PCB Vision System

A client needed automated PCB defect detection at their assembly station. Requirements: real-time inference on modest hardware, fully offline-capable, boots to ready without manual setup.

YOLOv8 was trained on their labeled PCB dataset. Two export paths were benchmarked: ncnn (the standard Raspberry Pi choice) and ONNX with Coral Edge TPU offload. The Coral path delivered 2-3x the throughput. Tkinter was chosen for the interface: lightweight, no browser dependency, no network stack. A systemd service auto-starts the detection pipeline on boot with a branded splash screen before the live feed. The whole experience is a sealed appliance.