Covering Six Faces and the Inside With Fewer Capture Points
A cylinder block has to be checked on six external faces, the deck, bottom, cold side, thrust side, gear side and flywheel side, plus the internal cavity. The part is large and arrives on a conveyor; good parts are released and defective ones blocked. A person walking six faces plus the cavity within one cycle is not realistic, fine defects are missed at a high rate, and the check leaves no structured data behind.
Six faces plus the inside: count points before resolution
The first account in multi-face coverage is the number of capture points: how many views each face needs, which direction enters the cavity, and how adjacent views overlap. Once the point count is fixed the cycle budget has somewhere to land; if the count cannot be reduced, higher resolution will not make the coverage finish. The fisheye lens reduces the point count here: a wide angle significantly enlarges what one capture covers, which is what makes six faces plus the cavity feasible within a 90-second cycle, subject to site acceptance results.
Fisheye trades point count for edge distortion
The wide angle is bought with edge distortion. The closer to the edge of a fisheye image, the more pronounced the distortion, and a defect of the same size presents different pixel morphology at different positions. That has two consequences: defect size has to be interpreted with position compensation, and as the point count falls, each point carries more responsibility, so imaging quality in edge regions must be verified separately. A buyer who sees fewer points and a faster cycle but overlooks edge consistency will find commissioning painful.
Segmentation prevents misses, the second pass prevents unfair calls
Deep-learning instance segmentation identifies and classifies defect regions pixel by pixel. Its strength is sensitivity: suspicious regions are almost all found. High sensitivity also means more false calls, so a classic algorithm performs a second precise judgment on each suspect region. The division is clear: segmentation keeps misses controlled, the second pass keeps false calls down. Thresholds in both stages are tuned continuously against site defect samples rather than fixed once.
How many pixels is 0.4 mm
Detecting a black spot as small as 0.4 mm rests on a physical premise: at that size the defect must occupy several pixels in the image before a segmentation network can separate it. Camera resolution is therefore a hard constraint, not a margin item. Field of view, working distance and pixel count together decide whether 0.4 mm can be seen, and finer defect requirements raise the resolution requirement. No algorithm gets around that.
One block, one record
Inspection data binds to the block's QR code and uploads to MES, with defect type, position and verdict stored in structured form. Every block's result is tied to its identity, giving a traceable one-block-one-record file that can be searched by block and supports batch traceability and process improvement. The value of structure shows up later: a pass or fail record cannot support defect distribution analysis, while a record with type and position can answer which defect keeps appearing in which region.
The metrics only mean something together
Overall accuracy above 99.5 percent, defect interception above 99 percent, escape rate no more than 1 percent, and stable operation around the clock, subject to site acceptance results. The three belong together: high interception means defective parts are largely stopped, low escape means few problems pass, and accuracy reflects overall reliability. Projects where one metric looks good while the other two fail are not rare in defect detection.
Questions on coverage and false calls
Does a fisheye make defects harder to judge?
In edge regions it does, so position compensation is needed and edge imaging is verified separately at acceptance. The centre is less affected.
Why keep a classic algorithm for a second pass?
Because segmentation sensitivity is bought with false calls. The second pass reviews suspect regions against rules, and the combination holds misses and false calls at once.
Is fisheye mandatory for six faces?
No. Conventional lenses with more points can also cover the part, at the cost of points and cycle time. A fisheye buys coverage with fewer points, which suits tight cycles.