Light-Discontinuity Detection

• Light-discontinuity detection is the identification and quantification of abrupt spatial or temporal changes in light intensity, crucial for photonic devices and imaging systems.
• It employs rigorous mathematical models and advanced device architectures, such as nonlinear waveguide arrays and dynamic vision sensors, to capture minimal light perturbations.
• Algorithmic pipelines integrating ResNet-based feature extraction and attention mechanisms enhance detection accuracy in both physical and digital imaging domains.

Light-discontinuity detection refers to the identification and quantification of abrupt spatial or temporal changes in light intensity within a system or across an imaging modality. This concept underpins a wide spectrum of applications, from discerning physical light jumps in photonic devices, to event-based vision sensors encountering hardware-driven non-idealities, to computer vision systems segmenting illumination boundaries in images. The field encompasses rigorous mathematical modeling of device physics, algorithmic detection pipelines, and empirical evaluation on specifically annotated datasets.

1. Physical and Mathematical Principles

Light-discontinuity detection exploits system responses that are sensitive to minute changes in illumination. In nonlinear optical waveguide arrays, the discrete nonlinear Schrödinger equation (DNLS) provides a precise framework for describing the propagation and interaction of light in coupled waveguides. Specifically, in a laterally-driven finite waveguide array, the amplitude response of the interior sites is governed by simultaneous cubic and linear couplings, leading to multiple steady-state branches and the emergence of a saddle–node bifurcation. Near this bifurcation, the system exhibits ultrasensitivity to tiny input-discontinuities, with the minimum detectable disturbance scaling as $|A_\mathrm{crit} - A|^{1/4}$, where $A$ is the injected amplitude and $A_\mathrm{crit}$ is its bifurcation point (Karjanto et al., 2017).

In semiconductor Dynamic Vision Sensors (DVS), discontinuities arise in the time sequence of event triggers. The physical origin is the finite parasitic capacitance ($C_J$) associated with the photodiode, which introduces a non-instantaneous (non–first-order) response to changes in photon flux. The measurable outcome is a finite $\Delta t_e$, a delay window where no events can be triggered, directly proportional to $1/\mu$ ($\mu$ being the rate of light-intensity change), independent of absolute luminance (Jiang et al., 2024).

2. Device Architectures and Signal Pathways

Photonic Waveguide Arrays

Light-discontinuity detection in such arrays is realized by lateral injection into boundary guides. The system’s core—three central nonlinear waveguides—exhibits a response such that even sub-threshold perturbations can flip the intensity profile, provided the operating point is within the ultrasensitive regime near $A_\mathrm{crit}$ (Karjanto et al., 2017). The propagation of the excitation and its conversion into a measurable transition depend on inter-site coupling ($Q$), on-site detuning ($V$), and Kerr nonlinearity.

Dynamic Vision Sensors (DVS)

In a DVS pixel, the flow of information is as follows:

• Incident intensity $L(t)$ generates a photocurrent $I_\mathrm{pd}(t) \approx \alpha L(t)$.
• $I_\mathrm{pd}(t)$ modifies the photodiode voltage across $C_J$, which, after amplification and buffering, produces a voltage $V_\mathrm{d}(t)$ encoding log-changes in intensity.
• $V_\mathrm{d}(t)$ is assessed by threshold comparators; events (ON/OFF) are fired as $V_\mathrm{d}(t)$ crosses $\pm\Theta$.
• The finite $C_J$ means step changes in $L(t)$ only take effect after a charging interval $\Delta t_e$, resulting in a discontinuous event-interval PDF (Jiang et al., 2024).

3. Algorithmic and Model-Based Detection

Edge Detection in Image Domains

Illumination-edge detection is formalized in the RINDNet architecture, where discontinuity maps are learned as semantic features distinct from reflectance, normal, and depth edges. The pipeline consists of three stages:

1. A shared ResNet-50 backbone for feature extraction.
2. Dedicated decoders (including one for illumination) combining spatial and semantic cues.
3. Per-type decision heads for score maps, refined by an attention mechanism encoding inter-type dependencies.

The illumination-edge output is produced by modulating the preliminary score map $O^i$ with the attention map $A^i$, as $Y^i = \sigma(O^i \odot (1 + A^i))$ (Pu et al., 2021).

Event-Driven Signal Processing

Simulators and filtering algorithms for DVS data must account for the $\Delta t_e$ gap introduced by $C_J$, which can cause event clustering and temporal bias, especially in dim, slowly-varying scenes. Algorithms may compensate by subtracting $\Delta t_e$ from timestamps or explicitly modeling dead-time intervals (Jiang et al., 2024).

4. Experimental Benchmarks and Evaluation

Light-discontinuity detection is evaluated both empirically and via simulation:

• In waveguide detectors, analytical criteria for the optimal injection amplitude are derived by asymptotic and bifurcation analysis of the DNLS steady-state equations, with calibration guided by sensitivity scaling laws (Karjanto et al., 2017).
• For DVS, empirical measurement utilizes event cameras (e.g., DAVIS240) exposed to controlled input ramps (varying $\mu$). Histograms of inter-event intervals robustly exhibit a “gap” $\Delta t_e$ matching analytic predictions. The relationship between $\Delta t_e$ and $\mu$ is linear, as validated by time-resolved statistics (Table 1 below) (Jiang et al., 2024).

$\mu$ (units/s)	$\Delta t_e$ (s)
50	$9.0 \times 10^{-3}$
100	$4.5 \times 10^{-3}$
200	$2.0 \times 10^{-3}$
300	$1.5 \times 10^{-3}$
400	$1.1 \times 10^{-3}$
500	$0.9 \times 10^{-3}$

• In image-based edge detection, the BSDS-RIND dataset provides per-pixel ground-truth for illumination discontinuities. RINDNet’s illumination-edge performance is measured using ODS, OIS, and AP. On the test set, ODS = 0.280, OIS = 0.337, and AP = 0.168—outperforming established generic and specialized baselines (Pu et al., 2021).

5. Annotation, Loss Functions, and Learning Protocols

Robust training hinges on curated ground-truth and specialized loss design:

• BSDS-RIND annotations were produced with tandem annotators and achieve $\mathrm{IoU}=0.92$ for illumination edges. Multi-label regions, though present in 2% of pixels, were accommodated in loss weighting (Pu et al., 2021).
• RINDNet’s balanced edge-loss addresses class imbalance, and a focal-attention loss (with $\alpha_1$, $\beta$, $\gamma_1$, $\alpha_2$, and $\gamma_2$ parameters) supervises both dense edge-labels and attention maps. End-to-end training aggregates these losses.

6. Design and Application Implications

The practical realization of light-discontinuity detection mandates alignment of hardware, simulation, and algorithmic inference:

• In waveguide arrays, the drive amplitude $A$ must be fine-tuned to operate as close as possible to $A_\mathrm{crit}$, while taking into account practical limitations such as losses and higher-order nonlinearities. Device sensitivity can, in theory, be made arbitrarily high as $A \to A_\mathrm{crit}$, but is ultimately bounded by extrinsic factors (Karjanto et al., 2017).
• DVS pixel design must minimize $C_J$ and ensure adequate feedback bandwidth to reduce $\Delta t_e$. Adaptive thresholding and explicit modeling of dead-time intervals enhance both hardware and software performance in low-light, slow-variation scenarios (Jiang et al., 2024).
• For computational detection, disentangling illumination-discontinuity from reflectance and geometry—by architectural separation and attention mechanisms—enables transfer of methods to new domains, such as microscopy or HDR scene analysis, with suitable adaptation of backbone and decoder selections (Pu et al., 2021).

7. Outlook and Extensions

Light-discontinuity detection remains a fundamentally multidisciplinary challenge. In photonic systems, further exploration of bifurcation-enhanced detection may open avenues for ultra-low-threshold optical sensors. In event-based vision, advances will focus on hardware-software co-design to address the residual non–first-order artifacts. In computer vision, architectural innovations in multi-type edge detection and attention-based fusion mechanisms will enable domain-agnostic models for diverse imaging tasks. Extensions to spatiotemporal detection, robust to cross-modal ambiguities, and informed by physically consistent annotation schemes, represent compelling future research directions.