--- title: "Astro Denoise — a noise-reduction benchmark on DC2 images" source: "https://univerlab.org/astro-denoise/" --- [UniverLab · open laboratory](/)EXP-006 # Astro Denoise Frontier knowledge hides behind the noise… astro-denoise is a **research proposal** for a modular, reproducible benchmark of noise-reduction methods on simulated Vera **Rubin** Observatory (LSST DC2) images. Any method — classical filter, trained network — plugs into the same protocol and runs on the same patches, and is scored not by how clean the image looks, but by _what it does to the science_: the **completeness** and **purity** of the faint-source catalog, compared against the DC2 truth catalog. **BM3D** and a **U-Net** are the first two references being explored — the platform is designed to grow as more methods are added. noise reduction → a benchmark scored by the science → comparable methods [Read the proposal (PDF) ↗](https://papers.univerlab.org/astro-denoise/proposal.pdf) ![](/bench/machines/astro-denoise-night-450.webp) EXP-006 · research · Sol 100 (2026-04-10) · roadmap: · Genesis · 2026 · Sol 100 (2026-04-10) ## A master's thesis in progress. This began as a master's thesis and is still taking shape. The ecosystem has worked since July 2026: an orchestrator with a common command-line protocol (FITS image in and out, metadata in JSON), a no-processing control condition, and two separately packaged methods, BM3D for the classical family and a U-Net for deep learning. The experiments run in the open, right here, but there is work ahead before this is a finished benchmark. Objectives ## Benchmarking noise-reduction methods on simulated DC2 images of the Vera Rubin Observatory Develop a reproducible benchmarking platform that compares noise-reduction methods on simulated DC2 images of the Vera Rubin Observatory, evaluating for each method the recovery of faint sources, the resulting image quality and its computational cost. 1. 1completeA modular, extensible ecosystem with a common command-line protocol, where each method is packaged separately with its own environment and a new method joins without changing the orchestrator. 2. 2in progressReliable source-recovery metrics (completeness, purity, completeness by magnitude and photometric accuracy, by one-to-one matching against the DC2 truth catalogue) and image-quality metrics (PSNR, SSIM and background noise), each checked against at least two independent sources in the literature. 3. 3in progressTools that measure each method's computational cost (time per region, peak memory and regions per unit of time) from outside its process, reported with a description of the hardware. 4. 4in progressThe methods under test: BM3D, and a U-Net trained only on DC2 regions outside the evaluation set, with its training recipe and weights published. 5. 5pendingA reproducible benchmark on a curated, versioned DESC DC2 dataset of multiband regions from several sky areas; each version is identified by checksums and declares its regions, and every result is reported against one version. Repositories * [UniverLab/astro-denoise](https://github.com/UniverLab/astro-denoise)The bench: orchestrator, science metrics and the control condition. * [UniverLab/astro-denoise-bm3d](https://github.com/UniverLab/astro-denoise-bm3d)BM3D, the classical baseline. * [UniverLab/astro-denoise-unet](https://github.com/UniverLab/astro-denoise-unet)The U-Net, the deep-learning baseline. Open questions * Q1Can a denoiser raise completeness without hurting purity — or does it invent sources that were never there? * Q2Which family of methods — classical filters or learned models — is more likely to help at LSST DC2 depth, and by how much? * Q3Can the whole benchmark be repeated with the same scientific results on the same version of data and code? * ![Universidad de La Salle](/logos/unisalle.png) * ![LSST](/logos/lsst.png) * ![Vera C. Rubin Observatory](/logos/verarubin.png) * ![Globus](/logos/globus.png) Cross-experiment ## Research Research artifacts that cut across the lab’s experiments — papers, datasets, benchmarks, models. The first entries come from this track; the list grows only as work is actually published. * in preparation ### Papers Publications and preprints produced by lab experiments. * in preparation ### Datasets Open datasets, starting with astro-denoise benchmark data. * in preparation ### Benchmarks Reproducible benchmark suites with documented assumptions. * in preparation ### Models Trained models released alongside their training recipes. Star chart: Yale Bright Star Catalogue (CDS V/50); constellation lines from d3-celestial by Olaf Frohn (BSD-3-Clause). FAQ ## FAQ What is astro-denoise? A research project benchmarking noise-reduction methods on simulated Vera Rubin Observatory (LSST DC2) images. It compares three conditions: a no-processing control, BM3D as the classical baseline and a U-Net as the deep-learning baseline. The goal is to recover faint galaxies without inventing ones that were never there — scored by science recovery (completeness + purity), not visual quality. What is noise reduction in astronomy? Astronomical images contain unavoidable noise from the detector and sky. Noise reduction aims to lower this noise while preserving real sources. The risk: a method can smooth out real faint galaxies or invent fake ones that were never there. What is LSST? The Legacy Survey of Space and Time — a 10-year astronomical survey conducted by the Vera Rubin Observatory in Chile. It will image the entire visible sky repeatedly, producing the deepest wide-field catalog of galaxies, stars, and transient events ever made. What is DC2? Data Challenge 2 — a simulated dataset that mimics what LSST will produce, created by the Dark Energy Science Collaboration (DESC). It includes realistic noise, PSF, and atmospheric effects, with a truth catalog telling you exactly which sources are real. Field notes