An existing JIVISION assembly-inspection case records 126 candidate image regions before filtering and 8 retained regions after filtering in the same verification scene. Based on the existing public record and its revision record, this article examines region generation, comparison on the original image, and the distinction between region counts and final decisions. It reports no new field tests.
Coarse detection and refinement in the original record
The case describes locating new differences at lower resolution, returning to original images to analyze candidate-region length, width, direction and contrast, and applying final rules. Surface masks and alignment also affect region generation. Before-and-after comparisons therefore need the same input, reference, recipe and processing chain. Counting boxes in unrelated screenshots is insufficient.
Separate coordinate spaces from decision stages
| Stage | Processing supported by the public record | Items to check during review |
|---|---|---|
| Difference localization | Locate new differences at lower resolution | Scale relationships and the associated input |
| Candidate-region analysis | Analyze region features on original images | Correspondence between contours, features and rules |
| Evidence display | Map coordinates back to the original captured image | Whether displayed positions match the actual anomaly |
| Final decision | Output pass, fail or pending review separately | Whether region processing is confused with final status |
The review items in the table are checks on the recorded processing chain. They do not establish that this article reran tests or obtained records for individual candidate image regions. The documented coordinate mapping makes the position of retained evidence a different question from the number of regions.
A count change does not establish accuracy
The recorded candidate-region counts describe a change in the set of image regions for the original scene. They do not establish that every excluded region was a false detection or every retained region was a real defect. Without labels for individual image regions and independent samples, do not calculate false-positive rates, missed-defect rates or accuracy, or assign excluded-region counts to individual rules.
Evidence needed for later retesting
Associate original images, reference and recipe versions, coordinate transforms, filtering reasons, retained candidate regions and final status in future tests. This is a proposed recording checklist, not a claim about existing delivered fields. The historical results remain limited to the recorded images, hardware, parameters and software. Independent batch samples, sustained operation and site-interface acceptance for production introduction remain pending verification.
Source: Existing assembly-inspection case, public record dated 2026-08-30. Related services: Vision algorithms and model engineering, project consultation.