Edge Computer Vision and Robotics QC

Real-time edge inspection system linking defect scoring directly to robotics rejection actions.

2020 - 2021

  • OpenCV
  • Edge Computing
  • Robotics
  • GPIO

This system automated visual quality decisions on a live production line by combining camera-based scoring, threshold logic, and actuator control. The key objective was deterministic low-latency rejection with traceable decisions.

Edge quality-control architecture graph showing camera capture, OpenCV scoring, reject actuation, and trace logging.
Architecture graph: camera frames are scored at the edge, rejection decisions trigger actuators, and events are logged for quality traceability.

Problem

Manual inspection could not keep pace with throughput variability and introduced inconsistent reject decisions.

Constraints

  • Decision latency had to stay within the line cycle time.
  • False positives had to be controlled to avoid unnecessary product waste.
  • Every rejection event needed traceability for root-cause analysis.

Architecture

  • Edge execution close to cameras and actuators to minimize response latency.
  • OpenCV-based feature extraction with threshold calibration tuned to product classes.
  • GPIO actuation path for deterministic reject commands.
  • Structured event logging to map visual evidence to reject actions.

Tradeoffs and Failures

  • Aggressive thresholds reduced escapes but increased false rejects.
  • Conservative thresholds lowered waste but risked passing borderline defects.
  • Lighting and camera-angle drift required ongoing recalibration.

Engineering Impact

  • Replaced manual spot checks with deterministic machine-assisted decisions.
  • Connected inspection output directly to mechanical response.
  • Improved post-incident debugging with decision logs tied to timestamps and events.

Outcomes

  • Faster reject loop under real line load.
  • More consistent reject criteria across shifts.
  • Better quality incident traceability and operational feedback.

What Made This Approach Different

The design focused on closed-loop behavior (detect -> decide -> actuate -> log), not just model accuracy in isolation.