Blinkies: Intermittent Signals in Multi-domain Research

• Blinkies are a polysemous term defining intermittent events used as signals in eye-based control, acoustic sensing, nanophotonics, and solar physics.
• They enable diverse applications from real-time human-computer interaction—with methods like event-based vision and SVM classification—to clinical diagnostics and animal behavior analysis.
• Research leverages blinking dynamics through probabilistic models and hardware innovations to reveal hidden structural and state information across physical and biological systems.

“Blinkies” is a polysemous research term whose meanings depend strongly on domain. In human–computer interaction and vision, it denotes systems that detect or exploit eye blinks as intentional commands, temporal signatures, or physiological probes. In distributed acoustics, Blinkies are low-rate sound power sensors that communicate through modulated LED emissions captured by a camera. In nanophotonics and single-emitter science, blinking refers to stochastic photoluminescence intermittency, spectral wandering, or transitions between bright and dark states. In solar physics, the related term “blinkers” denotes transient brightenings linked to magnetic activity in the quiet Sun (Sumathi et al., 2010, Scheibler et al., 2019, Chen et al., 2020, Shokri et al., 2022). The common denominator is the use of intermittent events—physiological, optical, or radiative—as carriers of information about control, state, or microscopic dynamics.

1. Terminological scope and domain-specific meanings

The literature uses the term in at least three technically distinct ways. First, blink-derived sensing treats eye closure as an observable with behavioral or control semantics. Second, Blinkies in acoustics are hardware nodes that measure local sound power and export it optically. Third, blinking in optics and materials science is an intermittency phenomenon of emitters rather than a sensor platform. Solar “blinkers” form a fourth, astrophysical usage centered on transient EUV brightenings (Sumathi et al., 2010, Scheibler et al., 2019, Pathoor et al., 2018, Shokri et al., 2022).

Usage	Defining signal	Representative papers
Blink-based sensing and control	Eyelid motion or blink propagation	(Sumathi et al., 2010, Lenz et al., 2018, &&&10&&&, Guttmann-Flury et al., 23 Jul 2025)
Acoustic Blinkies	Sound power mapped to LED emissions	(Scheibler et al., 2019, Liu et al., 18 Sep 2025)
Photoluminescence blinking	ON/OFF intermittency or spectral fluctuations	(Chen et al., 2020, Gallagher et al., 2024, Davanco et al., 2013)
Solar blinkers	Transition-region transients	(Shokri et al., 2022)

A recurring misconception is that these usages denote a single device class. The literature instead shows a family of concepts organized around “blinking” as an observable. In some cases it is the desired signal; in others it is an artifact to reconstruct, suppress, or reinterpret. This suggests that “blinkies” functions less as a unified taxonomy than as a cross-domain label for intermittent dynamics with high diagnostic value.

2. Eye blinks as control signals and visual primitives

A direct HCI formulation appears in a vision-based game-control system that detects long voluntary eye blinks and interprets blink patterns for communication between human and machine. Facial features, especially the nose tip and eyes, are detected and tracked in realtime, with nose-tip coordinates translated into mouse-pointer motion and left or right eye blinks mapped to left or right mouse clicks. The system uses a Six-Segmented Rectangular filter, integral images, template matching, and an SVM for face verification; motion detection in the eye ROI provides blink recognition; and a user-defined blink length groups prolonged closures into a single command. It runs at 30 frames per second on inexpensive USB cameras and is reported to show similar accuracy in extreme lighting conditions compared to normal lighting, without special lighting or calibration templates (Sumathi et al., 2010).

Event-based vision recasts blink detection as a temporal pattern-recognition problem. A purely event-based face detector uses eye blinks as a unique natural dynamic signature of human faces, exploiting asynchronous event streams of the form $ev=(x,y,t,p)$. The method builds a generic temporal blink model from a wide population of users, correlates local ON/OFF activity within 250 ms windows against this model, and accepts a blink only when two candidates satisfy temporal and spatial consistency constraints across the two eyes. After detection, a probabilistic Gaussian tracker updates face position on every incoming event. The reported system works indoors and outdoors, tracks multiple faces in real time, tolerates scale changes greater than $5\times$, and operates on a single CPU core at about 5.5 W, with blink detection initiating trackers at about 60% recall (Lenz et al., 2018).

These works treat blinking not as incidental ocular motion but as a low-dimensional, high-specificity signal. In frame-based systems, the main problem is separating voluntary blinks from spontaneous ones; in event-based systems, the central advantage is microsecond temporal resolution and low data redundancy. The two approaches converge on the same operational thesis: eyelid motion provides a robust supervisory signal for face-localized inference.

3. Biomedical, behavioral, and neurotechnology uses of blinks

In reading research, blinks are primarily a source of data loss. A reconstruction method based on symbolic sequence dynamics addresses this by recovering horizontal eye position during blinks from the eye movements immediately before closing and after opening. Velocity is symbolized into a four-letter alphabet, transition statistics of short symbolic words are compared to pre- and post-saccade “grammar,” and a critically damped one-dimensional oscillator is used to reconstruct the trajectory during the blink. Entropy analysis indicates a meaningful sequence length typically around $L \approx 60$, and validation with artificial blinks shows superiority over simpler interpolation approaches. The reconstructed data show no significant deviation from the usual behavior observed in readers, which is especially consequential for older adults and reading-impaired subjects with elevated blink-related data loss (0802.2201).

Clinical ophthalmic assessment uses blinks as diagnostic endpoints rather than missing data. Bapp, a mobile application built with Flutter and integrated with Google ML Kit, performs on-device real-time analysis of eyelid movements from live or prerecorded video. It uses eye openness probability, with a blink detected below 0.75, a complete blink below 0.25, and blink termination above 0.98, augmented by Eye Aspect Ratio for subtle and partial events. In validation on 45 videos from real patients, manually annotated by ophthalmology specialists from EPM-UNIFESP, the detailed summary reports 98.4% precision, 96.9% recall, 97.7% F1-score, and 98.4% accuracy, while the abstract reports an overall accuracy of 98.3% (Bonesso et al., 18 Nov 2025). The application also exports processed video and frame-wise CSV data.

Animal-welfare monitoring extends the same logic to nonhuman facial action units. Automated horse blink detection and classification has been evaluated with three methods: a frame-based YOLOv12 detector, an optical-flow magnitude thresholding approach, and a fine-tuned VideoMAE model. On a public EquiFACS-annotated dataset, the study reports a macro-F1 score of 0.898 for three-class blink classification and 0.926 for binary blink detection. Half-blinks are identified as the most difficult class both for automated systems and for humans, with the human baseline reaching a macro-F1 of 0.76 and substantially lower agreement for half-blinks than for “none” or full blinks (Alves et al., 3 Jun 2026).

In EEG preprocessing for BCI, blinking is explicitly revalued from nuisance to signal. The Adaptive Blink-Correction and De-Drifting algorithm uses blink propagation patterns to identify bad EEG channels in the Eye-BCI multimodal dataset. Blink candidates are detected from frontal EEG, validated through expected attenuation from FP1 to CZ, and compared against neighboring channels using Blink-Related Potentials and Longest Common Subsequence Similarity. Across 31 subjects and 63 sessions, classification accuracy averages 93.81% with a 95% confidence interval of [74.81%; 98.76%], exceeding ICA at 79.29% and ASR at 84.05% (Guttmann-Flury et al., 23 Jul 2025). A 5-minute window is described as a good trade-off for sufficient blink statistics and system responsiveness.

Taken together, these studies show that the same physiological event can serve four incompatible but complementary roles: command primitive, missing-data interval, clinical biomarker, and electrophysiological calibration probe. This suggests that blink analysis is governed less by the blink itself than by the inferential context in which it is embedded.

4. Blinkies as distributed acoustic sensors

In acoustics, Blinkies are low-power sound power sensors whose measurements are communicated optically by modulated LEDs and harvested by a camera. A multimodal blind source separation framework combines such Blinkies with a conventional microphone array under a joint probabilistic model. Microphone observations are modeled in the STFT domain, separated sources follow a time-varying spherical Gaussian distribution, and the non-negative blinky power matrix is assumed to have low-rank structure. Alternating updates analogous to independent vector analysis and Itakura–Saito NMF decrease the negative log-likelihood. In numerical experiments, the median separation performance is reported to be up to 8 dB more than that of independent vector analysis, with SDR improvements up to 4 dB in some scenarios and markedly reduced variability; the method can extract a source with one quarter the power of three interferers in cases where conventional IVA fails (Scheibler et al., 2019).

A later acoustic event classification framework uses Blinkies as edge devices that convert audio into LED light emissions recorded by a single camera. The optical channel is severely bandwidth-limited: at 30 fps, each LED conveys only 15 Hz, which the paper characterizes as about 0.2% of the normal audio bandwidth. To address this, a pre-trained autoencoder encoder is used as the embedding function. The latent dimension is fixed by hardware constraints to

$$L = 15\,\mathrm{Hz}\times 4\,(\mathrm{LEDs})\times 5\,\mathrm{s}=300.$$

Noise robustness is introduced by injecting Gaussian noise into the latent representation during autoencoder pre-training, and the encoder is designed for Raspberry Pi 4 deployment, with about 517,508 parameters, about 2 MB model size, and a 3.4 MB inference-time footprint. In simulation on ESC-50 under a stringent 15 Hz bandwidth constraint, the reported macro-F1 scores under the Resample+Distort channel are 0.336 for Sound Power, 0.308 for End-to-End, 0.523 for a vanilla Autoencoder, and 0.536 for the proposed noise-robust Autoencoder (Liu et al., 18 Sep 2025).

The acoustic literature makes clear that Blinkies are not microphones in reduced form. They omit phase and high-rate spectral detail, but gain deployment flexibility, relaxed synchronization requirements, and compatibility with ad hoc spatial layouts. A plausible implication is that their main algorithmic value lies in regularizing ill-posed audio inference with spatially distributed power cues rather than in replacing conventional acoustic sensing.

5. Photoluminescence blinking, intermittency, and control in emitters

In nanophotonics and emitter physics, blinking refers to stochastic intermittency of light emission rather than an instrument. Plasmonic nanojunctions provide a metallic example: intrinsic photoluminescence from single gold nanojunctions shows ubiquitous spectral fluctuations comprising a stable baseline that follows plasmonic resonances and a blinking component that appears and disappears randomly. The blinking manifests as intensity jumps, spectral wandering, and occasional line narrowing on timescales from milliseconds to seconds. It is activated by optical power, especially at 532 nm, shows no clear correlation with temperature from 4 K to 300 K, and leaves Raman and dark-field scattering spectra measurably unchanged. The proposed mechanism is light-induced formation of domain boundaries and quantum-confined metallic emitters, driven by photoexcited carriers and gold adatom–molecule interactions, so that metal luminescence becomes an optical proxy for atomic fluctuations in plasmonic cavities (Chen et al., 2020).

For colloidal semiconductor nanoplatelets, morphology strongly affects blinking mechanism. In CdSe and CdSe/CdZnS nanoplatelets, Auger recombination is ruled out as the source of blinking irrespective of morphology. Core-only and smooth-shell structures show B-type blinking, with intensity fluctuations unaccompanied by major lifetime changes, consistent with hot carrier trapping. Rough-shell structures exhibit additional nonradiative channels associated with shell defects or traps, and polarization-resolved spectroscopy reveals exciton fine-structure splitting on the order of several tens of meV, with values around 25 meV discussed for rough-shell variants (Hu et al., 2018). In colloidal CdSe/CdS core–shell quantum dots, ultrafast mid-infrared pulses at 5.5 $\mu$m and 150 fs can deterministically suppress blinking by switching a charged, low-quantum-yield grey trion state to a bright exciton state; the study reports significant flicker reduction and up to about 30% PL enhancement in thin-shell dots at optimal field strength (Shi et al., 2021). Near-field measurements further show that cycling among internal bright and dark states tunes the balance between tip-induced field enhancement and energy transfer, and that suppressed blinking near metal surfaces is due to fast energy transfer (Shafran et al., 2011).

Blinking also persists in solid-state single-photon sources. In GaN emitters, a photo-induced blinking regime appears when high laser power activates an additional trap state, leading to exponential on/off statistics and persistent changes in photon dynamics; the paper notes that about 5% of emitters show this above threshold (Berhane et al., 2017). In epitaxial InAs quantum dots embedded in circular Bragg grating and microdisk cavities, $g^{(2)}(\tau)$ measured across eleven orders of magnitude reveals blinking from tens of nanoseconds to tens of milliseconds. Multi-level rate-equation fits with several dark states produce radiative quantum yields significantly below unity, including 78%, 86%, and 53% for three representative devices, even when direct time-trace histogramming is inconclusive (Davanco et al., 2013).

Methodologically, blinking analysis has itself become a research object. In fluorescently labeled DNA, a two-state hidden Markov model separates noisy ON/OFF trajectories and finds that ON-duration probability density is well described by an exponential function, while OFF-duration probability density is well described by a log-normal function, both supported by Kolmogorov–Smirnov tests (Furuta et al., 2024). This contrasts with power-law dwell-time statistics in single CsPbBr$_3$ nanocrystals, where average exponents $\alpha_{\mathrm{ON}}\approx 1.09$ and $\alpha_{\mathrm{OFF}}\approx 1.47$ are accompanied by positive correlations between successive ON times and successive OFF times. The reported memory persists for about 80 switching cycles, corresponding to roughly 2 s at a 15 ms bin time, and is insensitive to surface capping ligand and embedding polymer, which the authors interpret as evidence that intrinsic traps dominate intermittency and its memory (Hou et al., 2019).

Perovskite systems further show that blinking mechanisms are not uniform even within one material class. In single CsPbBr$_3$ quantum dots, change point analysis applied to widefield data from 1308 lecithin-capped and 1317 oleic-acid/oleylamine-capped dots shows that lecithin-capped dots are 7.5 times more likely to be non-blinking and spend 2.5 times longer in their most emissive state, with DFT indicating stronger ligand binding for lecithin and therefore lower likelihood of ligand desorption during dilution (Gallagher et al., 2024). Yet single CsPbBr$5\times$0 nanocrystals have also been reported to show blinking memory largely insensitive to ligand and polymer environment, implying an intrinsic-trap contribution (Hou et al., 2019). A more radical deviation appears in individual perovskite quantum dots where low-emitting dark states exhibit better, not worse, single-photon purity: state-resolved measurements show $5\times$1 decreasing from 0.155 to 0.120 while exciton quantum yield decreases by a factor of about 8 and biexciton quantum yield by a factor of about 10. The proposed explanation is a self-trapped-exciton mechanism that selectively suppresses biexciton formation (Olejniczak et al., 11 Feb 2026). This directly contradicts the common expectation that dark states always degrade antibunching.

Finally, blinking can extend beyond nanoscale confinement. Entire MAPbBr$5\times$2 microcrystals with volumes of 0.1–3 $5\times$3 show multilevel photoluminescence blinking under ambient conditions with exceptionally high intra-crystal Pearson correlations, typically above 0.90 and with an ensemble mean of $5\times$4. More than 95% of 125 isolated microcrystals exceeded 0.9, and correlated blinking was observed even across fused polycrystalline grains. The proposed mechanism is efficient long-range carrier migration to a small number of transient nonradiative traps, implying mesoscopic electronic communication over micron distances (Pathoor et al., 2018).

6. Solar blinkers and broader interpretive themes

In solar physics, blinkers are transition-region transients detected automatically in SDO/AIA 304 Å images. Over ten years of solar cycle 24, one study reports 7,483,827 blinkers, alongside 2,082,162 EUV coronal bright points and 1,188,839 X-ray coronal bright points. The blinker birthrate is given as about $5\times$5. About 80% of blinkers occur at supergranular boundaries, 57% are associated with ECBPs, and 34% with XCBPs. Daily blinker counts correlate with quiet-Sun magnetic poles via a power law with exponent $5\times$6, normalized maximum intensity follows a power law with photospheric magnetic flux with slope about 1.31, and the monthly blinker occurrence is very strongly anti-correlated with sunspot number, with Pearson coefficients between $5\times$7 and $5\times$8 (Shokri et al., 2022).

The solar usage is terminologically separate from acoustic Blinkies and photoluminescence blinking, but conceptually it reinforces a common research pattern: intermittent brightening is used to infer hidden structure. In the Sun, the hidden variable is magnetic organization across atmospheric layers; in nanoemitters, it is carrier trapping, charging, self-trapping, or lattice restructuring; in biomedical systems, it is ocular state, eyelid biomechanics, or channel integrity; and in acoustic Blinkies, it is latent sound-field structure communicated through a narrow optical bottleneck (Shokri et al., 2022, Chen et al., 2020, Guttmann-Flury et al., 23 Jul 2025, Liu et al., 18 Sep 2025).

Across these literatures, blinking is alternately treated as control channel, artifact, biomarker, proxy observable, and failure mode. This suggests that the enduring significance of blinkies lies not in any single apparatus, but in the repeated discovery that intermittent events—properly modeled—encode structure that static measurements often miss.