Red Glue on Solder Pads: 5 Minutes to Replace 8 Hours of ROI Setup

Solder-pad contamination on PCBs — red glue overflow, component squeeze-out, dispenser splatter — leaves under 0.5 mm² of glue residue on the solder pad. It's a hidden killer. ICT passes it, the board ships looking fine, then weeks into the field a joint cracks open under vibration or thermal cycling.

This defect should be caught at AOI. But traditional rule-based AOI struggles here, and it's not because the detection algorithm isn't advanced enough — it's because the setup step is wildly expensive. Let's break it down.

1. The Real Pain Isn't the Algorithm — It's ROI Setup

The traditional AOI SOP for catching pad contamination: manually draw an ROI (region of interest) on every solder pad, then set the RGB threshold, area ratio, and binarization parameters inside that region.

The problem is scale. A medium-complexity PCBA has hundreds to thousands of pads. Drawing boxes one by one, tuning parameters, running trial passes — bringing up a single new board burns 8–16 engineer-hours. New panel, board revision, new product? Repeat the entire process.

What actually happens on the floor: most mid-size EMS plants resort to "set ROI on critical pads only, let the rest go" — escape risk gets baked straight into the system.

If the "set ROI" step could be automated, the underlying color-extraction algorithm (RGB threshold + area ratio) is actually a stable, reliable tool for red-glue detection.

2. Why Not Just Use Pure End-to-End AI?

So why not skip rules entirely — board in, all defects out?

Technically possible, but two real problems for the factory floor:

① Engineers lose control — model judgment is a black box. Want to tighten judgment slightly, or weight a high-risk zone? No handle to turn.

② Results aren't explainable — when audit asks "why was this board flagged NG?", "the AI said so" doesn't hold up.

Pure AI shines in "scenarios humans can't write rules for" — pad contamination isn't that case. The judgment logic is clear (non-metallic blob inside a pad region = NG); what needs solving is the upstream problem of "where is the pad region?"

3. Use AI Where It Hurts Most, Leave Detection to Classical

DaoAI's PCBA AOI now supports a pad-contamination module. The workflow splits into two stages:

Stage ①: AI auto-draws detection boxes

The model looks at the entire PCB once, automatically identifies the position of every solder pad, and draws all ROIs for you.

• New board onboarding: 8–16 hours → 5–10 minutes
• No coordinate memorization, no per-panel reset
• Engineers review and fine-tune the AI-drawn boxes — not start from zero
• No defect samples needed: this AI learns "what a pad looks like" (from PCB layout), not "what red glue looks like" — sidestepping the biggest data bottleneck of pure end-to-end AI

"Pad shapes vary too much for rules to cover" is exactly the kind of scenario where AI has a real advantage.

Stage ②: Color extraction does the detection

Inside each ROI, classical RGB threshold extraction + binarization + area ratio:

• Engineer picks the red-glue color range on the RGB triangle
• Sets the area ratio threshold (e.g., red-glue pixels > 10% of ROI = NG)
• Adjusts brightness binarization to match line lighting

For the engineer: readable, tunable, explainable. Reuses existing AOI operating experience — zero learning curve. For audit and SPC reports, every NG has a clear numerical basis.

4. Three Approaches Compared — Pad Contamination Scenario

"Auto-drawing the box" is where traditional AOI truly hits the wall (rules can't keep up, setup cost explodes) — using AI to break this bottleneck recovers the most engineer hours.

For pad contamination, pure rule-based setup is too expensive. AI's job is "where is the pad", not "what does red glue look like" — leave the latter to mature color-extraction algorithms, where engineers keep tuning parameters the way they already know how.