The existing JIVISION assembly-inspection case records three reference images classified as acceptable with zero scratch candidates, and a site sample with a diagonal scratch classified as defective. This review uses the published case and the revision record dated 2026-08-30 to explain what these inputs establish. It adds no site tests or sample statistics.
Reference images and evaluation samples serve different purposes
The original case found that a reference image contained the same abnormal texture as the site workpiece, causing a real scratch to be classified as acceptable. Subsequent work rebuilt references without the target anomaly and changed the inspection area and processing chain. Check the defect status of reference images themselves. Images used to configure a system do not automatically constitute independent evaluation samples.
Group available evidence and missing evidence separately
| Input group | What the original record supports | Conclusion not established |
|---|---|---|
| Three reference images | Acceptable results with zero scratch candidates under the original conditions | No false calls on unseen acceptable production batches |
| Site scratch sample | Defective result for the original diagonal-scratch sample | Coverage of every defect type and severity |
| Independent acceptable and defective samples | Required as further verification before production introduction | Completed independent batch-evaluation results |
| Boundary cases and invalid inputs | Quality gating and pending-review handling described in the case | Completed statistical verification of all boundary cases |
The grouping explains existing evidence and further verification needs. It does not report possession of a complete sample inventory, per-image labels or a new training and evaluation split. The original record does not provide independent sample counts and per-sample results sufficient to calculate overall accuracy.
Avoid mixing images from the same source in later evaluation
Later work can identify samples by workpiece, capture batch, recipe and image origin, distinguishing reference creation, parameter tuning and independent evaluation. Crops, augmented images or repeated exports of one original do not automatically constitute independent samples. This is a recommendation for further verification, not a sample-separation procedure already documented for the original project.
Keep samples associated with states in statistics
Before evaluation, agree on defect classes, visibility, labeling methods, acceptable and defective sample counts, and how pending-review and invalid inputs enter statistics. Record original images, reference versions, rule versions and results separately. Existing results remain limited to the original images, hardware, settings and software. Independent batch samples, continuous operation and site-interface acceptance remain pending verification. This article reports no new accuracy, missed-defect rate or production-acceptance result.
Fact source: Existing assembly-inspection case. Related services: Vision algorithms and model engineering, project consultation.