Computer Vision Development in Toronto

We developed and deployed a visual quality inspection system over 14 weeks. (1) In-line defect detection: high-speed camera systems installed on 35 production lines across 12 plants — 4-8 cameras per line capturing images of every part at production speed (120+ parts per minute). Lighting: custom LED illumination arrays designed for each part geometry — strobed lighting eliminating motion blur and creating consistent, repeatable illumination regardless of ambient conditions. Models: deep learning defect detection models trained on 4.2 million annotated images from the manufacturer's historical quality records. Defect types: surface defects (scratches, dents, tool marks, porosity, cracks — detected at sub-millimetre resolution), dimensional anomalies (warpage, incorrect trim, hole position deviation — detected through silhouette and feature measurement), and material defects (rust, contamination, coating defects — detected through colour and texture analysis).

Real-time classification: each part classified as pass, fail, or suspect within 85ms — faster than the production cycle time, enabling 100% inspection without slowing production. (2) Stamping press inspection: purpose-built inspection stations at 15 stamping presses — the highest-speed, most challenging inspection environment. Die monitoring: continuous imaging of the stamping die surface between strokes — detecting die wear, buildup, and damage that caused progressive quality degradation. The system predicted die maintenance needs 200-500 strokes before defect onset — enabling proactive die change during scheduled breaks rather than reactive change after defect discovery. First-piece verification: automated first-piece inspection after die setup or changeover — comparing the first production parts against CAD specifications within 30 seconds (replacing 15-minute manual first-piece inspection). (3) Assembly verification: camera systems verifying assembled modules — component presence (correct parts installed), orientation (parts correctly positioned), and completeness (all fasteners, clips, and secondary operations completed).

Poka-yoke integration: inspection system integrated with assembly station PLCs — stations locked and alerted operators when the vision system detected missing or incorrect components, preventing defective assemblies from advancing to the next operation. (4) Quality analytics: defect trend analysis — real-time monitoring of defect rates by type, part number, production line, shift, and time period. SPC integration: vision system measurements feeding automated SPC charts — detecting process drift and out-of-control conditions before they produced out-of-specification parts. Root cause correlation: defect patterns correlated with process parameters (press tonnage, cycle time, material lot, die age) — accelerating root cause investigation from days to hours. OEM reporting: automated generation of quality reports in each OEM's required format — PPAP documentation, quality performance dashboards, and 8D investigation data pre-populated from vision system records.

(5) Edge architecture: all inspection processing performed at the edge (NVIDIA Jetson Orin modules at each inspection station) — no dependency on network connectivity or cloud for real-time inspection decisions. Central analytics: inspection results streamed to a central cloud platform (AWS ca-central-1) for cross-plant analytics, model updates, and OEM reporting. Model updates: new defect detection models trained on cloud and deployed to edge devices during maintenance windows — keeping detection capabilities current with new part introductions and evolving defect types.