Back to Portfolio Source

Results · Embedded Imaging

Automated Bacterial
Colony Counter

An embedded imaging system for Raspberry Pi 5: an RP1 register-level PWM driver for the LED illumination, and an OpenCV pipeline that segments and counts colonies, with a statistical anomaly layer and an ML layer that activates once enough colonies are labelled. Below are the blind-evaluation results; the memo has the full write-up.

PythonOpenCVRaspberry Piscikit-learnFlask
Blind Validation
Technical Memo

The system scored blind on 250 synthetic plates, with ground truth loaded only after detection returned. A watershed bug found during the run was fixed and the set re-run; the memo documents the change. The memo reports the real numbers, including the two sets that missed target and the segmentation failure mode behind them, and sets them next to published error rates for OpenCFU and ColonyDoc-It.

0.869
Pooled F1 across 250 blind-scored plates (target 0.90)
1.000
Precision on every plate, zero false colonies
3 / 5
Test sets at or above target
15.75%
Mean count error (5.31% on passing sets)