Inspection archives
Inspection images for AI licensing
An inspection archive needs more than a picture count. Describe the inspected items, the decisions attached to them and the conditions under which the images were captured before discussing an AI licence.
Last checked
Count inspected items separately from images, and distinguish a recorded rejection from a verified defect.
What would a reviewer be trying to recognise?
Start with the intended task: identifying an unusual surface, locating a defect, reading a marking or checking an assembly. These are different questions. A folder labelled rejects does not explain which task its images support.
As a published research example, MVTec AD distinguishes defect-free training images from test images and provides annotations for anomalies. Its own terms restrict commercial use. It is a reference for understanding dataset structure, not material to copy into a sale, a Fairlode partner or evidence of current buyer demand.
For your archive, ask which decisions and supporting records actually exist. Do not commission a new labelling project until a relevant requirement and the cost of meeting it are clear.
Separate images, views and inspected items
Establish what one example represents. A single component may have several camera views, repeated captures or a video sequence. Those files should not automatically be described as independent inspected items.
Create an internal description of the grouping structure: item identifier, inspection run, camera position and any repeat inspection. State whether that relationship is recorded or would need to be reconstructed. Keep the identifiers themselves out of the initial enquiry.
An illustrative archive might contain several views of each part before and after rework. Without the grouping information, the before and after images could be mistaken for different parts. Ask the eventual technical reviewer how those groups would be kept together when preparing training and evaluation material; do not choose a split before the task is agreed.
Find out what each label proves
Ask who assigned each label, when and using which instructions. Separate an automated inspection result, an operator decision and a later confirmed finding. If an image was rejected because the camera view was unusable, it should not silently become an example of a defective product.
For a defect label, establish whether it applies to the whole image or a marked region. Check whether uncertain and disputed decisions have their own status. Record changes to the inspection criteria rather than presenting every year's labels as equivalent.
Use the dataset quality guide for the broader checks. Here the important question is narrow: can someone trace the label back to the inspection decision that produced it, and understand its limits?
Describe the capture conditions
Prepare a simple inventory of cameras, views, resolution, lighting arrangements and any image processing, using non-sensitive descriptions. Note changes in equipment or setup during the archive period. Record whether you retain original captures, compressed versions, cropped images or only screenshots of an inspection result.
Ask whether the visual appearance differs between product batches, inspection stations or time periods. A reviewer needs to know which conditions are represented before making claims about wider use. Avoid saying that an archive is representative unless you can explain of what, and how that was checked.
Do not improve images by silently replacing originals with sharpened, resized or AI-altered copies. Any agreed transformation should have a recorded purpose and an identifiable version.
List what is missing and what must stay out
Determine whether accepted items are retained as well as rejected ones. If the system saves only exceptions, say so. Record deleted periods, missing views and product types outside the proposed scope. Do not manufacture normal examples to fill a gap in the historical collection.
Review whether images or embedded metadata reveal people, customer identifiers, drawings, serial numbers, locations or confidential manufacturing methods. Ask the appropriate adviser to assess the permissions and restrictions before sharing anything. A cropped preview can still disclose sensitive details.
Keep regulatory inspection records in the systems and retention processes your business already uses. This article is about describing a possible collection, not changing a regulated workflow or authorising removal of inspection evidence.
Prepare the outline a reviewer can use
A first outline can cover the broad item category, inspection purpose, date range, approximate item and image counts, number of views, label types and known omissions. State whether originals and annotation histories exist. Use approximate bands only where they can be shared safely.
Include a clear uncertainty, such as unverified historical rejection reasons, instead of describing the archive as AI-ready. Read the provenance guide to organise where the images and labels came from without forwarding them.
Fairlode starts with descriptions, not image uploads. Do not attach photographs, screenshots or annotation files to establish interest. Use the existing readiness checklist, then ask whether there is a relevant requirement and what a separately approved review would involve. An interesting archive is not an offer or a valuation.
General information, not legal advice, a valuation or a buyer offer. Illustrative examples are not customer case studies.
Sources and further reading
Sources checked 22 October 2026. Historical research is background, not evidence of current buyer demand.