Computer vision for in-line quality control

Context

Componentes Industriales del Bajío is a Tier 1 auto-parts supplier (stamped and injected parts) for automakers established in Mexico. With three production lines running 24/7, quality inspection depended on human staff at line's end, checking parts by eye or manual gauges.

The challenge

• Defect-detection rate (burrs, surface cracks, part misalignment) inconsistent across shifts — dependent on inspector fatigue.
• Quality claims (PPM) from automaker clients above the contractual target.
• Line speed limited by manual inspection time (bottleneck).
• Automotive client requirement: full per-part traceability (IATF 16949) and auditable inspection evidence.

The solution: edge computer vision

Instead of relying on manual inspection or sending images to external servers, the plant implemented a computer-vision system with proprietary models running at the edge, inside the production line itself.

ARCHITECTURE

• Capture: high-resolution industrial cameras with structured lighting at 3 critical line points.
• Edge inference: the defect-detection model (convolutional network trained on historical OK/NOK part images) runs on edge devices next to the line, with no cloud-connectivity dependency — processing happens directly on the device, reducing latency, improving data privacy, and enabling faster reactions.
• Real-time decision: a part classified NOK is automatically rejected via actuator before moving to packaging — no human intervention needed in the normal flow.
• Traceability: every inspection is recorded with image, timestamp, and result, linked to the batch number — evidence ready for IATF audits.
• Continuous retraining: edge cases (low model confidence) route to a human inspector; those decisions feed the model's monthly retraining.
• Proprietary model, no third-party dependence: by not sending product images or design specs to external services, the plant protects the automotive client's intellectual property.

Results at 6 months

Metric	Before	After
Inspection speed	~4 sec/part (manual)	<0.8 sec/part
Defects detected before packaging	~82%	99.2%
Client quality claims (PPM)	Base 100%	-71%
Parts reworked due to false rejects	Base 100%	-45%
Manual inspection hours per shift	8	2 (low confidence only)

Inspection staff transitioned to an exception-validation and continuous-model-improvement role, instead of repetitive piece-by-piece inspection.

Key success factors

• Proprietary model running at the edge, not a closed "smart camera" from a vendor — allows tuning the model to each client's specific parts.
• Well-designed lighting and capture from the start — most machine-vision failures on the floor come from poor capture, not the model.
• Human in the loop for low-confidence cases, preventing the system from making blind decisions on ambiguous parts.
• Traceability as a byproduct, not a separate project — every inspection generated auditable evidence with no extra effort.

Lesson to replicate

The differentiator wasn't just "putting a camera with AI" — it was designing the system to decide on the line, without depending on the cloud, with proprietary data capture and a clear retraining process — especially relevant in manufacturing, where latency and protecting client designs are critical.