing. A separate automated review, independent of the processing chain, checked the images in two rounds. In the first round, a composite image failed because one inserted line could not be verified reliably; it was removed and the genuine information block from the carton was used instead. Before any image was replaced, the prior state was backed up so it could be restored. After publication, the delivered files were downloaded and compared with the team's own copies. At reduced resolution, the mean absolute greyscale difference was well under one grey level. The published file is the last link in the chain and received the same scrutiny as earlier steps.
What quality teams can reuse
- Define the measurand. When an image carries mandatory text, the characters matter more than the appearance of the image.
- Keep a separate reference from the earliest reliable processing step. Never compare an output only with itself.
- Use at least two measurement principles, such as an OCR text diff and a pixel-level ink check.
- Characterise OCR error on the reference before trusting its verdicts.
- Treat any disagreement as a stop, not as a score to be averaged away.
- Test classical filters as rigorously as AI tools.
- Add an independent second review and remove anything that cannot be verified.
- Back up before every change and pixel-compare the published files.
- Improve the input: frontal, unwrapped, frame-filling photographs in daylight make later steps easier.
The boundary matters. Generative models remain useful when nothing in the image has to be read. For images carrying mandatory information, the case supports classical processing plus a text comparison against the original. That is not a claim that one pipeline fits every product category; it is a demonstration that the reference, the second principle and the stop rule are what make the result auditable.
Tsipursky's question for quality leaders is whether, when machine and qualified person disagree, the organisation knows who was right, why, and what changed afterward. Processed images need the same answer before publication. A clean background is not evidence of faithful text. An independent reference, a disagreement log and a verifiable final file are.
