Home
Case Studies
Quality Control with Computer Vision

Manufacturing Computer Vision

How QualityVision Achieved 99.5% Defect Detection Accuracy While Inspecting 500 Parts Per Minute with Deep Learning

An advanced computer vision system leveraging deep learning and edge computing to automate quality inspection of manufactured components, detecting surface defects, dimensional variations, and assembly errors with superhuman accuracy at production line speeds.

Client: QualityVision

https://QualityVision.xcodeit.io/dashboard

VisionQC AI

Computer Vision

Defect Detection

99.2%

Manual QC

-60%

False Positives

<0.5%

Units/Hour

1,200

Live Inspection - Camera 01

Defect detected

REC

60 FPS

Model: DefectNet v2.1

Recent Detections

#45821 Pass

#45822 Defect

#45823 Pass

#45824 Pass

Today's Pass Rate 98.7%

99.5%

Defect Detection Accuracy

500

Parts Per Minute Inspection

80%

Fewer Customer Complaints

6 Months

Return on Investment

The Challenge

QualityVision, a Tier 1 automotive supplier producing injection-molded plastic components, faced a quality crisis that threatened major customer relationships. Their manual visual inspection process, relying on human inspectors examining parts under magnifying lights, was simultaneously too slow and too inconsistent. With production lines running at 500 parts per minute, inspectors could only sample-check 10% of output, meaning 90% of parts shipped without inspection. Even for the sampled parts, human inspection accuracy averaged only 92%—inspectors fatigued after hours of repetitive scrutiny, missed subtle defects like micro-cracks or slight color variations, and applied inconsistent standards. This resulted in a steady stream of customer complaints: 200 field failures per month traced back to defects that should have been caught at inspection. Major automotive OEM customers were threatening to pull contracts if quality didn't improve dramatically. Adding more inspectors wasn't viable—it would triple labor costs and still couldn't achieve 100% inspection at line speed. The company had tested traditional machine vision systems, but they generated excessive false positives (flagging good parts as defective), couldn't adapt to part variations, and required weeks of reprogramming when designs changed. QualityVision needed an intelligent, adaptive quality inspection solution that could examine every single part at full production speed with accuracy exceeding human capabilities while remaining flexible enough to handle product variations and design changes without extensive reprogramming.

Our Solution

XCodeIT developed an AI-powered computer vision inspection system using deep learning convolutional neural networks trained on millions of labeled images of good and defective parts. Deployed on NVIDIA Jetson edge computing platforms, the system performs real-time inference at line speed while integrating seamlessly with existing production equipment and quality management systems.

Engineered a custom inspection station with 6 industrial cameras capturing synchronized images from multiple angles as parts pass through at 500 parts per minute. LED dome lighting and polarized filters eliminate glare and shadows. Trigger sensors synchronized with conveyor motion ensure precise image capture timing. The imaging system achieves 50-micron resolution, detecting defects invisible to human inspectors.

Trained PyTorch-based convolutional neural networks (ResNet-50 architecture with custom classification heads) on a dataset of 2+ million labeled images encompassing 47 distinct defect types: cracks, voids, flash, short shots, sink marks, color variations, contamination, and dimensional deviations. The models achieve 99.5% accuracy with extremely low false positive rates (<0.3%).

Deployed the inference pipeline on NVIDIA Jetson AGX Xavier edge computing modules, processing 500 parts per minute with <120ms latency per part. Edge deployment eliminates cloud connectivity dependencies, ensures consistent performance, protects proprietary part designs, and enables real-time defect notifications to production line controllers via OPC-UA industrial protocols.

Integrated the vision system with pneumatic reject mechanisms that automatically divert defective parts into categorized bins based on defect type. This enables root cause analysis (certain defects correlate with specific mold cavities or process parameters) and simplifies rework/scrap decisions. Good parts continue to packaging without human intervention.

Implemented a quality feedback loop where production engineers can review flagged parts, correct false positives/negatives, and these corrections automatically feed into model retraining pipelines. Models are retrained weekly with new data, continuously improving accuracy and adapting to process drift or design changes without manual reprogramming.

Built a comprehensive analytics platform that tracks defect rates by part number, production shift, mold cavity, material lot, and defect type. Statistical process control charts identify trends before they become quality escapes. Pareto analysis highlights the most impactful quality improvement opportunities. All inspection images are archived with searchable metadata for customer audit requirements.

Technologies Used

Python PyTorch OpenCV NVIDIA Jetson OPC-UA C++

The Results

99.5%

Defect Detection Accuracy

Deep learning models significantly outperform the 92% accuracy of manual inspection, catching subtle defects humans miss while virtually eliminating false positives. This accuracy has been validated through rigorous testing against known defect samples and confirmed by zero customer quality escapes in the six months post-deployment.

500

Parts Per Minute Inspection

100% inspection at full production line speed—every single part is examined by AI across multiple angles in under 120 milliseconds. This was physically impossible with human inspectors who could only sample-check 10% of production. The system never fatigues, maintains consistent standards 24/7, and scales effortlessly across multiple production lines.

80%

Fewer Customer Complaints

Field failure rates plummeted from 200 defect-related complaints per month to fewer than 40. Customer satisfaction scores increased dramatically, and the major OEM customer threatening contract cancellation instead expanded their business by 30%. The system has created a reputation for quality leadership that's becoming a competitive differentiator.

6 Months

Return on Investment

Total system cost (cameras, edge computing, development, integration) was $380K. The system saves $65K monthly through eliminated customer complaint costs, reduced scrap, labor reallocation, and avoided contract penalties. ROI achieved in 5.8 months, with ongoing savings of $780K annually. The technology is now being deployed across five additional product lines.

"XCodeIT's computer vision system literally saved our business. We were on the brink of losing our largest customer due to quality issues our inspectors simply couldn't catch consistently. Now we inspect every single part with accuracy that exceeds human capability, at full production speed. The false positive rate is so low that our operators trust the system completely. Even better, the continuous learning means it keeps getting smarter. This isn't just quality control—it's a competitive weapon that lets us guarantee quality levels our competitors can't match."

Dr. Michael Chen

Project Details

Industry: Manufacturing & Quality Assurance
Services: Computer Vision, Deep Learning, Edge Computing, Industrial Automation
Duration: 6 months
Team Size: 5 specialists (Computer Vision Engineers, ML Engineers, Embedded Systems Developers)