In a recorded JIVISION project for inspection of assembled electronic parts, expanding the scratch inspection area also introduced candidates from board edges, traces, pads, holes and image boundaries. The review therefore examines references, part position and the inspection surface before interpreting counts of detected regions and final decisions. This article draws on the published project report and its project records; it adds no new onsite test conclusion.
How the problem emerged
An early test classified a scratched onsite sample as acceptable. The original project report records similar abnormal texture in the reference image and rules covering only a small fixed region. After rebuilding the reference and expanding coverage, scratches entered the candidate set, but normal structures also produced responses. Changing one threshold could not explain reference contamination, part movement and non-inspection regions together.
Examine structural texture separately
| Item | Problem identified in this project | Review direction |
|---|---|---|
| Reference image | The reference contained the target abnormality | Build references from agreed defect-free samples and record versions |
| Part position | Translation and scale differed from the reference | Check localization, registration and valid overlap first |
| Inspection surface | Edges, traces and holes entered the candidate set | Define surface masks and excluded regions from the task |
| Decision chain | Candidate counts and final quality decisions had different meanings | Record detected regions, filtering, evidence and rule outputs separately |
These checks correspond to the original project report's reference management, registration, surface masks and quality gating. The inspection task determines whether structural regions may be excluded. Do not mask a region containing required defects merely to reduce false calls.
Recorded staged results
The original project report records the following observations for its images, parameters, hardware and software configuration. Filtering of detected regions, reference-image behaviour and the specified defective sample are separate observations. They do not establish accuracy, a miss rate or performance across other products.
| Recorded item | Staged result |
|---|---|
| Possible anomaly regions before filtering | 126 |
| Possible anomaly regions after filtering | 8 |
| Number of reference images | 3 |
| Detected scratch regions in each reference image | 0 |
| Specified onsite scratched sample | Reject decision |
No new tests were performed for this article, and the original conclusions are unchanged. Production introduction still requires independent batch samples, sustained operation and site-interface acceptance. Different textures, defect definitions and mounting conditions require new samples and verification methods.
Inputs needed to apply the lesson
Provide raw normal and defective images, references, product versions, inspection regions, pose variation and quality rules. Record absent detections, filtered regions, misplaced evidence and incorrect final decisions separately to identify the stage responsible.
Source: report on inspection of assembled electronic parts, public record dated 2026-08-30. Related services: quality inspection and process error prevention and project enquiries.