Industrial-Grade Quality Assurance Platform
This isn't just defect detection—it's a complete QA platform with business intelligence, database analytics, batch processing, and automated reporting.
Why manual inspection is failing modern electronics
Human operators spend 8+ hours daily inspecting circuit boards under magnification, checking for microscopic defects that could cause product failures.
After hours of repetitive work, concentration drops. Subtle defects like "Mouse Bites" or hairline "Opens" get missed.
Different operators have different judgment. This subjectivity creates quality variance across shifts.
When defective boards reach customers, the cost is catastrophic: recalls, warranty claims, and brand damage.
Eliminate human error through AI automation
Reduce inspection time: 2 min → 1.2 sec
Ensure 100% consistency with objective analysis
Three integrated systems working as one
Smart Alignment: Automatically corrects rotated/shifted boards using ORB feature matching
EfficientNetB0: Deep learning model identifies 6 distinct defect types with high precision
Smart Costing: Real-time calculation of repair costs ($1.50 - $11.25)
BER Logic: Beyond Economic Repair engine auto-flags SCRAP when cost > 75% value
SQLite Backend: Permanent storage with timestamps, defects, and costs
Batch Processing: Process 50+ boards simultaneously via ZIP upload
Unlike academic projects, this system answers critical business questions: "Should we repair this?" "How much will it cost?" "Is our yield improving?"
Solving the rotation & misalignment challenge
In real factories, PCBs are never perfectly aligned. Boards can be:
Simple image subtraction would show EVERY edge as a "defect".
ORB (Oriented FAST and Rotated BRIEF) finds stable keypoints (corners, edges) on both the template and test board.
These points define a mathematical transformation (Homography Matrix) that "warps" the test board to align perfectly.
❌ Without Alignment
Edges don't match, false positives
✓ With Alignment
Perfect overlay, only real defects
ROI (Region of Interest) Extraction Pipeline
Contains defects + dust + lighting variations
Clean defect candidates
Bounding boxes → AI input
From 92% to 97.8% accuracy through architectural design
🔒 Frozen Backbone
Feature extraction only
• Mouse Bite • Open Circuit
• Short • Spur
• Spurious • Missing Hole
Standard transfer learning struggled with similar defects (e.g., "Spur" vs "Spurious Copper") because simple output layers can only draw linear decision boundaries.
The Dense 256 layer provides computational capacity to learn non-linear boundaries. Dropout prevents memorization. Result: System distinguishes subtle geometric differences.
| Architecture | Validation Accuracy | Training Time | Status |
|---|---|---|---|
| Baseline (Simple Head) | 92.0% | 45 min | ❌ Insufficient |
| Heavy Head (Dense 256) | 97.8% | 52 min | ✓ Target Exceeded |
Solving the "Context Myopia" problem
The AI consistently misclassified "Open Circuits" (gap in trace) as "Shorts" (bridge between traces). With tight cropping, both looked like generic "blobs" without context.
AI sees only a blob. Can't see if it's a gap or bridge.
15% False Positive Rate.
AI gets wider context (64px padded). Sees surrounding traces.
0% False Positives.
UI Box: Tight crop (5px padding) for operator display.
AI Box: Forced min 64x64px + 20px padding for inference.
Result: +15% accuracy improvement on Open Circuits.
Beyond detection: automated business logic
Mouse Bite: $11.25
Open Circuit: $2.25
Short: $1.50
Missing Hole: SCRAP
Real-time labor/material calculation
IF Repair Cost > $37.50
(75% of board value)
OR "Missing Hole" detected
THEN Status = SCRAP
Prevents uneconomical repairs
Open → "Solder jumper wire"
Short → "Remove bridge with blade"
Mouse Bite → "File edge smooth"
Actionable technician instructions
From temporary prototype to permanent production system
Every inspection is permanently logged for ISO 9001 compliance and traceability.
| Field | Description |
|---|---|
| Timestamp | Exact date/time |
| Defect List | JSON list of classifications |
| Health Score | 0-100% quality rating |
| Status | PASS / FAIL / SCRAP |
Enables longitudinal analysis, trend detection, and serves as a legal record for quality audits.
Line chart visualizes Health Score over time. Alerts managers to process drift.
Bar charts show PASS/FAIL distribution. Critical for hitting production targets.
Calculates total financial impact of scrapped inventory.
Filterable logs exportable to CSV for external reporting.
From single-board testing to production-scale automation
Time for 50 boards: ~25 minutes
Effort: High (Manual clicks)
Risk: Easy to skip boards
Time for 50 boards: ~90 seconds
Effort: Minimal (One click)
Reliability: 100% processed
Enables end-of-shift quality reports. A manager can upload a day's production (200+ boards) and get a complete defect analysis and yield report in under 5 minutes.
From defect detection to professional certification in < 1 second
Instant creation of formal inspection certificates.
Traditional reporting: 15 mins/board
Our System: < 1 second/board
Board ID: PCB_001
Date: 2026-01-03
Defects: 2
Cost: $13.50
✓ PASS - REPAIR
| Open Circuit | $2.25 |
| Mouse Bite | $11.25 |
Rigorous testing on 590 held-out images
| Defect Type | Accuracy | Recall | Criticality |
|---|---|---|---|
| Open Circuit | 98.0% | 100% | 🔴 Critical |
| Short | 99.0% | 100% | 🔴 Critical |
| Mouse Bite | 96.8% | 96.8% | 🟡 Major |
| Missing Hole | 99.0% | 100% | 🔴 Critical |
The system never missed a defect that could cause electrical failure (Opens, Shorts). This is the key metric for safety-critical apps.
Human: 2 mins/board (30/hr)
AI: 1.2 secs/board (3000/hr)
Current limitations and planned enhancements
Extreme factory lighting variations can reduce subtraction accuracy. Fix: Histogram equalization.
CPU-based feature matching limits throughput to 1.2s. Fix: GPU/CUDA acceleration (Target: 0.3s).
Currently optimized for green solder masks. Fix: Multi-color dataset augmentation.
Connect directly to RTSP/GigE industrial cameras for live conveyor belt scanning (Zero-touch operation).
Migrate database to AWS/Azure for multi-factory analytics and centralized quality monitoring.
"Human-in-the-Loop" feedback. When operators correct the AI, the model retrains overnight.
From manual inspection to intelligent automation
Accuracy • Zero Critical Escapes
Consistent, objective analysis eliminating human fatigue.
Integrated financial logic and automated work orders.
Batch processing and persistent enterprise database.
Presented by
Atharva Mundke
Internship Project
GitHub: PCB-Defect-Detection-and-Classification/atharvam_1732