Process guide / current prototype
How the inspection works
The software checks whether a barrel head can be inspected, then proposes the centres of the small flat metal pieces projecting from its outer rim. An operator confirms the proposals on the image.
First, the inspector finds the barrel centre and its circular rim. It checks all around that rim; a piece can appear at any clock position because the barrel may rotate between photographs.
The straight wood lines give a useful hint about the two possible ends where pieces tend to appear. The detector uses that hint to rank nearby groups, but the wood lines do not decide the position by themselves.
All reported angles are measured clockwise from the top of the original image: 12 o’clock is 0°, 3 o’clock is 90°, 6 o’clock is 180°, and 9 o’clock is 270°.
Make detail visible
Clearly dark or bright photos receive a restrained exposure correction. Moderately soft photos can receive mild sharpening. The original photo remains available for comparison.
Find the barrel
Wood-to-conveyor and inner rim edges locate the centre and rim. If stained wood makes that fit unreliable, a second pass follows the outer hoop edge. The inspector checks that the rim is sufficiently visible, inside the frame, and clear enough to inspect. Edge-fit results are sent for review.
Follow the outer rim
The circular search strip is straightened into a horizontal view. This keeps the search close to the actual edge, including when the barrel is not a perfect circle. A wider strip is available for pieces recessed into a worn rim or projecting beyond the usual edge.
Recognise the metal pieces
A small locally trained model scans overlapping views of the straightened rim. Its examples include clean barrels, stained barrels, and worn metal. When the first pair is uncertain, another model checks the same view, followed by the wider view if needed.
Check the group
Real pieces are expected close together, usually two to four within a short arc. The inspector compares both clock angle and distance from the rim, so close front and rear pieces are kept apart while repeated views of one piece are merged. Weak highlights need extra evidence, and a spaced tab can still be kept if a nearer false hit scores higher. Edge, wood-grain and conveyor checks provide fallback evidence when the models see too little; conveyor-derived positions are marked for review.
Report centres and rotation
Each proposed piece receives an image position and clockwise angle. The two highest-scoring pieces define a midpoint on the image. Its angle is measured clockwise from 12 o’clock; the reported clockwise rotation is the amount that would bring that midpoint to 12 o’clock. The overlay joins the selected pieces and marks their midpoint. If fewer than two pieces are found, no rotation is reported. Confirm the selected pieces before using the rotation.
| Suitable | The photo passed the image checks and the software found a plausible group. Confirm the piece centres by eye. |
|---|---|
| Check pieces | The barrel can be inspected, but the proposed pieces are uncertain or no confident group was found. The reason is shown beside the image. |
| Rejected | The barrel boundary, visibility, exposure, or clarity failed an image check. The reason is reported; the image is not scored for piece detection. |
| Clockwise rotation to 12 o’clock | The turn needed to move the midpoint of the two highest-scoring proposed pieces to the top of the image. This is a proposed alignment, dependent on both selected pieces being correct. |
This prototype combines conventional image checks with a small locally trained piece detector. Images are processed on this workstation; the inspector does not send them to a cloud service.