kjfk/galaxy-bar-detection-resnet50

Galaxy Bar Detection — ResNet50 (IITK Astronomy Project)

A CNN that measures stellar bar strength in simulated disc galaxies from a single image — and classifies barred vs unbarred as a by-product.

⭐ Final model: regression_resnet50.weights.h5

Regresses the physical A2/A0 Fourier bar-strength directly from a 224×224 grayscale stellar-density image. Classification = predicted A2/A0 ≥ 0.19.

Metric (450-image test set, held-out snapshots)	Value
Classification accuracy	96.0%
ROC AUC	0.981
Bar recall	92.0%
Strength error (MAE)	0.022 in A2/A0 units (range 0–0.28)
Face-on (0°) accuracy / MAE	100% / 0.014

It reproduces the physics across the whole simulation

Run over 182 snapshots spanning the simulation, the CNN's predictions track the independently-computed A2/A0 curve through the entire bar-formation history — MAE 0.0144 — from images alone (bar_evolution.png).

It generalises beyond its training range

Trained only on inclinations ≤ 40°, evaluated 0–90° on held-out snapshots (tilt_breakdown.png):

Inclination	0°	20°	40°	50°	60°	70–80°	90°
Accuracy	100%	89%	89%	89%	78%	~70%	82%

Performance holds at 50° (never trained there) and degrades gracefully toward edge-on, where a bar is physically hidden by projection.

---

The key finding: reframing beat data

The binary classifier trained on the same images scored 86.4%. Switching the target from a bar/no-bar label to the continuous A2/A0 value took it to 96.0%. An ablation isolates the cause (ablation_attribution.png):

Setup	Same 450-image test
Binary classification, v3 data	86.4%
Regression, identical v3 data	95.6%
Regression + 1,500 extra mid-strength images	96.0%

+9.2 points came from the reframing alone; the extra data added +0.4 (within noise). Binary labels forced the discarding of "mid-strength" galaxies (0.10 < A2/A0 < 0.22) — exactly the region containing the 0.19 decision boundary — leaving the classifier blind where it most needed to see. Regression removes the need to discard anything and supervises the full continuum.

---

Loading (weights-only — a Keras full-model serialization bug)

from tensorflow import keras
from tensorflow.keras import layers

base = keras.applications.ResNet50(weights=None, include_top=False, input_shape=(224,224,3))
inp = keras.Input((224,224,3))
x = keras.Sequential([layers.Identity()])(inp)      # placeholder (augmentation at train time)
x = keras.applications.resnet50.preprocess_input(x)
x = base(x, training=False)
x = layers.GlobalAveragePooling2D()(x)
x = layers.Dropout(0.3)(x)
out = layers.Dense(1, activation="sigmoid")(x)
model = keras.Model(inp, out)
model.load_weights("regression_resnet50.weights.h5")

# prediction * 0.30 = A2/A0 ;  >= 0.19 means "barred"

Input spec (must match exactly, or predictions are invalid): 224×224 stellar-density map, log10 clipped to 1e7–1e9 then normalised 0–1, grayscale replicated to 3 channels, 40 kpc field of view, galaxy centred and face-on aligned before tilting.

Training

ResNet50 (ImageNet) two-phase transfer learning: freeze backbone + train head (8 epochs, lr 1e-3), then unfreeze and fine-tune (≤18 epochs, lr 1e-5, early stopping on val MAE). Huber loss, random rotation/flip/translation augmentation. Splits are by simulation snapshot, never by image, so no galaxy appears in both training and test. Verified no overfitting: train MAE ≈ test MAE.

Labels are physical, not manual

Every image inherits the A2/A0 bar strength computed from the N-body particle data of its source snapshot (GADGET simulation, 502 snapshots) — no human labelling anywhere in the pipeline.

Earlier binary models (the journey)

bar_resnet50_v{1,2,3}.keras — v1 76.5% (tilts to 60°, label noise from invisible bars) → v2 86.8% (tilt capped at 40°) → v3 86.4% (AUC 0.944). Adding test-time augmentation + 3-seed ensembling reached 91.8%. A morpho-kinematic variant (velocity + dispersion channels) was tested and did not help (83.3%) — morphology alone suffices at ≤40°. Across architectures on the regression task: ResNet50 96.0% ≫ ConvNeXt-Tiny 91.6% ≫ EfficientNetV2-B0 84.0%; ensembling diluted the strongest model rather than improving it.

Limitations

One simulated galaxy observed at many epochs — generalisation to other galaxies is untested. Sim-to-real transfer (telescope noise, PSF, foreground stars) is also untested. Valid domain: inclinations ≲ 50°, images built with the exact render recipe above.

Code, dataset pipeline, all 15 figures and full analysis: https://github.com/GuptaOum/IITK_astronomy_project