Field-of-View Enhancement Techniques

• Field-of-View enhancement is a multidisciplinary set of techniques that expand imaging and sensory ranges using optical, computational, and neural methods.
• Innovative approaches include neural étendue expanders, off-aperture diffractive elements, and adaptive optical systems to maintain image fidelity over wider angles.
• Key trade-offs involve balancing increased angular coverage with potential resolution loss, driving research in adaptive, multi-modal, and context-aware strategies.

Field-of-view (FOV) enhancement refers to a suite of methodologies—spanning optics, computational imaging, machine learning, acoustic signal processing, and visual neuroprosthetics—designed to expand the angular or spatial extent over which sensory, imaging, or perception systems can acquire, reconstruct, or render information. Increasing the FOV is critical for diverse applications in optics, computational photography, microscopy, medical imaging, virtual/augmented reality, robotics, autonomous systems, auditory scene analysis, and visual prosthesis design. The underlying principles and trade-offs differ by domain, but typically involve addressing physical, engineering, or algorithmic constraints related to wavefront propagation, sampling, spatial encoding, or system hardware.

1. Optical and Computational Principles of FOV Enhancement

Classic optical systems are fundamentally constrained in FOV by physical properties such as lens geometry, numerical aperture, sensor pixel pitch, aperture location for coding elements, and refractive index contrasts. For example, retinal FOV is increased by filling the ocular "image space" with optically dense media; Snell's law ($n_1\sin\theta_1 = n_2\sin\theta_2$) ensures that peripheral rays are refracted inward, thereby increasing angular coverage and shifting images toward the fovea (Doshi et al., 2010).

In digital and holographic imaging, the FOV is limited by the étendue ($G = 4A\sin^2\theta$), which is the product of illuminated area $A$ and the solid angle $2\theta$ subtended by the detected or emitted light. Holographic displays are further restricted by the SLM pixel pitch: the maximal diffraction angle is $$\theta_s = \arcsin\left(\frac{\lambda}{2\Delta_s}\right)$$, implying that the native étendue is $G_s = 4A \sin^2\theta_s$ (Tseng et al., 2021). FOV can be expanded by engineering static or learned optical encoders to increase the diffracted angle while maintaining or even improving image fidelity.

For coded aperture imaging, the placement of diffractive optical elements (DOEs)—not merely on- but off-aperture—de-couples global from local wavefront control. Off-aperture DOEs unmix incident rays based on angle, providing local control over aberrations and improving fidelity at wide angles (Wei et al., 30 Jul 2025). Both ray and wave propagation must be co-optimized, especially for hybrid refractive-diffractive systems, using differentiable models accounting for off-axis aberrations and depth-dependent blur.

In microscopy, spatially variant aberrations limit conventional pupil-plane adaptive optics to a small isoplanatic patch. Placing the deformable mirror conjugate to the aberration source (conjugate AO) can extend the uniform correction across a much larger specimen area (Mertz et al., 2015).

2. Learning-Driven and Programmable FOV Expansion

Leveraging data-driven learning to expand FOV is prominent in holographic, computational, and video-imaging contexts. Neural étendue expanders, physically realized as nanofabricated diffractive elements with submicron pixel pitch, are trained end-to-end (together with SLM phase patterns) using natural image datasets. The learning objective minimizes the difference between the reconstructed holographic intensity and the target image under a perceptually informed low-pass filter:

$\text{minimize}_{\mathcal{E},\mathcal{S}_{1\dots K}} \sum_k \left\| \left|\mathcal{F}(\mathcal{E} \odot \mathcal{U}(\mathcal{S}_k))\right|^2 - T_k \right| \ast f \right\|_2^2$

where $\mathcal{E}$ is the neural expander modulation, $\mathcal{S}_k$ the SLM encoding for target $T_k$, $\mathcal{F}$ is the Fourier transform, $\odot$ is the Hadamard product, $\mathcal{U}$ upsamples to neural expander resolution, and $f$ is a retinal low-pass filter (Tseng et al., 2021). This yields order-of-magnitude FOV increases with peak PSNR exceeding 29 dB on retinal-resolution images.

Data-driven scene extrapolation also appears in NeRF-enhanced outpainting. Here, a Neural Radiance Field (NeRF) is pretrained to render wide-FOV synthetic images based on a sparse set of posed photographs. These NeRF-generated wide-FOV image pairs are then used to train a scene-specific outpainting network; this two-stage pipeline ensures the extrapolated FoV remains faithful to real scene geometry, outperforming purely generative or naive crop-based outpainting for downstream vision tasks (Yu et al., 2023).

In video, temporal coherence and scene continuity are exploited to infer beyond-FoV content. Clip-recurrent transformers with decoupled cross-attention (DDCA), as in the FlowLens system, fuse optical flow–guided warping with transformer-based feature propagation across video clips. This enables “online video inpainting,” reconstructing both peripheral and occluded scene regions, significantly improving semantic understanding and object detection outside the native view (Shi et al., 2022).

3. Domain-Specific FOV Enhancement Techniques

Optics and Sensors

• Meta-Optics and Metalenses: Planar, subwavelength-structured metasurfaces (and diffractive optical elements) allow nearly 180° diffraction-limited FOVs, enabling compact wide-angle sensors for thermal imaging, surveillance, and AR/VR. The phase profile is polynomially engineered for minimal spot size over large angles: $\Phi(\rho)=M \sum_{i=1}^{12} A_i \rho^{2i}$ (Wirth-Singh et al., 2023). Metasurfaces also underpin flat wide-FOV analytical designs, with phase compensation derived from first principles (Yang et al., 2021).
• Multi-Focus Light Sheet Microscopy: Overlapping several axially offset Gaussian light sheets, each with optimal Rayleigh zones, produces a uniform, extended FOV (e.g., 450 μm at axial FWHM $<7.5\,\mu$m across the field), overcoming the classical trade-off between resolution and FOV in conventional single-focus light sheet fluorescence microscopy (Li et al., 2021).

Medical and Scientific Imaging

• Tomography and X-ray Microtomography: Lateral offset of the center-of-rotation in the acquisition geometry doubles the effective FOV without pixel-size sacrifice, exploiting redundancy-weighted backprojection to preserve spatial resolution and support multi-contrast imaging (attenuation, phase, dark-field channels) from the same data (Allan et al., 14 Feb 2025). In CT, field-of-view extension is achieved via linear extrapolation in sinogram space with subsequent deep artifact suppression (U-Net), providing plausible EFoV reconstructions suitable for clinical planning (Fournié et al., 2019).
• Holography and Display: The field-of-view in digital holography is fundamentally dictated by the hologram's numerical aperture, not merely the pixel pitch. High-HNA hologram synthesis (by increasing object size at close synthesis distance and suppressing aliasing) yields larger angular FOVs, further improved by upsampling and aperiodic binary masking (Chae, 2019). For high-fidelity stereoscopic VR/AR, neural étendue expanders enable étendue expansions of $64\times$, supporting immersive, large-eyebox holographic visualization (Tseng et al., 2021).

Acoustic and Speech Processing

• FoV-aware Speech Enhancement: FoVNet partitions the acoustic scene into spatial blocks, applies blockwise beamforming, and selectively enhances all speech sources within a configurable angular sector (the “FoV”), employing learnable in- and out-of-FoV embeddings and an ultra-low-complexity DNN post-filter for efficient smart-glasses–based augmented hearing (Xu et al., 12 Aug 2024).
• Mixture of Experts for Binauralization: In dynamic binaural audio rendering, a signal-dependent mixture of experts framework adaptively blends multiple filter outputs, one per candidate spatial region, using convex optimization over instantaneous residuals. This allows real-time tracking and selective enhancement or suppression of sound from arbitrary directions, supporting continuous talker motion and “world-locked” spatial audio in AR/VR (Mittal et al., 16 Sep 2025).

4. FOV Enhancement in Prosthetics, Robotic, and Perception Systems

Visual prostheses face a strict trade-off between FOV and spatial resolution, given a fixed number of stimulation sites (phosphenes). Simulation in VR shows diminishing returns for increasing FOV: spreading a limited channel set over a wide view impairs object recognition and slows response times due to reduced angular resolution. Empirically, concentrating phosphenes in the central field yields the best usability, even at the cost of peripheral coverage (Sanchez-Garcia et al., 28 Jan 2025).

For environmental perception and robotic applications, learning-augmented FOV expansion is critical for navigation and safety. Both NeRF-based FOV extrapolation (Yu et al., 2023) and transformer-based video inpainting (Shi et al., 2022) are shown to improve semantic reasoning and occlusion prediction beyond what is possible with single-frame raw sensor input.

In simultaneous wireless information and power transfer (SWIPT), the effective FOV of resonant beam alignment is physically limited by retroreflector and resonator design. Embedding a telescope in the transmitter compresses angular deviation at the sensitive gain medium by a factor of $|n_2|$, tripling the practical FOV for mobile receiver alignment over meter-scale links (Han et al., 8 Aug 2024).

5. Key Trade-offs, Limitations, and Future Research

A recurring trade-off in FOV enhancement is between angular extent and spatial, temporal, or spectral resolution. Increasing FOV by spreading finite spatial (optical, neural, or array) resources can degrade fine detail, introduce aberrations, or lower contrast, unless compensated by computational post-processing or adaptive encoding strategies. For data-driven or neural-optical methods, learning must integrate perceptual priors to ensure that frequency bands critical to human vision or hearing are preserved within the expanded FOV.

Emerging research areas include:

• Integration of metasurfaces and neural expanders for further boosting FOV without device complexity (Tseng et al., 2021).
• Adaptive, context-dependent allocation of FOV and resolution in prosthetics based on user intent or attention (Sanchez-Garcia et al., 28 Jan 2025).
• Multi-modal FOV expansion for joint imaging and depth or semantic understanding (e.g., RGBD, phase-contrast, or multi-channel acoustic) (Wei et al., 30 Jul 2025, Allan et al., 14 Feb 2025, Xu et al., 12 Aug 2024).
• Scalable, real-time scene outpainting and beyond-view video synthesis leveraging both geometry- and data-driven priors for navigation, robotics, or safety-critical systems (Shi et al., 2022, Yu et al., 2023).
• Array-geometry-agnostic angular audio enhancement for flexible, wearable, consumer devices (Mittal et al., 16 Sep 2025).

6. Comparative Overview of Methodologies

Domain/Modality	Key Enhancement Approach	Notable FOV Achieved / Improvement
Holography/Display	Neural étendue expander	$64\times$ expansion, >29 dB PSNR (Tseng et al., 2021)
RGBD Imaging	Off-aperture DOE hybrid E2E design	$>5$ dB PSNR @ 45°, RGBD at 28° (Wei et al., 30 Jul 2025)
Microscopy	Conjugate AO (deformable mirror)	FOV scaling to full DM projection (Mertz et al., 2015)
Tomography (X-Ray/CT)	Offset COR + redundancy weighting/U-Net	$\sim 2\times$ FOV, preserved resolution (Allan et al., 14 Feb 2025, Fournié et al., 2019)
Meta-Optics	Polynomial phase meta-lens	80° thermal imaging, $4.2\text{–}4.6$ mrad res (Wirth-Singh et al., 2023)
Speech Enhancement	FoVNet + MCWF in smart-glasses	Adaptive block FOV, 50 MMACS, $>0.2$M params (Xu et al., 12 Aug 2024)
Binauralization	MoE signal-dependent experts	Real-time, continuous FOV, array-agnostic (Mittal et al., 16 Sep 2025)
Video Perception	FlowLens transformer + optical flow	Quantitative mIoU gains $\sim$15% (unseen) (Shi et al., 2022)
Visual Prosthesis	VR-based empirical trade-off quantification	Diminishing return at $<2.3$ phosphenes/deg (Sanchez-Garcia et al., 28 Jan 2025)
SWIPT (Power/Comm.)	Telescope in resonant beam system	FOV expanded to 28° (3$\times$ baseline) (Han et al., 8 Aug 2024)

7. Concluding Remarks

FOV enhancement is a cross-disciplinary objective, achieved through a dynamic interplay of physical optics, computational design, learning-based optimization, and perceptual or task-related constraints. Whether via neural-optical synergy, spatially adaptive encoding, transformer-based inpainting, or hardware-in-the-loop optimization, these methods collectively redefine the attainable boundaries for immersive, high-resolution, and task-specific sensing in sight, sound, and beyond. Advances in this area continue to erode the classical constraints imposed by device physics, sampling theory, and algorithmic bottlenecks, enabling ever broader, richer access to spatial, temporal, and perceptual information.