Custom Background Removal for Print

1. The Challenge

Processing had become the bottleneck

A production Direct-to-Film (DTF) print pipeline removes the background from every customer upload before printing. Processing had become the bottleneck: each image took tens of seconds, and large files had to be capped at 24 MP.

The hard constraint: Print quality could not be traded for speed — thin strokes, fine filaments and soft edges all have to survive onto transparency cleanly. Any solution that gained speed by losing 2 px of edge fidelity was not acceptable.

2. How We Approached It

Profile before optimising

Measured every stage end-to-end to find where the time was actually going, instead of guessing or buying bigger servers.

Purpose-built algorithm

Designed a custom background-removal algorithm for print artwork instead of relying on a general-purpose segmentation tool.

Built for very large images

Re-engineered the data path for multi-core, memory-efficient processing, so 100+ MP files run as smoothly as small ones.

Quality as a hard constraint

Validated every speed-up against fine-detail test artwork, so no gain in speed cost anything in print quality.

Image normalisation −75%

before / after — bars drawn to a common scale (12 s = full width)

12 s

3 s

Background removal >8× Faster

before / after

8 s

<1 s

3. Toolkit

Computer vision OpenCV NumPy pyVips

Service Python FastAPI

Practice Micro-profiling Parallel processing Quality benchmarking

4. Outcomes

Metric	Before	After	Improvement
Image normalisation	12 s	3 s	−75% Time
Background removal	8 s	<1 s	>8× Faster
Max input resolution	24 MP	100+ MP	4× Ceiling
Fine-edge fidelity	Baseline	2 px detail kept	No quality loss

Strategic Business Value

Faster turnaround: Artwork is ready for production in seconds, not tens of seconds — throughput multiplied without adding servers.
New product headroom: Large-format, high-DPI jobs now run without manual pre-processing, opening a new tier of orders.
Lower compute cost: Larger files on the same servers, with no hardware upgrade — profiling eliminated the waste rather than scaling around it.

Note: Implementation details are proprietary to the client. Available for deeper technical discussion under NDA.