Mechanically Supersampled Radiography

• Mechanically supersampled radiography is an imaging strategy that uses controlled sub-pixel shifts to acquire interlaced samples and enhance spatial resolution.
• It employs mechanical motion combined with forward modeling and iterative reconstruction to fuse data from non-uniform detector positions.
• Applications include planar radiography, dual-energy CBCT, interferometry, and X-ray ghost imaging, each improving detail and contrast under various system trade-offs.

Searching arXiv for papers on mechanically supersampled radiography and closely related X-ray supersampling methods. I’m querying arXiv for mechanically supersampled radiography, wobble correction, analyzer-less interferometry super-resolution, and supersampled scanning microscopy. Query: "mechanically supersampled radiography" Mechanically supersampled radiography is an acquisition and reconstruction strategy in which controlled or measured mechanical motion is used to obtain interlaced or irregular samples of an X-ray transmission field at sub-pixel offsets, followed by reconstruction on a grid finer than the native detector pitch. In the recent arXiv literature, the concept appears in planar radiography with mechanically moved high-Z photon-counting detectors, in dual-energy cone-beam CT through sub-pixel shifted binning of a dual-layer flat-panel detector, in analyzer-less X-ray interferometry through detector phase stepping, in scanning transmission X-ray microscopy through high-rate position readout and position-tagged count reconstruction, and in X-ray ghost imaging through a fast-spinning modulation mask coupled to a single-pixel detector (Weigt et al., 28 Jul 2025, Su et al., 2022, Dey et al., 21 Jan 2025, Finizio et al., 19 Dec 2025, Zeng et al., 2 Jan 2026).

1. Conceptual basis

In its most general form, mechanically supersampled radiography uses controlled mechanical motion, such as sub-pixel shifts of the detector, object, or optics, to synthetically increase sampling resolution beyond the detector’s native pixel pitch, and then uses an appropriate forward model and inverse reconstruction to recover a higher-resolution image (Dey et al., 21 Jan 2025). The essential premise is that the detector’s nominal grid need not coincide with the physical sampling grid of the measurements. If the relative motion is known or estimated, measurements acquired at different sub-pixel phases can be fused into a finer effective sampling lattice.

A closely related formulation appears in supersampled scanning microscopy, where the sample position is read out at a rate significantly higher than the vibration spectrum and the transmission image is reconstructed from the recorded list of positions and detector counts. In that framework, image pixels are defined a posteriori from irregularly sampled position–count data rather than from a fixed acquisition raster (Finizio et al., 19 Dec 2025). The same logic transfers directly to radiography whenever transmission measurements can be tagged with sufficiently accurate mechanical state information.

The literature also makes clear that not every mechanically induced displacement is already a supersampling method. In large-scale transmission radiography, detector wobble produces non-uniform, time-varying sampling of the beam profile and line integrals, but the cited wobble-correction work treats that motion as a nuisance to be normalized away, not as an explicit super-resolution reconstruction problem. The paper nonetheless notes that accurate per-sample wobble tracking could, in principle, support non-uniform sampling reconstruction on a finer grid (Rogers et al., 2016). A related misconception appears in grating interferometry: translating the detector produces sub-pixel sampling of the object, whereas translating only the grating changes fringe phase without giving the lung-tissue parameters sub-pixel resolution (Dey et al., 21 Jan 2025).

2. Acquisition geometries and mechanical encodings

One established realization is mechanically moved planar photon-counting radiography. In that setting, a set of radiographic frames is acquired with a photon-counting detector that is physically moved between exposures along a prescribed trajectory; each exposure forms a conventional pixelated image, but the detector is at a slightly different sub-pixel position relative to the object; and a model-based iterative reconstruction reconstructs a high-resolution image on a finer grid than the native detector pixels (Weigt et al., 28 Jul 2025). The same study distinguishes short-range supersampling, optimized for maximal resolution gain through displacements on the order of one pixel, from long-range supersampling, in which trajectories span multiple pixels and simultaneously average out pixel defects, ASIC border effects, and spectral inhomogeneities.

A second acquisition geometry is sub-pixel shifted binning in dual-layer flat-panel detectors. The suRi framework replaces conventionally aligned binning with 1D or 2D sub-pixel shifted binning, specifically half-pixel shifted binning in the reported study, to double the spatial sampling rate during dual-energy CBCT acquisition. Because the shifted binned pixels cover overlapping but non-identical sets of native pixels across layers, the stacked measurements can be written as a composite sampling matrix acting on an underlying higher-resolution projection (Su et al., 2022).

A third geometry is detector phase stepping in analyzer-less interferometry. For each phase step, the detector is shifted by a sub-pixel translation in the fringe direction, so multiple low-resolution views sample the same underlying fringe at different phases. These phase-stepped acquisitions are then rasterized to a synthetic high-resolution signal and fed to an iterative reconstruction of attenuation, differential-phase, and dark-field images (Dey et al., 21 Jan 2025).

Scanning implementations use the same principle in continuous motion. In supersampled STXM, the sample is scanned with relaxed instantaneous positioning requirements while the true sample position is measured interferometrically at 4 kHz, detector counts are accumulated between position samples, and each count increment is assigned to the pixel corresponding to the measured position rather than the nominal scan coordinate (Finizio et al., 19 Dec 2025). X-ray ghost imaging provides a more radical encoding: a fast-spinning brass mask with a 32 × 1024 random binary annular pattern rotates at up to 200 revolutions per second, presenting 1024 distinct patterns per revolution to a single-pixel detector. Spatial information is thereby multiplexed into a high-rate temporal signal, yielding X-ray visualization of moving objects at up to 200 frames per second with a resolution of 225 μm (Zeng et al., 2 Jan 2026).

3. Forward models and reconstruction frameworks

The reconstruction literature converges on explicit forward models that incorporate shift-dependent sampling, blur, and counting statistics. In mechanically supersampled planar radiography with a high-Z photon-counting detector, each frame is modeled as a Poisson measurement with mean

$$\lambda_{m,i} = I_{0,m,i} \exp\big(-[A^{(m)} \mathbf{x}]_i\big),$$

where $A^{(m)}$ is the frame-specific system matrix and the in-plane detector shift $\boldsymbol{\delta}_m$ is absorbed into the shifted ray geometry, so that

$$[A^{(m)} \mathbf{x}]_i = \sum_j a_{m,i,j}(\boldsymbol{\delta}_m)\,x_j.$$

The reported implementation uses Maximum Likelihood Expectation Maximization with a distance-driven projector/backprojector, optionally combined with total variation regularization (Weigt et al., 28 Jul 2025).

In supersampled scanning microscopy, the primitive measurements are count increments $\Delta C(t_i)$ associated with average measured positions $p_x(t_i), p_y(t_i)$. Reconstruction onto a chosen image grid is then performed by binning the counts and dwell times for all intervals whose measured positions fall inside a given pixel:

$$C_k = \sum_{i \in \mathcal{I}_k} \Delta C(t_i), \qquad \tau_k = \sum_{i \in \mathcal{I}_k} \Delta t, \qquad \hat I_k = \frac{C_k}{\tau_k}.$$

This formulation makes the grid a reconstruction choice rather than an acquisition constraint (Finizio et al., 19 Dec 2025).

Analyzer-less interferometric supersampling adds a nonlinear fringe model. The reference fringe for phase step $k$ is written as

$$g_r^k(x)=A \cos\left[\frac{2\pi}{p_D}(x+t_k)\right] + B,$$

and the object-modulated fringe as

$$g_{obj}^k(x)=\left\{A e^{-S(x+t_k)} \cos\left[\frac{2\pi}{p_D}(x+t_{ph}(x)+t_k)\right] + B\right\}e^{-\mu T(x+t_k)}.$$

After blurring by source and detector point-spread functions and sampling on under-resolved detector pixels, the unknown fields $A^{(m)}$0, $A^{(m)}$1, and $A^{(m)}$2 are recovered by Poisson maximum-likelihood optimization with gradient descent (Dey et al., 21 Jan 2025).

Ghost-imaging realizations use a multiplexed linear model,

$A^{(m)}$3

where each row of $A^{(m)}$4 is a flattened random binary mask pattern and $A^{(m)}$5 is the corresponding bucket-measurement vector. The reported reconstruction minimizes a total-variation-regularized objective,

$A^{(m)}$6

which is solved by TVAL3 (Zeng et al., 2 Jan 2026).

These formulations share a common structure: mechanically induced diversity enters either as explicit spatial shifts in the system matrix or as a sequence of coded measurements, and resolution enhancement is obtained by inversion of the joint acquisition model rather than by image-space interpolation alone.

4. Position metrology, calibration, and degradation modeling

Successful mechanically supersampled radiography depends on recovering the actual sampling positions with sub-pixel precision. In the planar PCD study, the detector shifts are estimated from the image data by Enhanced Correlation Coefficient registration applied to STC-calibrated projections. This removes any requirement for high-precision mechanical encoders and allows tolerance to imperfect stage accuracy, repeatability errors, thermal drift, and slight deviations from planned trajectories (Weigt et al., 28 Jul 2025).

A closely related registration strategy appears in ImPASS, where successive low-resolution frames satisfy

$A^{(m)}$7

and phase correlation estimates $A^{(m)}$8 from the normalized cross-power spectrum, followed by centroiding of the correlation peak for sub-pixel localization. In the reported microscopy experiments, commanded diagonal steps of $A^{(m)}$9 produced measured average translations of about $\boldsymbol{\delta}_m$0 detector pixels per step, and the method was used to compensate for backlash and other mechanical imperfections (Caron, 29 Aug 2025).

When the motion is not controlled but merely observed, calibration becomes more elaborate. In cargo radiography, Beam Position Detectors sample the cross-section of the X-ray beam, Gaussian models are fitted in air-only scans, sensor sensitivities and detector endpoint offsets are calibrated, and a Random Regression Forest estimates instantaneous beam position together with uncertainty. Those estimates are then fused into a wobble-correction pipeline derived from the image-formation model (Rogers et al., 2016). Although that framework is presented for precision restoration rather than super-resolution, it supplies exactly the ingredients required by a supersampling-oriented reinterpretation of wobble.

Radiographic super-resolution work also stresses that realistic degradation models matter. AID-SRGAN introduces a composite degradation model for radiographic images, a denoising module with an attention mechanism, and a separate-joint training strategy for super-resolution from noisy low-resolution radiographs (Huang et al., 2022). This suggests that mechanically supersampled acquisition alone does not eliminate the need to model blur, noise, and artifacts: sub-pixel sampling only recovers frequencies that survive the physical transfer function, and reconstruction quality depends on accurate representation of detector PSF, source blur, and scanner-specific corruption processes.

5. Detector platforms and system trade-offs

Mechanically supersampled radiography is tightly linked to detector design. In planar radiography, high-Z photon-counting detectors offer high intrinsic contrast and dose efficiency because X-rays are directly converted to charge carriers, but spatial resolution remains limited by pixel pitch and detector physics. The reported 75 μm-pitch GaAs detector illustrates the central trade-off: smaller pixels improve spatial resolution but can degrade spectral performance through more charge sharing, higher capacitance, and stronger non-uniformities, whereas larger pixels improve spectral fidelity but cap attainable resolution (Weigt et al., 28 Jul 2025).

Dual-layer flat-panel detectors expose a different trade-off. In medical CBCT, the native receptor array is often binned to increase signal readout speed by over 4–9 times, sacrificing spatial resolution by at least 50%–67%. The suRi method addresses this bottleneck by combining dual-energy acquisition with half-pixel shifted binning, preserving high readout speed while reintroducing fine spatial sampling through reconstruction (Su et al., 2022).

Ultrafast CMOS development broadens the design space further. The cited CMOS program reports more than 10× quantum-efficiency improvement for >10 keV X-rays through a photon-attenuation layer, a compact block-wise readout architecture reaching 76 kfps, and a burst-mode imager with sequential transfer gates, targeted frame rates of at least 20 Mfps, charge transfer in about 12 ns, and input-referred noise below $\boldsymbol{\delta}_m$1 (Yue et al., 2023). Although that paper emphasizes data-enabled position super-resolution rather than mechanical supersampling, its detector characteristics are directly relevant to any shift-based scheme that must acquire many short, low-noise frames.

At the opposite architectural extreme, X-ray ghost imaging replaces a 2D detector entirely with a single-pixel detector and a mechanically modulated illumination basis. This bypasses the conventional trade-off between pixel count and readout time, since the single-pixel detector is read continuously at 1 MHz while the spinning mask provides the spatial degrees of freedom (Zeng et al., 2 Jan 2026).

6. Demonstrated performance, applications, and limitations

The strongest direct evidence for mechanically supersampled planar radiography comes from the high-Z photon-counting study. Using a 75 μm-pitch GaAs detector and MLEM reconstruction, the reported edge-spread 10%–90% width improves from about 93.75 μm with linear interpolation to about 31.25 μm with iterative reconstruction, and the MTF10% improves from about 9.43 lp/mm, corresponding to an effective resolution of about 106.1 μm, to about 30.98 lp/mm for short-range supersampling and about 29.6–29.7 lp/mm for long-range supersampling, corresponding to about 32.3–33.7 μm. Long-range trajectories also increase CNR in both transmission and spectral-ratio images and reduce graininess, spectral non-uniformity, ASIC border artifacts, and sensor defects. In comparison with a clinical mammography system, the reconstructed images show sharper detail and more homogeneous contrast at comparable or reduced dose (Weigt et al., 28 Jul 2025).

In CBCT, suRi reports that synthesized monochromatic CT imaging demonstrates a spatial image resolution improvement of 46.15% relative to FBP from 1×2 binned data, and 16.33% relative to an iterative aligned-binning baseline, while essentially recovering the resolution of true 1×1 data from shifted-binning measurements (Su et al., 2022). In analyzer-less interferometry, simulations with a fringe period $\boldsymbol{\delta}_m$2 and 30 μm or 50 μm detector pixels show that iterative super-resolution reconstruction remains robust even when the detector does not satisfy the Nyquist condition required by traditional fringe-recovery algorithms (Dey et al., 21 Jan 2025).

Mechanically informed motion correction also yields substantial precision gains when the motion is unintended. In wobble-corrected cargo radiography, traverse-mode wobble RMS is reduced from 0.0185 to 0.0054, corresponding to correction of 87% of wobble-induced error, and PSNR improves by 21.3 dB in air-only scans after wobble correction (Rogers et al., 2016). In supersampled STXM, the method achieves overheads of about 10%–15%, largely independent of dwell time, removes cryostat-induced vibrations, and restores near-diffraction-limited imaging of a Siemens star using a reconstruction grid as fine as 2.5 nm (Finizio et al., 19 Dec 2025). In X-ray ghost imaging, the spinning-mask architecture reaches 200 fps at 225 μm resolution without detector dead time, and moving objects remain recognizable up to 7.68 cm/s (Zeng et al., 2 Jan 2026).

The main limitations are consistent across modalities. Resolution enhancement remains bounded by the physical PSF and blur support; mechanical shifts do not create recoverable information beyond what the system transfers (Yue et al., 2023). Accurate shift estimation, detector calibration, and blur modeling are essential; errors in motion tracking or PSF specification directly degrade reconstruction (Weigt et al., 28 Jul 2025, Dey et al., 21 Jan 2025). Dose and SNR remain coupled to the number of shifted frames or coded measurements, so finer supersampling usually means either more exposures or lower per-frame counts (Su et al., 2022, Zeng et al., 2 Jan 2026). Motion during acquisition, stage instability, and computational cost are recurrent practical constraints. These factors explain why the field has developed along two complementary lines: physically grounded acquisition schemes that create genuine new samples, and reconstruction methods that explicitly encode mechanics, detector response, and count statistics rather than relying on interpolation alone.