Cutting the Polished Region Before Any Algorithm Decides
A machined cylinder block arrives on the line conveyor and moves on once released from it. What has to be checked is not one kind of thing: hole position and size are geometric, porosity and voids on the polished face are appearance defects, scratches on the seal band are thin weak features, and a missing operation is a question of whether a step was performed. Four different classes on one station, covered by one algorithm, usually ends with one of them sacrificed.
Decide where to look before deciding how to judge
The first step is not an algorithm but extracting the polished-face inspection region. A block has both polished and unpolished surfaces, and the cast texture and structure of the unpolished area produce a large number of defect-like disturbances in the image. Without limiting the search to polished faces, porosity detection faces false defects across the whole frame and the false-call rate cannot be brought down. The order matters: region first, algorithm second. Reversed, it just hands noise to the model.
Geometry to classic algorithms, appearance to a model
Hole position and size are explicit geometric items, and classic vision is stable and explainable: the decision can be measured and stated. Porosity, voids and seal-band scratches vary in form, and rules cannot express them fully, so deep learning trained on samples is more effective. The split follows the nature of the question, not its difficulty: deep learning is not used everywhere because it is stronger, and for geometric quantities the classic approach is more reliable.
Template matching on the photographed face locates the inspection regions so they stay consistent across workpiece poses. That is the common input for both algorithm families.
Glare on a polished face is an angle problem
Polished surfaces are strongly reflective, and under ordinary lighting porosity and scratches are masked by specular reflection. The defect is not absent from the image; that area is simply blown out and carries no information. High-brightness light with controlled capture parameters brings small defects back into the image.
The key is the pairing of light and viewing angle. Simply raising brightness makes the reflection stronger; changing the incidence angle moves the specular lobe out of the camera's view. A better model cannot repair that.
45 seconds across several faces
The cycle is 45 seconds and covers all inspected faces, subject to site acceptance results. Hole and defect detection run in parallel inside their own regions, with capture and computation pipelined. Multi-face layout has to reconcile shooting angle with occlusion of the inspection regions, which is why face-by-face imaging is more common than one panorama: a single shot always has a face hidden behind something.
Archiving by model and date, so a query is possible
Inspection images are stored by model and date. With many block variants and batches, a quality dispute has to retrieve the images from that model and date rather than searching a pile of records. Archiving rules follow the inspection recipe and switch automatically on a mixed line, so every image matches its model and production date. This is often treated as an extra feature; in practice it is what closes the traceability chain.
Where both algorithm families struggle
Porosity and voids sometimes look alike: porosity is a point depression, a void is a missing edge or area. Deep learning separates them from labelled samples, and difficult cases can be rechecked against a regional template. That shows the split is not a cure-all. The hard part is not which algorithm handles which defect but separating the two on borderline samples, and that is solved with labelling quality and a recheck mechanism.
Questions on the division
Why not deep learning for everything?
Hole position and size must be quantifiable and explainable, where the classic method is more stable. Deep learning earns its place on variable defect morphology.
Can region extraction exclude real defects?
It can if the boundary is wrong. Region extraction follows the actual polished contour, and defects near the boundary may fall outside, so boundary cases are checked explicitly at acceptance.
What changes with another model?
Region templates, capture points and archiving rules. The algorithms stay; the area they work in and the parameters change.