A German online retailer’s packaging-photo project has produced a practical test of a quality principle: an automated image result is only as reliable as the independent reference used to check it. The case, described by Justin Flick of SOMUVA UG, focused on carton back panels carrying mandatory product information. The aim was to turn phone photographs into clean, straightened images on a white background for product pages without altering any printed detail.
Clean appearance, unchanged information
The panels include manufacturer and importer names and addresses, warning statements and contact details such as telephone numbers. Improving sharpness, lighting and background was acceptable; changing the information was not. In measurement terms, the printed text was the measurand, not the visual appearance of the image.
Generative model changed plausible details
The first attempt used a generative image model. The team supplied a straightened carton image and instructed the model to keep every character, number and symbol unchanged. The result looked professional, with crisp print, even lighting and a clean background. Comparison with the source revealed errors: the digits “2801” had become “2001”, a phone number contained incorrect digits, and commas and full stops had been exchanged in places. Because the changes were plausible, they were difficult to catch by eye. The images were discarded and none was published. The team concluded that a generative model tends to create text that looks plausible rather than reproducing the printed characters, and that visual inspection is too weak to serve as the reference for mandatory text.
Deterministic processing keeps photographed pixels
The replacement was a deterministic processing chain. The carton face is rectified with a perspective transform from four corner points. Uneven illumination is estimated and divided out, then white balance is adjusted to a target tone. Denoising is gentle and edge-preserving. Text areas are marked as protected regions so their pixels pass through background clean-up unchanged. The processed panel is then placed on a white background with a soft shadow. Nothing on the carton is redrawn or re-typeset; every printed character in the final image comes from the photographed pixels.
Two independent checks against a reference
For each image, the team kept a reference made only from rectification plus illumination and white-balance correction, before denoising or clean-up. Each processed carton face was compared with this reference before background placement. The first check was an OCR diff: both images were upscaled and read with an open-source OCR engine using German and English language data, then word sequences were aligned and deviations listed. The second was an ink check that did not depend on reading. Pixels clearly darker than the carton background in the reference were treated as print; if such pixels became light in the processed image and most of a connected group disappeared, the group was flagged as a possible lost character.
The gate also showed that conventional filters are not automatically safe. 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 team noted that a filter defining what counts as noise may classify a punctuation mark as noise.
OCR has its own measurement uncertainty
OCR itself introduced uncertainty. On small print, it misread the reference image, for example taking a “5” for an “S”. A raw OCR diff therefore reported disagreements that processing had not caused. Taken at face value, such errors could produce false rejects; ignored wholesale, they could hide real changes and produce false accepts. The team compared this with a gauge study, where variation from the measurement system must be separated before variation is attributed to the part. The ink check and a separate review step helped identify the source of disagreements. The working rule was that disagreements should halt release until resolved rather than being averaged into a similarity score. This follows the cited advice from Gleb Tsipursky to identify the source of a disagreement rather than assuming the person or model must be wrong.
Publication checks and practical lessons
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; the team removed the composite and used the genuine information block from the carton. Previous states were backed up before replacement. 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 delivered file is the final output and was checked as rigorously as every earlier step.
For quality teams, the practical points are to define the measurand, keep a separate reference from the earliest reliable processing stage, use more than one measurement principle, characterise OCR errors on the reference, test classical filters as rigorously as AI tools, add an independent second review, back up before changes and compare published files. Better input also helps: frontal, unwrapped, frame-filling photographs taken in daylight make later processing easier. Generative models remain useful where no text has to be read. For images carrying mandatory information, the team now uses classical processing plus a text comparison against the original. Tsipursky’s question for quality leaders—whether they know who was right, why, and what changed after a machine-human disagreement—applies equally when a processed image and its reference disagree.
