Why weak-feature recovery fails in production

Every vision engineer has watched it happen: the connector inspection that ran clean for weeks starts missing pins. The pins are there. A human can see them. The tool cannot. The usual response — lower the threshold, widen the search region, raise the exposure — works for a shift and then breaks something else, because the underlying problem was never the threshold.

Weak pins and leads do not fail because the algorithm is too dumb. They fail because the signal is buried in lighting variation, clipping, neighbor overlap, oxidation, plating differences, and sensor noise — and because production teams need a defensible measurement, not just a found blob. A recovery strategy that cannot explain why it found a feature will eventually report a feature that is not there, and that failure mode is worse than a miss.

Why the naive fixes break

• Global thresholds assume the whole image shares one illumination state. High-density arrays never do — corner pins and center pins live in different lighting worlds.
• More gain or exposure rescues dim features by clipping bright ones, trading a visible failure for a silent one two rows away.
• Bigger search regions invite neighbor capture: the tool locks onto the adjacent pin and reports a plausible-looking, wrong position. No error is raised. The CSV looks fine.

The workflow that survives

The weak-lead recovery approach I run in production is CPU-based and deliberately boring:

• Local ROI per expected site. The nominal grid (from the drill drawing or CSV model) tells you where each pin should be, so each recovery runs in a small window with its own statistics.
• Contrast normalization inside the ROI. Every window is stretched to its own dynamic range before any decision is made — corner pins and center pins now compete on equal terms.
• Adaptive weak thresholding. The threshold derives from the ROI's own histogram, then the top-percent bright pixels are selected explicitly rather than hoping a fixed cut catches them.
• Weighted centroid, not blob center. Intensity-weighted centroids degrade gracefully as signal fades; binary blob centers jump.
• Confidence as a first-class output. Each recovery exports a confidence score, CSV fields, and an optional overlay graphic. Low-confidence hits route to review instead of silently passing.

The output is the product

The recovery math matters less than what leaves the tool. A weak-feature result that ships with measured position, deviation from nominal, confidence, and an overlay bitmap can be defended in a customer meeting, correlated against AOI, and audited after an RMA. That is the difference between an algorithm and a production system.

Field reality

When a connector inspection program moves to a new factory — Vietnam, Guadalajara, a new line in Thailand — weak-feature behavior is the first thing that changes. Different ambient light, different plating vendor, different operator handling. A workflow that exposes confidence and overlay evidence survives that transition and the escalations that follow. One that hides the decision behind a single pass/fail bit does not.

← Fixtureless grid inspection case study · All notes · Next: from CSV to acceptance →