Image analysis pipelines for production imagery

The company needed to compare how six ink formulations spread after landing on a test surface. Each printed droplet was roughly 150 microns across, and a reliable decision required measuring a meaningful sample rather than inspecting a handful by eye.

Manual measurement would have taken weeks. Generic image tools measured one droplet at a time, while dust and scratches were visually similar enough to droplets to skew a conventional contour analysis.

The pipeline first located the bounded test strip, divided it into six color regions with engineered margins, and used OpenCV contour analysis to identify plausible droplets and draw bounding circles around them.

To remove false positives, we developed a “Smaller Neighbor” algorithm. It compared nearby detections and discarded the smaller candidate, using the droplets' regular spacing to separate real printed dots from dust and scratches.

The resulting Python application processes each high-resolution test image, isolates all six ink colors, filters the detections, calculates mean diameters, and exports both annotated evidence and structured measurements.

It ran on a laptop and processed hundreds of test strips, giving the engineering team a repeatable analysis they could inspect rather than an opaque model output.

A task estimated to take weeks was reduced to minutes, saving more than $33,000 per year in labor.

The measurements showed that the cyan ink spread substantially more than the other formulations, giving the company evidence to replace it before production use.