Skip to content

Case Study: Conservative MegaCam Denoising with Noise2Noise

This page walks through a runnable example that trains a small neural network on CFHT MegaCam calibration darks and then applies it carefully to science data. The model is trained only on darks, so it cannot reconstruct arbitrary scenes. The example keeps that boundary explicit and uses the model in a way that preserves sources.

The implementation is examples/example_megacam_cr_denoise.py.

What the example demonstrates

The pipeline uses:

  1. FitsCutoutDataset to read matching patches from compressed calibration files without converting them to another format.
  2. A custom FITSTransform that normalizes each patch by its own median and median absolute deviation.
  3. StackDataset and make_loader to build paired PyTorch batches.
  4. A compact, fully convolutional network trained with an L1 Noise2Noise loss.
  5. Reading of whole CCDs, inference, metrics, JSON summaries, Markdown tables, and an optional figure showing each window before and after.
flowchart LR
    D1["Training dark A"] --> P["FitsCutoutDataset + StackDataset"]
    D2["Training dark B"] --> P
    P --> T["SelfNorm median/MAD"] --> L["make_loader"] --> N["Blank field U-Net"]
    N --> DT["Reserved dark test"]
    S["Science CCD"] --> N
    N --> C["Conservative cosmic ray repair"]
    C --> M["Source and background metrics"]

Why train on darks?

With the shutter closed, the intended image is blank. The detector still adds dark current, hot pixels, read noise, and cosmic ray hits. Two darks are therefore noisy observations of the same empty field, with independent read noise and independent hit positions. The example trains on paired patches from two different darks, which teaches the network to predict blank for a given detector.

That training is a useful calibration exercise. A dark does not give a clean image of a live field, so the example never writes the blank prediction over science data; doing so would erase stars, galaxies, and diffuse structure along with the artifacts.

By default the last dark file is set aside as a test exposure. The remaining darks train the model, and CCDs 31 through 40 of those training files serve as a separate convergence check. The reserved dark gives the model a real exposure it never saw during optimization.

Biases provide a read noise control through --mode bias. The default --mode both runs the dark and bias experiments together when both sets are present.

The two inference policies

Reserved dark: predict the blank field

A dark really is a blank field, so the network prediction is used for the whole test CCD. Dark current means the prediction remains a calibration proxy rather than pixel-perfect truth. The generated dark_test_metrics.md and dark_summary.json report:

  • the cosmic ray and sharp outlier fractions before and after cleaning;
  • the dark RMS about the median before and after cleaning; and
  • the file and HDU used for the test.

These numbers come from a real dark exposure rather than synthetic injections.

Science: conservative repair of isolated pixels

For a science CCD the network acts as a second opinion alongside the image itself. A pixel is repaired only when it passes both parts of the cosmic ray test:

  1. it lies more than 5 * robust_sigma above a 7x7 box background; and
  2. it lies more than 5 * robust_sigma above the median of its 3x3 neighbourhood.

The blank prediction must also place the pixel more than 4 * MAD above the predicted blank level. Pixels that pass are replaced by their local 3x3 median; every other science pixel is copied unchanged. This favours keeping sources over aggressive smoothing of read noise.

The science report measures the cosmic ray and sharp outlier fractions before and after cleaning, a count of faint structure, the background median, and aperture flux ratios at bright local maxima. Because no clean science target exists, these values are reported as diagnostics rather than accuracy scores.

The noise-injection test adds the difference of two training darks to a science window and measures how much the conservative path changes. It provides a stress test of that path.

Optional Astro-SCRAPPY comparison

Astro-SCRAPPY is a classical baseline for removing isolated cosmic rays. It is an optional extra, installed separately from torchfits. When it is available, pass --compare-astroscrappy to run it on the same science and reserved dark windows:

python examples/example_megacam_cr_denoise.py \
  --mode dark \
  --compare-astroscrappy

The script calls it with sigclip=5, sigfrac=0.3, objlim=5, four iterations, and a robust read noise estimate for each window. The baseline output lands in astroscrappy_science_metrics.md and astroscrappy_test_dark_metrics.md, and also appears in the dark summary JSON. If Astro-SCRAPPY is missing, the command prints a short skip message and the torchfits path still runs.

The comparison uses the same raw windows and the same cosmic ray diagnostics for both methods. Neither method can claim to have found every cosmic ray, because no labelled science exposure exists; the mask is a proxy in both cases.

Data and execution

The fetch scripts cache public CFHT files under benchmarks_data/cfht_megacam/:

bash scripts/fetch_cfht_megacam_sample.sh
bash scripts/fetch_cfht_calib_frames.sh
python examples/example_megacam_cr_denoise.py \
  --mode both \
  --compare-astroscrappy

Set TORCHFITS_EXAMPLE_FAST=1 for a bounded run that trains one epoch, evaluates a single CCD, and uses 1024x1024 windows. It still exercises the dataset, transform, loader, inference, and reporting paths.

Useful options:

--mode {dark,bias,both}       Which calibration experiment to run
--eval-hdus N                 Number of HDUs per evaluation file
--full-eval-files N           Number of science files to inspect
--compare-astroscrappy       Run the optional classical baseline
--inject-stars                Add the synthetic flux recovery appendix
--out-dir PATH                Directory for Markdown, JSON, and figures

A MegaCam file holds many large CCD extensions, so the full run takes a while. Start with --eval-hdus and --full-eval-files to keep a first run short and reproducible.

Reading the figure

The optional gallery image is a 512x512 crop chosen because it has the most cosmic ray candidates under the diagnostic mask. It is a focused patch, so a sparse region can look nearly empty. The figure shows localized pixel changes rather than overall photometry.

Each panel is stretched independently between its 1st and 99th percentiles, so the on-screen brightness has no calibrated meaning. For quantitative interpretation, use the Markdown metrics and the aperture flux and background values in the JSON summary. The "cosmic ray removed" value counts pixels that the before image proxy flagged and that the same proxy no longer flags after cleaning; it measures a change in that proxy rather than supervised detection accuracy.

Limitations

  • Training on darks teaches nothing about the morphology or flux distribution of stars and galaxies.
  • The science path repairs isolated, sharp positive excursions only. Trails and extended artifacts remain in place.
  • The cosmic ray masks are heuristic, because no clean version of the science frame is available.
  • Astro-SCRAPPY and the neural path use different algorithms; compare them with the source and background diagnostics rather than a single visual impression.

The main lesson is how the pieces compose: calibration frames feed a PyTorch training loop directly, and the same torchfits reads feed a careful, measurable inference result without exporting images in between.