Directly imaging an exoplanet has been famously compared to spotting a firefly hovering next to a coastal lighthouse from miles away. The core challenge is one of extreme contrast: a host star typically outshines its orbiting terrestrial planets by a factor of $10^5$ to $10^{10}$ (Goun et al.; Laginja et al.).
If a telescope were optically perfect and sat in a vacuum, starlight would form a predictable diffraction pattern that could be easily subtracted. In reality, atmospheric turbulence (for ground-based observatories) and minute thermal or mechanical distortions (for space telescopes) warp the incoming light waves. This creates a dynamic, evolving interference pattern known as star speckles. These speckles perfectly mimic the spatial appearance of faint planetary companions, completely burying the true planetary signal under structured optical noise (Bodrito et al.; Daglayan et al.).
To overcome this barrier, the field of high-contrast imaging operates on two fronts: physical hardware suppression to block the bulk of the starlight, and mathematical post-processing to separate residual speckles from genuine astronomical signals.
Physical Starlight Suppression and Wavefront Control
Before any algorithmic denoising can occur, the overwhelming glare of the host star must be physically attenuated to prevent detector saturation. This is achieved using two primary hardware mechanisms:
- Coronagraphs and Flight Masks: Internal masks placed within the telescope's optical path block the central starlight while allowing off-axis light from surrounding planets to pass through (Kasdin et al.; Riggs et al.).
- Starshades: External spacecraft deployed tens of thousands of kilometers ahead of the telescope cast a precise geometric shadow over the objective lens, blocking starlight before it ever enters the instrument (Hu et al.).
However, physical masks are only effective if the stellar wavefront remains perfectly centered and stable. To counteract real-time optical distortions, telescopes use Adaptive Optics (AO) systems equipped with deformable mirrors. Because these mirrors must adjust their shapes thousands of times per second across multiple wavelengths, computing the necessary corrections is a massive bottleneck.
To solve this, Goun et al. developed a reduced-dimensional system identification technique for multi-spectral wavefront control. By projecting the complex, high-dimensional behavior of the optical system into a lower-dimensional state space, the authors significantly accelerated the control loop, enabling the rapid stabilization required to achieve contrast ratios exceeding $10^9$.
Algorithmic Separation: Low-Rank vs. Sparse Trajectories
Even with advanced wavefront control, residual starlight leaks through, forming a quasi-static speckle field. To isolate a planet from these speckles during post-processing, astronomers exploit a key behavioral difference: field rotation.
When a telescope tracks a star over several hours, the orientation of the telescope's internal optics remains fixed relative to the instrument, meaning the speckle pattern stays relatively stationary. However, the astronomical field of view rotates. From the perspective of the detector, the speckles remain quasi-static, while any real planet moves along a predictable circular arc over time.
[ Image Stack Over Time ]
|
+---> Quasi-Static Speckles (Highly correlated across frames) --> LOW-RANK MATRIX
|
+---> Moving Planet (Appears in unique pixels along an arc) --> SPARSE TRAJECTORY
Daglayan et al. formalized this intuition by framing exoplanet detection as a matrix decomposition problem called Low-Rank Plus Sparse Trajectory Decomposition (LRPT). If you stack a sequence of images into a matrix, the highly correlated, slow-moving speckle background forms a low-rank component. Conversely, the planet is sparse because it only occupies a single pixel at any given timestamp, moving along a specific trajectory.
Rather than performing a simple one-pass subtraction, the authors introduced an iterative framework called the Alternating Minimization Algorithm with Trajectory (AMAT) (Daglayan et al.). AMAT continuously alternates between refining the low-rank model of the speckles and updating the estimated path of the planet. This joint optimization prevents the planetary signal from accidentally leaking into the background model, dramatically improving the final signal-to-noise ratio.
Alternative causal frameworks also exploit these spatial correlations. For instance, Gebhard et al. adapted Half-Sibling Regression (HSR) to high-contrast imaging. HSR uses pixels far away from the target location (which share the same systemic telescope noise but contain no planet signal) to causally predict and subtract the speckle noise at the target location, providing highly accurate planet brightness estimates.
Hybrid Denoising: Whitening Meets Deep Learning
In recent years, deep learning models like Convolutional Neural Networks (CNNs) have been deployed to recognize faint planetary signals. However, standard CNNs struggle when applied directly to raw high-contrast images. Because star speckles exhibit massive spatial correlations, a neural network easily mistakes a bright speckle cluster for a planet, leading to high false-alarm rates.
To bridge this gap, researchers have developed hybrid architectures that combine classical statistical modeling with deep learning:
- deep PACO: Flasseur et al. combined patch-covariance modeling with a CNN. The patch-covariance step estimates the local statistical structure of the speckle noise and uses it to "whiten" the image. Whitening strips away the heavy spatial correlations of the speckles, transforming the background into independent and identically distributed (i.e., white) noise. When the CNN processes this whitened data, it no longer has to learn to ignore complex speckle shapes; it only needs to detect the remaining coherent planetary signature.
- ExoMILD: Similarly, Bodrito et al. introduced a framework that models star speckles across multiple spatial scales and physical symmetries. By embedding these physical invariants directly into the statistical preprocessing layer, the model achieves high sensitivity even on highly distorted datasets.
By offloading the noise-correlation modeling to rigorous statistical frameworks, the downstream neural networks can operate near the fundamental physical limits of the telescope (Flasseur et al.).
Statistical Decision Frameworks at the Photon Level
At the bleeding edge of high-contrast imaging—such as imaging Earth-like worlds with the upcoming Habitable Worlds Observatory (Laginja et al.)—the light levels are so low that detectors must operate in photon-counting mode. Using specialized Electron Multiplying CCDs (EMCCDs), the camera records binary events: whether or not a photon struck a given pixel during an ultra-short exposure (Hu et al.; Dou et al.).
In this regime, traditional Gaussian noise assumptions break down completely. The data consists of binary streams contaminated by detector-induced multiplicative noise. To reliably extract signals without generating false positives, Hu et al. developed a Bernoulli Generalized Likelihood Ratio Test (GLRT). Instead of measuring continuous brightness, the Bernoulli GLRT calculates the statistical probability that a local cluster of binary photon arrivals matches the expected profile of a planet versus residual starlight artifacts.
Similarly, for processed images where residual noise distributions exhibit heavier tails than a standard Gaussian curve, Daglayan et al. demonstrated that modeling the noise via a Laplacian distribution yields a much higher true-positive rate with zero false positives.
Whether at the level of hardware wavefront control, matrix decomposition, or photon-counting statistics, the core philosophy of high-contrast imaging remains the same: by precisely modeling the physical and temporal behavior of starlight noise, astronomers can systematically peel it away to reveal the faint planetary signals hidden underneath.
Key Foundational Work
The Nancy Grace Roman Space Telescope Coronagraph Instrument (CGI) technology demonstration (2020) — contributed 2 concept elements to this direction.
Exoplanet Detection in Starshade Images (2021) — contributed 1 concept elements to this direction.
High-contrast integral field spectrograph (HCIFS): multi-spectral wavefront control and reduced-dimensional system identification (2020) — contributed 1 concept elements to this direction.