RT-Focuser: Remote Refocusing & Real-Time Deblurring

• RT-Focuser is a dual-domain system that combines non-mechanical optical remote refocusing with a real-time deep learning deblurring network.
• The optical module enables fast, 3D volumetric imaging by using precise alignment protocols to maintain diffraction-limited performance.
• The deep learning model employs a modular U-shaped architecture achieving competitive restoration metrics (PSNR ~30.67 dB, SSIM ~0.9005) with ultra-fast inference speeds.

RT-Focuser refers to two distinct advanced systems in image-based science and engineering: (1) a remote-refocusing optical module that enables fast, non-mechanical 3D imaging in microscopy, and (2) a real-time, lightweight deep neural network for image deblurring on edge devices. Both approaches are characterized by rapid focus modulation, but are applied in different domains—physical optics and computational restoration, respectively.

1. Fundamental Principles of Remote-Refocusing RT-Focuser Systems

Remote refocusing in optical microscopy enables swift axial (z-axis) scanning without sample or objective movement. The RT-Focuser module, as detailed in Hong et al. (Hong et al., 2023), comprises three serially-coupled microscope subsystems:

• Microscope 1: Primary objective (O1) and tube lens (TL1).
• Microscope 2: Secondary objective (O2) and tube lens (TL2), arranged in a 4-f relay.
• Microscope 3: Imaging objective (O3) and tube lens (TL3), coupled to the camera.

Optical refocusing is achieved by translating a "remote mirror" conjugated to the focal plane of O2, thereby shifting the imaging plane within the sample. The system maintains conjugate imaging of O1’s and O2’s pupils to avoid aberrations and preserve diffraction-limited performance. Critical theoretical requirements include matching lateral magnification ($M_{\rm lat}$) to the refractive index ratio ($n/n'$), and ensuring axial magnification ($M_{\rm ax}$) closely matches $M_{\rm lat}$ for undistorted volumetric imaging.

2. Construction, Alignment, and Characterization Protocols

Optimal construction hinges on precise selection of focal lengths and spatial relationships:

• $M_1 = f_{\rm TL1}/f'_{\rm O1}$ and $M_2 = f_{\rm TL2}/f'_{\rm O2}$, with $M_{\rm lat} = M_1 M_2 = n/n'$.
• The tube lens separation, $\Delta_{\rm TL1–TL2} = f_{\rm TL1} + f_{\rm TL2}$, is enforced using shear plates.

A standardized eight-step alignment protocol is used:

1. Set O1–TL1 separation.
2. Establish TL1–TL2 relay.
3. Axially align O2–TL2.
4. Place TL3 and camera.
5. Laterally position TL2 via repeated imaging and overlap tuning.
6. Measure lateral/axial magnification gradients.
7. Fine-adjust O2 axial position.
8. Verify $|M_{\rm lat}-M_{\rm ax}|<0.01$ and repeat doublet adjustment as needed.

Tolerance thresholds are highly stringent: TL2 must be positioned laterally within ±0.05 mm; axial errors in O1, O2, or the relay must stay below 0.5–1.5 mm to retain $>$80% of the diffraction-limited imaging volume.

3. Quantitative Performance Metrics and Sensitivity Analysis

The RT-Focuser module’s imaging performance is quantified using mean point spread function FWHM, normalized bead fluorescence signal, distortion, and the 3D diffraction-limited volume $V_{\rm DL}$.

Misalignment Type	ΔFWHM$_{xy}$ (μm)	Norm. Signal	$V_{\rm DL}$ ($10^6$μm$^3$)
Initial (aligned)	0.26	≥1	9.6
O1 axial ±1 mm	0.30	0.87	3.4
O2 axial ±2 mm	0.32	0.85	2.7
TL1–TL2 axial ±2 mm	0.33	0.92	2.4
TL2 lat. ±0.13 mm	0.40	0.70	0.6
TL2 $f_{\rm eff}$±1.3 mm	0.29	1.03	8.4

Axial/lateral misalignments sharply degrade $V_{\rm DL}$ (up to −93% for TL2 lateral misplacement). Acceptable imaging requires FWHM$_{xy} \le 0.26$ μm and normalized signal $S \ge 0.8$.

4. RT-Focuser: Real-Time Lightweight Deep Learning Deblurring Model

RT-Focuser, as proposed by Wu et al. (Wu et al., 26 Dec 2025), is a purpose-built, U-shaped SISO convolutional network optimized for edge-side image deblurring under motion blur conditions. The network architecture incorporates three specialized modules:

• Lightweight Deblurring Block (LD): Performs depthwise 3×3 convolution, pointwise channel expansion/compression, and optional sharpness normalization via Laplacian-based augmentation. Preserves edge information at low computational cost.
• Multi-Level Integrated Aggregation (MLIA): Aggregates multiscale encoder outputs, resizing them to common resolution, fusing by concatenation and channel-attention.
• Cross-source Fusion Block (X-Fuse): Decoder refinement by fusing upsampled decoder features, encoder stage features, and the original blurred input, integrating both high-frequency details and low-frequency context.

Mathematical formulations for these modules are provided, ensuring reproducibility and clarity for implementation.

5. Training Procedure, Loss Function, and Benchmarking

The model is trained on the GoPro dynamic scene deblurring dataset. Key aspects include:

• Input preprocessing: random crop to 256×256, normalization.
• Loss: $$\mathcal{L}_{\rm MSE} = \frac{1}{NCHW}\sum_{n=1}^N \left\| I_{\rm hat}^{(n)} - I_{\rm gt}^{(n)} \right\|_2^2$$.
• Optimizer: AdamW, cosine annealing of learning rate over 3,000 epochs.

Computational complexity is quantified as 5.85M parameters and 15.76 GMACs for a 256×256 RGB input. Empirical restoration results are PSNR=30.67 dB, SSIM=0.9005, with inference speeds of ∼154 FPS (RTX 3090), 146.7 FPS (iPhone 15 A16), and 15–23 FPS (Xeon CPU, various runtimes).

Comparative analysis reveals RT-Focuser achieves competitive quality with an order-of-magnitude improvement in speed and computational cost compared to legacy models (SRN, MIMO-UNet, EDVR, ERDN, etc.).

Model	PSNR	SSIM	Params (M)	GMACs	Time (s)
SRN	29.97	0.9013	8.06	109.07	2.52
EDVR	31.54	0.9260	23.61	33.44	0.21
ERDN	32.48	0.9329	45.68	2138.89	2.89
RT-Focuser	30.67	0.9005	5.85	15.76	0.006

6. Design Recommendations and Integration Guidelines

For optical RT-Focuser modules, select objectives and tube lenses such that $M_{\rm lat}=n/n'$ and build TL2 as a tunable doublet. Use iterative alignment protocols to maintain $|M_{\rm lat}-M_{\rm ax}|<0.01$ and lateral positioning to within 0.05 mm. Image 100 nm fluorescent beads across the desired volume to validate FWHM and signal criteria.

For the deep learning RT-Focuser, deployment is suited for real-time embedded contexts requiring high throughput and single-frame inference. The modularity allows extension with additional loss terms or hybrid attention designs where greater restoration quality is desired and latency tolerance exists.

7. Domain Significance and Future Directions

The dual RT-Focuser paradigm—physical remote refocusing in microscopy and algorithmic real-time deblurring—addresses a persistent bottleneck in both imaging science and real-world computer vision: fast, resolution-preserving acquisition or restoration. Optical RT-Focuser modules facilitate volumetric imaging for advanced biological studies, while the computational RT-Focuser enables reliable edge computing in time-sensitive applications such as robotics, UAVs, and mobile medical diagnostics.

A plausible implication is that continual evolution in both physical and computational RT-Focuser systems will further converge multimodal rapid-focus techniques, expanding 3D imaging and real-time restoration capabilities across disciplines (Hong et al., 2023, Wu et al., 26 Dec 2025).