Computer vision quality control system using Vertex AI Vision, Cloud Vision API, and Gemini Vision — detecting defects, anomalies, and quality deviations in real-time from production line cameras. 98.7% accuracy at 0.3 seconds per unit.
We implement a production-grade computer vision inspection system on Google Cloud — training custom Vertex AI Vision models on your defect taxonomy, deploying real-time inference at the edge or in the cloud, routing borderline cases to human review, and streaming all results to BigQuery for SPC analytics. The system integrates with your existing MES and quality management software.
Human visual inspection is limited by fatigue, subjectivity, and throughput — inspectors miss defects at high line speeds and apply inconsistent standards across shifts.
When inspection fails to catch defects on the line, the cost multiplies — warranty claims, returns, customer complaints, and brand damage that far exceed line-level prevention costs.
Rules-based inspection systems generate high false positive rates — rejecting acceptable product and triggering unnecessary line stoppages that reduce throughput.
Vertex AI Vision and Gemini Vision process camera feeds in real time — detecting surface defects, dimensional deviations, assembly errors, and cosmetic issues at line speed.
Defects are classified by type, severity, and location — enabling root cause analysis and targeted process adjustments rather than undifferentiated rejection.
Borderline cases below confidence thresholds are routed to human review with the camera frame and AI assessment highlighted — keeping humans in the loop for ambiguous cases.
All inspection results land in BigQuery — enabling SPC analysis, shift-by-shift trend monitoring, defect Pareto charts, and correlation with process parameters.
Confirmed human review decisions feed back into model training automatically via Vertex AI Pipelines — continuously improving accuracy on new defect types and product variants.
Vertex AI Vision can be trained to detect any visually distinguishable defect: surface scratches, dents, discoloration, missing components, incorrect assembly, label placement errors, dimensional deviations, and contamination. The specific defect classes are defined during model training using your labelled image dataset. Detection capability is directly related to the quality and diversity of the training data we build with your quality team.
The system is camera-agnostic and works with any IP camera capable of streaming RTSP or HTTP video. Industrial cameras (Basler, FLIR, IDS) provide the best results for high-speed inspection. Lighting requirements depend on the defect type — surface defects benefit from structured lighting; dimensional checks work with standard diffuse illumination. We advise on camera positioning and lighting during the site survey phase of implementation.
We work with your quality team during implementation to collect and label a training image dataset of acceptable product and known defect categories. Vertex AI Vision requires approximately 100–500 images per defect class for initial model training. The model improves continuously through the human review feedback loop. We handle model training, evaluation, and deployment as part of the implementation.
The system processes images at 0.3 seconds per frame, enabling inspection of up to 200 units per minute at one frame per unit. Higher throughput is achieved by deploying multiple camera stations or using GPU-accelerated inference nodes. For very high-speed lines, we design a multi-camera array with parallel inference to maintain 100% inspection coverage. Exact throughput is validated during the site assessment.
Catch defects before they reach customers. We implement a production-ready AI vision inspection system in 4–6 weeks.