Research map / Detection and segmentation / CNN features
A New Statistical Model of Star Speckles for Learning to Detect and Characterize Exoplanets in Direct Imaging Observations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ExoMILD, a multi-scale convolutional statistical model of speckle nuisance (with rotational and spectral symmetries) embedded in an end-to-end learnable detection and flux-estimation framework for SPHERE exoplanet imaging.
Problem
Detecting and characterizing faint exoplanets in direct imaging, where quasi-static stellar speckles dominate the signal, with reliable statistically grounded detection scores and flux estimates.
Approach
Extends the PACO patch-based Gaussian nuisance model into a convolutional mixture of learnable distributions across multiple patch scales, with a learnable linear projection, N-fold rotational symmetry and spectral homothety modeling, and a UNet denoiser trained by Gaussian negative log-likelihood on the detection score.
Key results
- In ASDI mode the proposed method reaches average AUC 0.761 vs 0.693 for PACO and 0.417 for cASDI over five SPHERE datasets.
- In ADI mode the proposed method reaches average AUC 0.645, matching MODEL&CO (0.645) and above PACO (0.520).
- Flux error (ARE) of 0.51 vs 0.56 for PACO and position RMSE of 0.11 vs 0.21 pixels across 1,351 synthetic exoplanets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed end-to-end method instead of PACO | Exoplanet detection in ASDI · Average across 5 test datasets | AUC: 0.693 ± 0.005 -> 0.761 ± 0.004 | Helped |
| Use the proposed end-to-end method instead of MODEL&CO | Exoplanet detection in ADI · Average across 5 test datasets | AUC: 0.645 ± 0.005 -> 0.645 ± 0.005 | No clear effect |
| Use the proposed statistical model instead of PACO | Exoplanet flux estimation · 1,351 synthetic exoplanets detectable by both methods | ARE: 0.56 -> 0.51 | Helped |
| Use the proposed statistical model instead of PACO | Exoplanet position estimation · 1,351 synthetic exoplanets detectable by both methods | RMSE: 0.21 -> 0.11 | Helped |
| Add N-fold rotational symmetries (N=1,2,4) (ablation) | Proposed statistical model · ADI | AUC: 0.554 ± 0.005 -> 0.575 ± 0.005 | Helped |
| Add multiple spatial scales through 64 × 64 patches (ablation) | Proposed statistical model with N=1,2,4 symmetries · ADI | AUC: 0.575 ± 0.005 -> 0.581 ± 0.005 | Helped |
| Add N-fold rotational symmetries (N=1,2,4) (ablation) | Proposed statistical model · ASDI | AUC: 0.713 ± 0.005 -> 0.723 ± 0.005 | Helped |
| Add multiple spatial scales through 32 × 32 patches (ablation) | Proposed statistical model with N=1,2,4 symmetries · ASDI | AUC: 0.723 ± 0.005 -> 0.726 ± 0.005 | Helped |
Surprising findings
- In ADI mode the proposed method is not better than MODEL&CO on average AUC and PCA beats PACO on some datasets, showing dataset-dependent behavior.
Limitations
- Current approach is limited to low spectral resolution data (two spectral channels), with higher spectral resolution left to future work.
- Flux estimation compared only against PACO due to computational constraints.
- Detection relies on synthetic injections for evaluation given lack of ground truth.
Experimental setup
Datasets: SPHERE SHINE-F150 survey observations (220) SPHERE test datasets HD 159911, HD 216803, HD 206860, HD 188228, HD 102647 HR 8799 SPHERE observations
Benchmarks: SPHERE synthetic injections on 5 HD stars HR 8799 real data
Metrics: AUC absolute relative error (ARE) RMSE
Baselines: cADI KLIP/PCA PACO MODEL&CO
Builds on: PACO MODEL&CO ANDROMEDA KLIP/PCA
Closest papers by meaning
- FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution (NeurIPS 2026 · similarity 0.84)
- Kepler Light Curve Classification Using Deep Learning and Markov Transition Field (Student Abstract) (AAAI 2024 · similarity 0.82)
- Learning What's Real: Disentangling Signals and Measurement Artifacts in Multi-Sensor Data, with Applications to Astrophysics (NeurIPS 2026 · similarity 0.81)
- STAR: A Benchmark for Astronomical Star Fields Super-Resolution (NeurIPS 2025 · similarity 0.81)
- The Multimodal Universe: Enabling Large-Scale Machine Learning with 100 TB of Astronomical Scientific Data (NeurIPS 2024 · similarity 0.80)
- Probabilistic Data-Driven Modelling of Astrophysical Transients: The Neural Process Family for Ultrafast and Class-Agnostic Light Curve Reconstruction (NeurIPS 2026 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.