Artificial Microsaccade Compensation

• Artificial microsaccade compensation is a method that mimics involuntary eye movements to prevent sensory fading and enhance signal robustness.
• It integrates hardware and software approaches, such as stochastic stimulation and geometric warping, to mitigate adaptation, motion blur, and jitter.
• Studies report improved accuracy and reduced artifacts, with up to 91.5% microsaccade artifact suppression and enhanced performance in diverse sensor systems.

Artificial microsaccade compensation is a class of methods inspired by the biological phenomenon of microsaccades—small, involuntary eye movements in animals with foveated vision—that stabilize, enhance, or prolong sensory perception in the presence of limits such as receptor adaptation, motion-induced signal loss, or external vibration. These techniques have been implemented across domains including electrotactile sensory substitution, event-based visual sensing, high-resolution ophthalmic imaging, and mobile robotics, and hinge on either actively introducing “microsaccade-like” stimulation or estimating and compensating for corresponding motions to suppress artifacts and preserve information.

1. Biological Basis and Conceptual Motivation

Microsaccades are small, rapid, involuntary eye movements occurring at 1–2 Hz in humans and many foveated animals. They serve to prevent perceptual fading that arises when sensory receptors adapt to static stimuli, thus ensuring the persistence of visual scene perception. In artificial systems, analogous phenomena include adaptation-induced fading in tactile or retinal prostheses, insensitivity of event cameras to edges aligned with motion, and motion blur or distortion in unstable video acquisition platforms. Artificial microsaccade compensation mechanisms mimic or model these biological roles to enhance the stability and persistence of artificial sensory signals (Burner et al., 3 Dec 2025, He et al., 28 May 2024, Wei et al., 2020, Chekhchoukh et al., 2013).

2. Mechanisms in Sensory Substitution Devices

Chekhchoukh et al. (2013) implemented microsaccade-inspired stochastic signal jumps in a 12×12 tongue electrotactile array for vision substitution (Chekhchoukh et al., 2013). The scheme operates as follows:

• Encoding: Each “taxel” (i,j) receives a voltage (1–10 V) per the 2D visual pattern. Saccades are periodically (0.5 Hz, exponentially distributed intervals) simulated by shifting each taxel's input randomly to one of its eight neighboring positions.
• Effect on Receptor Adaptation: A first-order adaptation model is posited,

$$\tau\,\frac{dR(i,j,t)}{dt} = -\left(R(i,j,t)-R_{\rm base}\right) + \alpha S(i,j,t)$$

with partial resets at each saccade:

$R(i,j,t_k^+) = (1-\gamma) R(i,j,t_k^-)$

This emulates a high-pass filtering effect, counteracting receptor saturation by spatially distributing stimulation across fresh mechanoreceptors.

• Experimental Validation: A 30% reduction in angular error was observed in orientation-judgment tasks when saccades were included under long-duration (10 s) tactile presentations, with no significant detrimental effect on short (≤5 s) stimuli.
• Generalization: Analogous stochastic jitter (rate and amplitude matched to relevant physiology) can be incorporated in retinal prostheses and any sensory array subject to fading, provided jump parameters respect the array spatial scale.

3. Hardware and Software Approaches in Event-Based Vision

The AMI-EV system introduces physical microsaccade-like motion to an event-based (neuromorphic) camera by mounting a precision wedge prism in front of the sensor and rotating it at high speed (720 rpm ≃ 12 Hz) (He et al., 28 May 2024). The technical details include:

• Hardware: The rotating prism redirects incoming light, causing all scene edges to traverse various orientations relative to the pixel grid, thereby generating events even for edges parallel to camera motion (which would conventionally be missed).
• Geometric-Optics Warping: The induced optical flow for each pixel is analytically modeled through Snell’s law and implemented in software. Each event is time-tagged with the prism's orientation and “de-rotated” (motion-compensated) to a common reference before being accumulated in downstream processing.
• Performance: The AMI-EV achieves 2× improvement in texture persistence and feature stability, higher event-stream uniformity (KDE variance reduced from 0.425 to 0.196), 40–60% more reliable corner tracking, and 10–20% boosts in high-level vision metrics (IoU, PDJ) relative to standard event and frame cameras under challenging, dynamic conditions.
• Limitations: Mechanical complexity, power draw, and compensation discretization (≈1.5–2 px) introduce constraints for precision tasks and miniaturized or high-G applications.

4. Real-Time Microsaccade Correction in Ophthalmic Imaging

Wide-field OCT angiography (OCTA) is sensitive to minute eye movements (microsaccades and blinks), producing motion artifacts that degrade capillary detection. An efficient, hardware-free, self-navigated correction algorithm was proposed (Wei et al., 2020):

• Instantaneous Motion Index (IMSI):

$$\mathrm{IMSI}(t_i) = \frac{\sigma[D_i(x,y)]}{\mu[D_i(x,y)]}$$

where $D_i(x,y)$ are en face image values for batch $i$. Suprathreshold IMSI ($>$0.25) signals a microsaccade or blink artifact.

• Correction Protocol: When a batch exceeds the motion threshold, the system logs the index and upon return to sub-threshold IMSI, automatically rescans that locus. The process is powered by a GPU-accelerated pipeline that completes per-batch processing and rescan logic within 30 ms, supporting real-time correction.
• Results: Suppression of 100% of blinks and 91.5% of microsaccade artifacts across 14 eyes (7 subjects), with qualitative preservation of fine capillary detail. The method operates with <5% increase in light exposure over uncorrected acquisition.
• Limitations: It is sensitive only to rapid, not slow, drift; ignores direct measurement of displacement; depends on patient fixation.

5. Computational Approaches for Camera Shake and Robotics

"Artificial Microsaccade Compensation: Stable Vision for an Ornithopter" presents a real-time, rotation-based image stabilization method for high-frequency-shake platforms (Burner et al., 3 Dec 2025):

• Mathematical Model: The method models orientation changes as elements of SO(3), with Gauss–Newton optimization minimizing

$J(R) = \sum_x \|I_{t+1}(x) - I_t(R x)\|^2$

over small 3D rotations.

• Algorithmic Pipeline: An inverse-compositional Lucas–Kanade method is used to estimate inter-frame rotation, followed by temporal low-pass filtering of viewpoints and recursive warping/averaging for producing stabilized output.
• Experimental Benchmarks: On a flapping-wing robot (12–20 Hz vibrations), “saccade” mode reduces RMS normal-flow errors by ≈90% and frame intensity change by ≈90% (relative to raw video), outperforming commercial stabilizers such as Adobe Premiere Pro Warp by at least 50% in objective metrics, with zero scene distortion and real-time CPU operation.
• Key Technical Features: No dependence on unreliable IMU data; robust to rolling-shutter and transmission dropouts; requires only prior intrinsics.
• Limitations: Sharpness loss (~10%) due to averaging; does not compensate for translation, which can be addressed with depth-aware enhancements.

6. Comparative Summary of Implementations and Evaluation

Domain	Core Mechanism	Key Benefit
Sensory Substitution	Stochastic spatial pattern jumps	30% error reduction after 10 s static
Event Cameras	Rotating prism, geometric warping	2× texture persistence, ↑ feature tracking
OCT Angiography	Real-time motion metric/rescan	91.5% microsaccade suppression
Robotics Video	SO(3) image rotation estimation	90% motion artifact suppression

Across these domains, the unifying principle is that artificial microsaccade compensation either injects microsaccade-like input (as in prostheses and event cameras) or continuously estimates and corrects for high-frequency “microsaccade” motion (as in OCT and robotics), yielding robust stabilization, increased perceptual persistence, and higher information throughput.

7. Extensions, Limitations, and Future Directions

A plausible implication is that artificial microsaccade compensation can be generalized to any sensor modality or robotic system in which sustained static stimulation, motion-aligned insensitivity, or high-frequency vibration leads to perceptual or inferential loss. Noted limitations across implementations include power and complexity in hardware solutions, information loss with frame averaging or resampling, and insufficient compensation for drift or translation. Potential extensions involve:

• Electro-optic/micro-mirror devices for rapid, low-power actuation (event vision) (He et al., 28 May 2024).
• Edge-preserving and deblurring fusion strategies to counteract sharpness loss (robotics video) (Burner et al., 3 Dec 2025).
• Depth-aware stabilization and automated calibration (robotics, computer vision).
• Cross-domain adaptation of variance-normalized real-time motion metrics (IMSI variants) (Wei et al., 2020).

These directions suggest a broadening technological and theoretical footprint for artificial microsaccade compensation, at the intersection of neuromorphic engineering, sensory augmentation, biomedical imaging, and robotic perception.