Dynamic Range Expansion Techniques

Updated 30 March 2026

Dynamic Range Expansion is a collection of strategies in signal processing and imaging designed to boost the measurable range between peak signals and noise floors.
Advanced algorithms, including deep residual networks, chain exposure synthesis, and diffusion priors, improve HDR recovery with significant gains in PSNR and speed.
Hardware innovations like modulo ADCs, dual-gain amplifiers, and hyperbolic windowing enable high-fidelity acquisition across diverse fields such as astronomy, biomedical imaging, and quantum sensing.

Dynamic range expansion refers to a collection of signal processing, imaging, and measurement strategies designed to extend the span between the largest and smallest reliably detectable values within a given system. The metric of dynamic range (DR) is ubiquitous across disciplines—imaging, spectroscopy, analog-to-digital conversion, interferometry, and high-precision metrology—expressed typically as the ratio between peak and noise floor (or minimum discernible signal), often in decibels. Dynamic range expansion is of fundamental importance whenever faint phenomena are observed in the presence of dominant signals or where detector, sensor, or quantizer limitations would otherwise truncate, clip, or saturate critical information. Recent advances encompass both algorithmic (e.g., deep learning methodologies, multi-exposure fusion, self-calibration), hardware-based (e.g., modulo/folding ADCs, dual-gain amplifiers), and quantum/squeezing-based approaches.

1. Mathematical Definitions and System-Specific DR Metrics

In imaging and signal acquisition, the dynamic range is canonically defined as

$\mathrm{DR} = \frac{S_{\max}}{S_{\min}}$

where $S_{\max}$ is the greatest quantifiable signal and $S_{\min}$ is the noise floor or least discernible change. For images, this takes the form

$\mathrm{DR}_{\text{image}} = \frac{\max_{i} x_i}{\min_{i:x_i>0} x_i}$

with $x$ the pixel intensity vector (Aghabiglou et al., 2022). In phase-imaging systems (QPI), DR is quantified as the largest divided by smallest measurable optical path delay (Toda et al., 2020); in radio astronomy as the ratio of peak flux to map rms noise (Sridharan et al., 20 Apr 2025). In ADC systems, DR is the ratio between largest recoverable voltage and rms noise at the digitizer input (Yin, 2015).

Unit conversions are system-dependent: for instance, in interferometry, DR is often limited by fringe wrapping (e.g., $2\pi$ phase) or detector full-well, while in electronic readouts, saturating the high-gain channel necessitates fallback to low-gain amplification.

2. Algorithmic Dynamic Range Expansion in Imaging

2.1 Deep Residual Architectures for Inverse Problems

State-of-the-art dynamic range expansion for inverse imaging—especially in radio astronomy—relies on residual deep network series (Aghabiglou et al., 2022). The core mechanism is a decomposition: $x = \sum_{k=1}^K r^{(k)}$ where each residual $r^{(k)}$ captures signal components missed by the prior stages. At each iteration, a back-projected data residual is computed and fed to a trained DNN ( $G^{(k)}$ ), yielding an updated estimate. Training proceeds stage-wise with fixed early networks, minimizing an $\ell_1$ -norm loss against ground truth. Empirical results show that as few as $K=4$ residual terms achieve SNR/logSNR on par with computationally intensive Plug-and-Play methods, while providing three orders of magnitude acceleration and robust recovery across a wide DR ( $10^4–10^6$ ).

2.2 Deep Chain Exposure Synthesis and CNN Expansion

The "Deep Chain HDRI" framework (Lee et al., 2018) decomposes the mapping from an LDR image to an HDR target into a chain of subnetworks, each effecting a single stop ( $\pm 1$ EV) shift in exposure, thereby synthesizing a virtual multi-exposure stack. These intermediate-exposure images are fused by classical Debevec–Maluesta HDR merging, leveraging the invertibility of camera response for physical plausibility. Loss functions combine pixelwise reconstruction and total variation. Quantitative evaluation yields 6–10 dB improvements in PSNR over classical ITM methods.

ExpandNet (Marnerides et al., 2018) uses a triple-branch architecture to fuse local, medium-, and global-scale features, enabling hallucination of missing intensities clipped by the LDR acquisition or tone mapping. This architecture avoids explicit upsampling and is validated to outperform both classical and standard U-Net baselines in metrics including PSNR, SSIM, and HDR-VDP.

2.3 Diffusion Priors for Extremes Restoration

The recent Sagiri approach (Li et al., 2024) introduces a two-stage process: an initial color and brightness mapping via a SwinIR-based transformer aligns the input's statistics with true HDR, followed by a masked diffusion prior for synthesizing semantically and texturally plausible content in saturated regions. Histogram, frequency-domain, and SSIM-based losses ensure perceptual realism and statistical fidelity. As a plug-in, the generative stage can enhance any LDR restoration output, markedly improving both perceptual and reference-driven scores.

3. Hardware-Based DR Expansion: Modulo ADCs, Dual-Gain, and Windowing

3.1 Modulo (Folding) ADCs and Unlimited Sensing

Digitization of analog signals with DR demands far beyond conventional quantizer rails necessitates modulo ADC architectures. The signal is analog-folded: $y(t) = \mathcal{M}_\lambda\{x(t)\} = x(t) - 2\lambda\,\big\lfloor\frac{x(t)+\lambda}{2\lambda}\big\rfloor$ This preserves all input energy within a window $[-\lambda,\lambda]$ , avoiding saturation or clipping. Unfolding proceeds by reconstructing the integer folds $k[n]$ via high-order differencing or convex programs, then inverting via

$x[n] = y[n] + 2\lambda k[n]$

The hardware realization now spans from wideband prototypes achieving eight- to hundred-fold DR increases (Mulleti et al., 2023, Li et al., 27 Nov 2025) to integrator-based designs that eliminate digital counters for bandwidth-unlimited folding and achieve up to 60-fold DR gain and ~30 dB SINAD improvements (Zhu et al., 2024). The sampling rate required is dictated by the signal's peak slew and the folding threshold ( $T\le 2\lambda/\alpha$ ), not by amplitude.

3.2 Dual-Gain Preamplification

In rare-event detection (e.g., LZ dark matter TPCs), DR expansion is essential for discriminating between single-photon events and large S2 pulses. This is implemented by splitting the PMT output into high-gain and low-gain amplifiers, digitizing both, and dynamically selecting the unsaturated channel (Yin, 2015). This architecture achieves coverage from ~1 keV (three photoelectrons) to several MeV (thousands of photoelectrons) without information loss.

3.3 Hyperbolic Spectral Windowing

DFT-based spectral analysis benefits from the hyperbolic window family (Stewart, 2014), which by an adjustable exponent $\alpha$ and warp parameter $s$ , enables tuning between main-lobe width (sensitivity, ENBW) and sidelobe suppression (DR). For fixed ENBW, increasing $\alpha$ steepens the roll-off and lowers the sidelobe floor, enabling detection of tones 40–60 dB below the peak—constituting a continuous, lossless dynamic-range expansion in the spectral domain.

4. Multi-Exposure and Adaptive Shifting in Physical Measurement

4.1 Multi-Exposure Fusion

Classical HDR techniques in X-ray ptychography employ tiered exposure times and beam stops to capture strong and weak scattering in separate frames, scaled and merged pixelwise (Rose et al., 2016). The dynamic-range improvement is proportional to the ratio of total exposures ( $t_{max}/t_{min}$ ), with real experimental gains accounting for noise, masking, and saturation, e.g., a 76× DR enhancement and spatial resolution boost from 50 nm to 18 nm.

4.2 Dynamic Range Shift with Wavefront Shaping

ADRIFT-QPI (Toda et al., 2020) adapts phase imaging DR by recording the static background phase via standard DH, programming an SLM to cancel this background, and then increasing illumination to measure the residual (small) dynamic variation via dark-field detection. DR increase scales as $1/\phi_{res,max}$ , yielding measured gains of 6.6× in sensitivity and ~44× faster acquisition.

5. High-Precision and Quantum-enabled DR Expansion

In quantum metrology, squeezing and deamplification strategies decouple the trade-off between local sensitivity ( $\Delta\phi$ ) and global dynamic range ( $DR$ ). Initial spin-squeezing (two-axis counter-twisting) lowers $\Delta\phi$ , while subsequent quantum deamplification expands the invertible (unambiguous) phase interval: $\text{DR} = \pi e^{r_2},\quad \Delta\phi = (1/\sqrt{N})e^{-r_1}$ with squeezing parameters $r_1$ , $r_2$ . Sequential deamplification composes DR enhancements multiplicatively and can be hybridized with quantum amplification to mitigate detector noise (Liu et al., 2024). This protocol is implementable within the coherence and squeezing limits of contemporary AMO platforms.

6. Application Domains and Impact

Dynamic range expansion underpins advances across multiple fields:

Radio astronomy: Enables high DR imaging (up to 45 dB at 8 GHz) critical for deep-field surveys, with self-calibration and pointing calibration strategies directly formalizing the solution interval choice to guarantee DR (Sridharan et al., 20 Apr 2025).
Analog hardware design: Folding ADCs and integrator-modulo architectures are enabling sub-Nyquist, high-fidelity capture of signals whose amplitude varies by orders of magnitude, critical in biomedical signal acquisition, radar, and spectrum analysis (Mulleti et al., 2023, Zhu et al., 2024, Li et al., 27 Nov 2025).
Imaging and computational photography: Modern neural architectures and diffusion-based priors are reconstructing information lost in saturated LDR captures to approach the true radiometric range of a scene (Lee et al., 2018, Marnerides et al., 2018, Li et al., 2024).
Quantum sensing: Decoupling local phase sensitivity from maximum measurable range allows atomic clocks and interferometers to operate robustly in regimes that formerly required exclusive trade-offs (Liu et al., 2024).

7. Limitations, Trade-offs, and Future Directions

While algorithmic and hardware methods now yield DR expansions approaching two orders of magnitude (and quantum protocols remove several classical trade-offs), limits remain due to reconstruction algorithm complexity, calibration accuracy, component nonidealities, and physical non-linearities. For modulo ADCs, the maximum fold rate is limited by the feedback/control-loop delay and analog nonidealities; quantum deamplification is ultimately bounded by decoherence. In practice, application-specific tuning (windowing parameter, gain setting, network depth) remains essential to balance sensitivity and DR given underlying architecture and acquisition constraints.

Ongoing research explores more robust digital unfolding in high-noise/low-oversampling settings, adversarially trained HDR expansions, and fully integrated quantum-classical dynamic-range adaptation across hybrid platforms.

References: