Problem
A materials lab makes silicon carbide (SiC) ceramic by siliconizing a porous carbon blank. Technologists tune the share of residual carbon (C) and free silicon (Si) for each application: more carbon for friction pairs, more SiC for strength. For every test sample they must learn the phase shares fast. Weighing after leaching out silicon and burning off carbon takes days, manual analysis of polished sections is slow, and threshold segmentation takes an hour or more per few images and has an error above 10%, because the SiC and Si phases look alike in some areas.
Example from the article: 279 colour optical-microscope images of polished sections (×200, about 4600 × 3400 px) had to be split into three phases (Si, C, SiC), with phase areas feeding process adjustments.
Solution in brief
Cluster the pixels of each image into three groups with k-means in Lab color space, which gives rough phase masks without manual annotation. Cut the images into 256 × 256 tiles and train a U-Net on these masks. The network returns masks and area shares per phase in a few minutes per batch of images.
Questions this case answers
- How to measure area shares of material phases on microscope images without manual annotation, by pixel clustering that supplies masks to a segmentation network
- How to make segmentation robust to brightness and contrast differences between images without relabeling
- How to measure the area share of each material phase on micrographs without manual labeling
- How to get segmentation masks for a new image domain without an annotated dataset
Tools and methods
- k-means in Lab space — labels three phases without annotators and tolerates lighting changes
- U-Net — segments tiles into three probability channels (Si, C, SiC)
- Binary cross-entropy, Adam, ReLU, early stopping — training setup
- Optical microscope images — the only input
Google or a one-time email code. No password.