Dislocation Extraction Pipeline

The problem

Materials scientists study dislocations, the line defects of a crystal, in electron microscopy films of aluminium under strain. Measuring their density and curvature was done by hand on a few chosen frames, while each film holds thousands. The lines cover about 3% of the pixels, two thirds of the corpus had no annotation, and the usual segmentation score rewarded masks that looked right but broke lines apart.

What I built

1. 1

   Built a human-in-the-loop annotation tool (FastAPI and React) with magnetic live-wire tracing, then grew the dataset from 318 to 890 image-mask pairs under a single line-width convention.

2. 2

   Replaced pixel overlap with topological metrics, each checked on hand-built failure cases, and hardened the training protocol to cut run-to-run dispersion 20 to 40 times.

3. 3

   Ran a controlled benchmark of 4 segmentation architectures with architecture as the only variable; MA-Net won on every metric.

4. 4

   Wrote a CPU-only instance separation library (Hungarian matching, integer programming, smoothing splines) with 387 tests and 94% line coverage.

5. 5

   Packaged the full chain as a desktop application with SHA-256 manifests, provenance-carrying exports and 130 automated tests.

Outcome

• A simple connected-component filter removed 75% of line fragmentation, a larger gain than the whole architecture search.

• 17,144 dislocations extracted over the corpus at 0.59 s median per image, turning a manual task into a repeatable measurement of density and curvature.

• Caught two silent defects that had invalidated a full training period, including a library argument ignored without error that injected 15 times too much noise.