U-Net Segmentation

There are no manual galaxy masks in Galaxy10, so segmentation is weakly supervised: an Otsu-based teacher produces three-class pseudo-labels, and a U-Net is trained to reproduce and regularise them.

Pseudo-labels (the “teacher”)

OtsuSegmenter (background percentile 5) builds a galaxy mask by Otsu threshold, hole filling and largest-component selection, then splits it:

Class	Definition
0 — background	outside the galaxy mask
1 — disk	galaxy mask, below the 80th percentile of galaxy pixel brightness
2 — “bulge”	pixels above that percentile

The “bulge” is a brightness quantile

Class 2 is by construction the brightest 20 % of the galaxy's pixels, so the Otsu bulge fraction is 0.200 for every morphology class. It is not a physical bulge/disk decomposition. The U-Net learns this brightness criterion; its predicted bulge fraction (median 0.100) correlates only very weakly with colour (see Physical Interpretation).

Pseudo-label pixel shares on the training set: 9.5% background, 72.4% disk, 18.1% bulge; class weights in the loss compensate the imbalance.

Model and training

U-Net (Ronneberger et al. 2015) with four encoder levels (64 → 512 filters), a 1,024-filter bottleneck with dropout 0.3 and bilinear up-sampling. Loss: 0.5 × Dice + 0.5 × cross-entropy (label smoothing 0.05, class weights 0.3 / 1.5 / 3.0). AdamW, learning rate 1 × 10⁻⁴, 20 epochs, FP16, best epoch by validation mean IoU.

Results

0.540

Test mIoU

against Otsu pseudo-labels

80.7%

Pixel accuracy

all test pixels

39.6%

Bulge recall

brightest-quantile class

Provenance. Model scores in this section are copied from the output cells of the AstroVision notebooks (and, for DINOv2 per-class recall, read from an exported figure). The documentation data pack contains no metrics file for them, so they cannot be re-derived from the pack; the cell of origin is listed for each value in _facts/astrovision/FACTS_INDEX.md §3 and shown when hovering over a number.

Pixel class (test)	Precision	Recall	F1
Background	0.722	0.611	0.662
Disk	0.825	0.935	0.876
“Bulge”	0.730	0.396	0.513

Notebook 06, per-pixel classification report.

What the mIoU does and does not mean

The mIoU is measured against the Otsu pseudo-labels, not against human masks, so it quantifies how well the network reproduces its teacher. The comparison table in the project README (random ≈ 0.167, Otsu ≈ 0.500, “+4 % over the teacher”) has no counterpart in the notebook outputs and is not reproduced here: the teacher scored against itself is trivially perfect. The network misses most of the “bulge” pixels (recall 39.6%) because a quantile threshold is a noisy target at the pixel level.

By morphology class, the U-Net mIoU ranges from 0.355 (Cigar Shaped Smooth) to 0.517 (Unbarred Tight Spiral).

U-Net mIoU by galaxy class

0 · Disturbed0.41

1 · Merging0.42

2 · Round Smooth0.42

3 · In-between Round Smooth0.39

4 · Cigar Shaped Smooth0.35

5 · Barred Spiral0.49

6 · Unbarred Tight Spiral0.52

7 · Unbarred Loose Spiral0.44

8 · Edge-on without Bulge0.39

9 · Edge-on with Bulge0.37

Other segmenters

DINOv2PatchSegmenter clusters patch features with k-means (k = 3, 4, 5 tried in the notebook) — a label-free alternative, shown qualitatively. SAMSegmenter (Segment Anything) is optional and was not run: its checkpoint was absent in the recorded notebook.