Interferometric Coherence Images

• Interferometric coherence images are defined as maps of spatial coherence using first- and second-order correlations to encode both amplitude and phase information.
• They are generated using intensity interferometry, ptychographic reconstruction, and deep learning techniques to enhance imaging in optical, radar, and quantum applications.
• The method improves noise suppression and phase retrieval in scattering media while balancing resolution with statistical stability and robustness.

Interferometric coherence images are image constructs in which the spatial distribution of coherence (either first- or second-order) is mapped, enabling reconstruction or analysis of physical observables from interferometric measurements. These images are foundational in optical and radar imaging, phase retrieval, noise suppression, synthetic aperture radar (InSAR) analysis, advanced quantitative phase microscopy, and robust imaging through random or scattering media. Coherence images are generated through processing of intensity or field correlations—typically through either second-order statistics (such as those utilized in intensity interferometry) or first-order mutual coherence analyses—yielding a representation that encodes both amplitude and phase information, or, in some applications, surrogate proxies for phase stability or signal quality.

1. Fundamentals of Interferometric Coherence Imaging

Interferometric coherence images are formed by evaluating the correlation between optical (or electromagnetic) field intensities or amplitudes at distinct spatial, temporal, or spectral points. In second-order coherence imaging (intensity interferometry), the principal object is the normalized intensity–intensity correlation function: $$g^{(2)}(r_1, r_2) = \frac{\langle I(r_1) I(r_2) \rangle}{\langle I(r_1) \rangle \langle I(r_2) \rangle}$$ For stationary, zero-mean Gaussian (thermal) fields, the Siegert relation connects $g^{(2)}$ to the first-order degree of coherence $g^{(1)}$: $g^{(2)}(r_1, r_2) = 1 + |g^{(1)}(r_1, r_2)|^2$ where $g^{(1)}(r_1, r_2)$ is the normalized field autocorrelation. For an optical field, $g^{(1)}$ is intimately connected to the mutual coherence and, under the van Cittert–Zernike theorem, is proportional to the spatial Fourier transform of the object's transmission or reflection function. Thus, ensemble-averaged intensity correlations as a function of detector position separation directly encode the squared modulus of the object’s spatial frequency spectrum. This is the backbone of intensity interferometric imaging and serves as the basis for "coherence images" in intensity fluctuation analysis (Wang et al., 2017).

2. Second-Order Ptychography: Formation and Reconstruction Pipeline

Second-order (intensity-based) ptychography extends traditional ptychographic iterative reconstruction into regimes of incoherent illumination by leveraging spatially resolved measurements of intensity–intensity correlation functions. The crucial steps, as crystallized in (Wang et al., 2017), are:

1. Acquisition: At each probe (illumination) position, multiple short-exposure intensity frames are acquired to statistically sample intensity fluctuations; typically $N \sim 10$ to $1000$ realizations per scan position to ensure sufficient averaging.
2. Computation of $g^{(2)}$: For each probe position $R_j$, compute

$g^{(2)}$0

for all relevant $g^{(2)}$1.

1. Fourier Transformation: Subtract $g^{(2)}$2 and perform a 2D Fourier transform to extract the amplitude $g^{(2)}$3, which estimates the exit-wave spectrum modulus per scan.
2. Iterative Phase Retrieval: An extended ptychographic iterative engine (ePIE) algorithm is used: the measured spectrally resolved amplitudes constrain the modulus of the estimated object’s spatial spectrum during each step, iterative object and probe updates proceed, and a loose support constraint defined by a mask $g^{(2)}$4 regularizes the solution and accommodates moderate probe mislocalization.
3. Convergence and Image Formation: The normalized residual error

$g^{(2)}$5

is minimized, and the converged estimate $g^{(2)}$6 represents a complex-valued image—yielding both amplitude and phase.

Resolution is dictated by the maximal observed correlation separation $g^{(2)}$7: $g^{(2)}$8. Signal-to-noise ratio benefits both from the scaling with the square root of the number of frames and the redundancy inherent to overlapping ptychographic sampling (Wang et al., 2017).

3. Relationship to Optical Coherence and Noise Suppression

Spatial and temporal degrees of coherence (DSTCI), characterized by spatial ($g^{(2)}$9) and temporal ($g^{(1)}$0) coherence lengths, play a critical role in image quality for interferometric modalities. The decomposition of arbitrary low-coherence fields into coherent modes allows explicit quantification of the impact of partial coherence on speckle and background noise (Shin et al., 2016). The key findings are:

• Speckle Noise Scaling: $g^{(1)}$1.
• Parameter Tuning: Decreasing both spatial and temporal coherence lengths drives synergistic suppression of speckle, with spatial coherence length contributing roughly four times the effect of temporal coherence length on speckle reduction.
• Implementation: Coherence is tuned via pupil scanning (spatial) and wavelength sweeping or filtering (temporal), with the optimum balancing maximal contrast and speckle suppression (Shin et al., 2016).

These principles are directly applicable in interferometric coherence image formation, optimizing contrast, and minimizing artefactual noise.

4. Extension to Noisy and Inhomogeneous Media

Interferometric (coherence-based) imaging is robustified against random media or strong noise through the use of local cross-correlation windowing and coherence gating. The two-point coherent interferometric (CINT) function: $g^{(1)}$2 preserves statistical stability even in highly scattering media (Borcea et al., 2021). With additional phase retrieval or eigenvector extraction from the two-point (off-diagonal) coherence kernel, both high resolution (comparable to the homogeneous medium limit) and robustness to clutter are achieved. Coherence images in these contexts are thus not mere measures of local field similarity, but operational intermediate representations enabling statistically stable inversion—critical in, for example, radar remote sensing and medical ultrasound (Borcea et al., 2021).

5. Machine Learning Approaches to Coherence Image Estimation

Deep learning, especially convolutional neural networks (CNNs), has been deployed to estimate coherence images from interferometric SAR data (Mukherjee et al., 2020, Sun et al., 2019, Mukherjee et al., 2020). These approaches share several features:

• Coherence Estimation: The normalized complex correlation

$g^{(1)}$3

is computed, commonly via a patch-based estimator ($g^{(1)}$4 sliding windows).

• Deep Architectures: Networks process real and imaginary SAR channels, often following extensive preprocessing (outlier saturation, normalization, denoising via autoencoders), and output per-pixel coherence probability maps.
• Label Generation: Spatially balanced and smooth labels are constructed using Markov random fields or reference stack-based estimators, mitigating overfitting to noisy or incoherent regions (Mukherjee et al., 2020, Sun et al., 2019).
• Hybrid Physics–DL Models: Advanced models such as CoHNet fuse neural estimators with physics-constrained surrogate models (e.g., RVoG inversion) to enforce physical plausibility and interpretability in the coherence–decoration–height inference chain (Mahesh et al., 14 Apr 2025).

Performance gains include lower root-mean-squared error (RMSE) in coherence and phase estimation, finer discrimination of phase discontinuities, and increased computational efficiency compared to classical (boxcar, NL-SAR) techniques (Mukherjee et al., 2020).

6. Variants and Specializations: Optical and Quantum Coherence Imaging

Interferometric coherence images underpin a wide range of specialized imaging modalities:

• Dynamic Full-Field OCT (D-FFOCT): Temporal analysis of pixelwise interferometric signals, specifically through calculation of per-pixel temporal standard deviation or autocorrelation, yields maps of metabolic dynamics with cellular and subcellular specificity (Apelian et al., 2016).
• Cepstrum-Based Interferometric Microscopy (CIM): The spatial-shifting cepstrum algorithm numerically decouples overlapping field spectra from multiple shifted holograms, supporting tripled field of view in phase imaging and robust recovery of three independent complex amplitude fields (Rubio-Oliver et al., 17 Jan 2025).
• Quantum Optical Induced-Coherence Tomography: Second-order induced coherence between photon pairs (signal and idler) enables depth-resolved imaging (tomography) even with undetected photons, as the interference visibility in the detected channel is dictated by the heralding efficiency; both time-domain (delay scanning) and frequency-domain (spectral FT) approaches are supported (Kim et al., 2023).

These variants extend coherence imaging into regimes where conventional techniques are inapplicable or inefficient, such as ultra-low light, inaccessible spectral bands, or high-throughput biological microscopy.

7. Limitations, Interpretation, and Practical Considerations

Several caveats and methodological clarifications are critical:

• Fringe Observability: The presence of strong fringes or spatially structured visibility in interferometric images does not necessarily imply microscopic (quantum or classical) coherence of sources. Overlapping point spread functions of a patterned intensity distribution can produce high visibility even for fully incoherent emission (e.g., in exciton imaging) (Stolz et al., 2017).
• Noise Robustness: Coherence gates (temporal, spectral, spatial) can suppress structured and broadband noise well beyond what is achievable via spectral filtering, as in optical coherence filtering under dominant noise ($g^{(1)}$5 object intensity) (Mohta et al., 10 May 2026). Experimental alignment, drift stability, and interferometric path delays must, however, be matched to the relevant coherence lengths to avoid loss of signal or excessive filtering.
• Resolution vs. Stability Trade-off: Masking of interferometric data to enforce coherence gating in strongly inhomogeneous or scattering media enhances statistical stability but necessarily degrades achievable resolution (cross-range $g^{(1)}$6, depth $g^{(1)}$7) compared to the homogeneous case (Moscoso et al., 2016, Borcea et al., 2021).
• Machine Learning Artifacts: Deep network models can exhibit class imbalance or edge misclassification without robustly constructed label sets and judicious architectural design. Physics-informed loss functions anchor outputs to plausible physical regimes (Mahesh et al., 14 Apr 2025).

Optimization of interferometric coherence images is therefore application- and context-specific, balancing hardware, measurement design, statistical averaging, and algorithmic reconstruction to achieve the desired trade-offs in resolution, signal-to-noise, and interpretability.