Ghost Noise: Mechanisms & Mitigation

• Ghost noise is a class of noise processes in correlation-based imaging, involving detector fluctuations, quantization errors, and nonlinear interactions.
• It arises from classical detector noise, quantization artifacts, and stochastic resonance, which collectively degrade image fidelity and SNR.
• Mitigation strategies such as colored noise pattern design, deep learning denoising, and adaptive correlation techniques enhance imaging performance under challenging conditions.

Ghost noise encompasses a class of fundamental noise processes and artifacts arising in correlation-based imaging and nonlinear signal detection, most notably in ghost imaging, quantum ghost spectroscopy, and related computational and physical systems. It is both a practical and theoretical construct, describing non-idealities such as detector noise, environmental background, quantization artifacts, and emergent spectral features generated through nonlinear interactions between input signals and noise. Ghost noise is intrinsically linked to the performance limits of ghost imaging protocols—classical and quantum—and has also motivated algorithmic solutions and pattern engineering strategies for noise resilience and information extraction in severely resource-limited regimes.

1. Noise in Ghost Imaging: Statistical Origins and Characterization

In computational ghost imaging (CGI), the canonical reconstruction is based on second-order fluctuation correlations between bucket detector measurements and a known (or measured) set of spatial illumination patterns. For $N$ realizations, the standard estimator is

$$\hat{T}(x,y) = \frac{1}{N}\sum_{i=1}^N [S_i - \langle S\rangle][P_i(x,y) - \langle P(x,y)\rangle]$$

where $S_i$ aggregates the object’s response to the $i$th pattern $P_i(x,y)$. In practice, each $S_i$ and $P_i(x,y)$ may contain noise terms—including additive background, pattern drift, dark current, and quantization error—so that the measured signals are

$$S_i^{\text{meas}} = S_i^{\text{true}} + N_i^S,\quad P_i^{\text{meas}} = P_i^{\text{true}} + N^P_{i,x,y}$$

where the $N_k$ are typically modeled as zero-mean, independent, white noise processes (Shimobaba et al., 2017). The superposition of these stochastic terms (here and in the finite-$N$ pattern sum) is termed "ghost noise."

Signal-to-noise ratio (SNR) metrics—defined as

$$\mathrm{SNR} = \frac{\mu_{\rm sig}}{\sigma_{\rm sig}}$$

with $\mu_{\rm sig}$ (mean over object region) and $\sigma_{\rm sig}$ (standard deviation)—are used throughout to quantify noise suppression (Nie et al., 2020, Nie et al., 2020).

2. Ghost Noise Mechanisms: Classical, Quantization, and Nonlinear Effects

Classical Detector and Environmental Noise

Classical ghost noise sources include Poissonian photon shot noise, additive background (ambient light, electronic noise), detector dark current, and uncorrelated environmental fluctuations. These sources contribute to both the mean and variance of reconstructed features, with uncorrelated noise tending to degrade SNR in both regular and ghost imaging. Notably, GI protocols can exploit the cross-correlation structure to suppress uncorrelated detector noise more effectively than direct imaging, especially in the low-photon, high-noise regime (Ganesan et al., 2022). In this context, "ghost noise" refers chiefly to those noise components that are erased or suppressed by the correlation process.

Quantization Artifacts and Dithering

Low-bit quantization (binary, few bits) in either the signal or reference arms introduces systematic signal-dependent distortions (quantization errors $\epsilon(y)$ correlated with underlying intensities), which can severely degrade ghost image contrast. Dithering—injecting random noise prior to digitization—decorrelates quantization error from the input, thus suppressing these ghost noise effects even under extreme bit limitations. An optimal dither amplitude of $a \approx 0.5\Delta$ to $1.0\Delta$ (one to two least-significant bits) is found empirically to maximize recovered SNR (Li et al., 2017).

Ghost Stochastic Resonance (GSR): Emergent Frequencies

A distinct literature describes "ghost noise" as the generator of emergent frequencies absent from any input (ghost stochastic resonance). Here, in nonlinear systems driven by multiple periodic inputs and noise, linear interference produces constructive peaks at the ghost frequency $f_{\rm ghost}=|f_2-f_1|$, and a threshold detector—mediated by noise—transduces these peaks into output events at $f_{\rm ghost}$ (Balenzuela et al., 2011). The ghost noise thus refers to the combination of background stochastic processes and nonlinearity that seeds new, "ghost" spectral features in the system's response.

3. Experimental Techniques for Ghost Noise Suppression and Exploitation

Pattern Engineering: Colored Noise and Correlated Structures

Recent advances demonstrate improved resilience to ghost noise by shifting from uncorrelated (white) to correlated (colored, e.g., pink $1/f$) noise patterns for illumination. Pink noise patterns engender strong positive cross-correlation between nearby pixels, boosting the signal-carrying term in the correlation estimator and reducing sensitivity to motion, blur, and environmental noise:

• In pink-noise CGI, SNR exceeds that of white noise by $4$–$5\times$ under matched sampling conditions, and robust imaging is possible with an order of magnitude fewer patterns, even for moving objects or persistent background interference (Nie et al., 2020, Nie et al., 2020).
• Orthonormalization of colored noise patterns (Gram–Schmidt) achieves sub-Nyquist sampling, maintaining SNR and image fidelity at sampling ratios $\beta = N/N_{\rm pixel} \sim 0.05$–$0.1$, while suppressing noise cross-talk (Nie et al., 2020).

Algorithmic Approaches: Deep Learning and Statistical Inference

Deep learning (U-Net, DnCNN architectures) provides effective, data-driven denoising for CGI. By training on pairs of noisy and clean CGI reconstructions, networks learn to suppress both classical and ghost noise artifacts, notably outperforming conventional post-processing filters and enabling high-quality CGI with reduced pattern count (Shimobaba et al., 2017). Noise2Ghost (N2G), a self-supervised method, partitions measurement data and exploits sub-reconstruction cross-prediction to robustly denoise even in the absence of ground truth, outperforming total variation, DIP, and implicit neural representation baselines particularly at extreme noise and compression (Manni et al., 14 Apr 2025).

Specialized Protocols: First-Photon, Differential, and Instant Ghost Imaging

• First-Photon Ghost Imaging (FPGI) uses the statistics of first-photon arrivals under structured patterns, enabling reconstructions with $<1$ photon per pixel—orders of magnitude below conventional noise thresholds. SNR optimization involves a trade-off between statistical (shot) noise and background noise, with an optimal pattern sparsity $Sp\sim0.01$ (Liu et al., 2017).
• Differential ghost imaging and quantum differential protocols engineer the combination of bucket and reference channels to suppress the impact of losses and detector noise. In low-brightness, quantum SPDC regimes, optimal linear combinations (optimized DGI) achieve SNR above classical and even canonical quantum GI (Losero et al., 2019).
• Instant Ghost Imaging (IGI) replaces the standard zero-lag correlation by a correlation of temporal differences across frames, effectively nulling slowly-varying background fluctuations and delivering resilience against arbitrarily strong optical noise (Yang et al., 2020).

4. Quantitative Comparisons and Noise Performance Metrics

Closed-form SNR expressions for GI and regular imaging (RI), including explicit dark-count noise dependence, have been derived for single-pixel object cases: $$\mathrm{SNR}_{\mathrm{GI}} = \frac{\lambda_s\sqrt{N}}{\sqrt{D_G(\lambda_s,\lambda_d,P)}}$$ where $D_G$ aggregates source and noise contributions, and $\lambda_s$, $\lambda_d$ are mean source and detector counts per shot, respectively. GI surpasses RI in SNR when $\lambda_d \gtrsim \lambda_s$ and detector noise dominates shot noise; GI’s correlation structure cancels uncorrelated dark counts that, in RI, directly limit image fidelity (Ganesan et al., 2022). Experimental studies confirm that with colored noise patterns or orthogonalized bases, SNR and PSNR can be enhanced by $10$–$15$ dB over standard illumination, with object recapture possible even under severe interference (ambient lamp, diffusers, detector noise) (Nie et al., 2020, Nie et al., 2020).

5. Ghost Noise in Quantum Ghost Spectroscopy: White vs. Colored Noise Response

In quantum ghost spectrometry protocols utilizing SPDC-generated frequency-entangled photon pairs, the effect of noise is governed by the spectral structure:

• White noise (frequency-independent detector background) is efficiently suppressed by coincidence detection and does not degrade spectral resolution; the system remains robust even at low count rates (Sansoni et al., 20 Mar 2025).
• Colored noise, especially from double-pair emissions in SPDC (multi-pair accidental coincidences), matches the source joint spectral amplitude $f(\omega_s, \omega_i)$, irreducibly broadens the ghost spectrum, and degrades resolving power. A colored noise fraction $\gtrsim 25\%$ can destroy resolution even for relatively large spectral separations. This asymmetry is experimentally confirmed and dictates the operational photon flux regime for practical quantum ghost spectroscopy (Sansoni et al., 20 Mar 2025).

6. Theoretical and Practical Implications Across Modalities

• Ghost noise defines the ultimate limits of ghost imaging and related modalities in photon-efficiency, dose minimization, and hostile noise environments.
• Pattern design (colored-noise ensemble, Gold matrices), adaptive algorithmic denoising, and optimized correlation operators collectively expand the operational regime into low-light, high-background, and information-sparse scenarios, fundamental to biological imaging, remote sensing, and X-ray/EUV spectral imaging (Zhao et al., 2019, Nie et al., 2020, Manni et al., 14 Apr 2025).
• In nonlinear systems, ghost noise through stochastic resonance provides a mechanism for extracting information at emergent frequencies, underlying phenomena in neuroscience, laser physics, and climatology (Balenzuela et al., 2011). Here, the interplay between noise intensity and system nonlinearity creates resonance at ghost frequencies with SNR optimized at a critical noise level $D_{\rm opt}$.

7. Open Problems and Directions

• Further theoretical work is required to precisely characterize the interplay between colored pattern statistics, measurement noise, and reconstruction SNR, particularly under severe undersampling or for dynamic, three-dimensional, or scattering objects.
• Efficient, hardware-compatible algorithms for real-time pattern generation, orthonormalization, and adaptive denoising in the presence of time-varying ghost noise remain underdeveloped (Nie et al., 2020, Nie et al., 2020).
• In quantum imaging and spectroscopy, mitigation strategies for colored noise (multi-pair emission) such as photon-number-resolving detection, tight temporal gating, or higher-order correlation protocols present promising but technically challenging avenues (Sansoni et al., 20 Mar 2025).
• Integration of self-supervised deep learning strategies and compressive sensing holds potential for overcoming increasing background and ghost noise in emerging imaging scenarios (e.g., nano-scale x-ray, in-vivo samples) (Manni et al., 14 Apr 2025).

Collectively, the ghost noise paradigm not only delineates the noise floor in correlation-based sensing but also catalyzes innovations in pattern design, algorithmic reconstruction, and quantum measurement protocols, forming a central unifying theme across classical and quantum information acquisition systems.