Memory-Guided Exposure Control

• Memory-Guided Exposure Control is an approach that uses historical sensor data to dynamically adjust exposure settings and capture optimal images in variable lighting conditions.
• It utilizes proportional feedback and minimal computational resources to prevent sensor saturation while accurately detecting transient events.
• Empirical evaluations show high dynamic range performance, robust artifact minimization, and effective operation in applications such as astronomical monitoring and HDR SLAM.

Memory-guided exposure control refers to adaptive algorithms that adjust the exposure time and sensor gain of imaging devices in real time by making use of stored measurements and internal state from previous exposures. This approach is essential for imaging scenarios involving rapidly changing scene brightness and high dynamic range, such as astronomical transient monitoring and online 3D mapping, where the system must avoid sensor saturation while capturing subtle photometric features. By leveraging memory of past exposures and configurable setpoints, memory-guided exposure control systems maximize information content and minimize artifacts arising from abrupt illumination changes. Recent developments have advanced these techniques for both targeted scientific imaging (e.g., flare monitoring in stellar observations) and robotics applications (e.g., map-aware exposure control in SLAM), resulting in robust, autonomous exposure management under computational and memory constraints (Alexandrov et al., 2018, Ramiaramanantsoa et al., 2021).

1. Principles of Memory-Guided Exposure Control

Memory-guided exposure control is predicated on maintaining and utilizing a minimal internal state that evolves over sequential exposures. Core state variables typically include the preceding exposure settings (exposure time τ, gain G) and the measured peak or representative signal in a defined image region (e.g., the PSF peak value, $\widetilde S_i$). This historical state is compared against a configurable setpoint ($S_{\rm SP}$), allowing the controller to issue proportional or otherwise adaptive adjustments on the next cycle.

The fundamental mathematical relationship for proportional adjustment is:

$$\tau_{\rm opt} = \frac{S_{\rm SP}}{\widetilde S_{\max,i}} \tau_i$$

where $\tau_i$ is the current exposure, $\widetilde S_{\max,i}$ is the measured peak after corrections, and $S_{\rm SP}$ is the target setpoint. Gain adjustments can be incorporated if the calculated optimal exposure falls outside the hardware-allowed range (Ramiaramanantsoa et al., 2021).

2. System Architectures and Memory Utilization

Implementations of memory-guided exposure control typically require minimal computational and memory resources, allowing integration even on resource-constrained platforms such as CubeSat payload processors. For instance, the SPARCS mission employs a BeagleBone Black platform with 512 MB RAM, running its exposure algorithm in Python under KubOS Linux. The memory footprint encompasses persistent logging of the latest parameters ($\tau_i, G_i, \widetilde S_i$) in a “restoring” file and temporary buffers for calibration data (bias, dark, flat frames) as well as bad-pixel and cosmic-ray masks (Ramiaramanantsoa et al., 2021).

A summary of memory resources in SPARCS is as follows:

Resource Type	Contents	Purpose
Persistent state	Current $\tau_i$, $G_i$, $\widetilde S_i$	Recovers last state after reset
Calibration frames	Bias/dark/flat per (gain, temperature)	Real-time calibration corrections
Quality masks	Bad-pixel/cosmic ray templates	Robustness to detector artifacts
Temporary subrasters	$16 \times 16$ px regions around target	Efficient local statistics computation

On each loop iteration, the algorithm "remembers" the previous exposure, using it to guide the update of exposure time and gain, constituting a feedback loop grounded in immediate experience, rather than inferences or predictions extending over multiple timescales (Ramiaramanantsoa et al., 2021).

3. Control Algorithms and Mathematical Formulation

Memory-guided exposure controllers primarily operate by acquiring an image with the current settings, processing it to extract a representative statistic, and updating parameters for the next frame. The SPARCS implementation utilizes the bias-subtracted peak pixel in the PSF as its primary control variable:

$$\widetilde S_{\max,i} = \frac{\tau_i}{\kappa_i}\left[\mathcal F_i (\Lambda_{e,i} + \beta_{e,i}) + \delta_{e,i}\right]$$

where $\kappa_i$ is the system gain, $\mathcal F_i$ is filter throughput, $\Lambda_{e,i}$ is target photon rate, $\beta_{e,i}$ is sky background, and $\delta_{e,i}$ is detector noise (Ramiaramanantsoa et al., 2021).

Proportional updates proceed as follows:

• If any pixel achieves ADC saturation, immediately assign $\tau_{\rm next} = \tau_{\min}$ and $G_{\rm next} = G_{\max}$.
• Otherwise, calculate $\tau_{\rm opt}$ as above, and:
  • If $\tau_{\min} \leq \tau_{\rm opt} \leq \tau_{\max}$, set $\tau_{\rm next} = \tau_{\rm opt}$, $G_{\rm next} = G_i$.
  • If $\tau_{\rm opt} < \tau_{\min}$, assign $\tau_{\rm next} = \tau_{\min}$, $$G_{\rm next} = \min(G_{\max}, (\tau_{\min}/\tau_{\rm opt}) G_i)$$.
  • If $\tau_{\rm opt} > \tau_{\max}$, set $\tau_{\rm next} = \tau_{\max}$, $G_{\rm next} = G_i$ (Ramiaramanantsoa et al., 2021).

Pseudo-code specifies initialization with maximum exposure and minimum gain, iterative image acquisition, processing, proportional control logic, and persistent logging on each cycle.

4. Cross-Channel and Map-Aware Coordination

A distinguishing feature in certain memory-guided exposure controllers is the capacity for cross-channel coordination. In the SPARCS system, independent NUV and FUV detectors each run the control logic, but detection of a flare onset in the NUV channel ("if $\tau_i = \tau_{\max}$ and $\tau_{\rm next} < \tau_{\max}$") triggers the FUV channel to abort its ongoing exposure, immediately synchronizing its next exposure parameters to those computed by the NUV channel. This mechanism ensures both rapid adaptation to transient events and protection against data loss from saturation in the complementary band (Ramiaramanantsoa et al., 2021).

In map-aware exposure control for HDR SLAM, a similar philosophy applies: exposure decisions are made in view of the global map state and predicted camera motion, with the objective of maximizing observation information gain. Although fine-grained algorithms (e.g., per-voxel statistics, RGB-D noise models, or explicit entropy objectives) are not detailed in available summaries, the approach fundamentally leverages a temporal memory of the map state and recent sensor readings (Alexandrov et al., 2018). This suggests that memory-guided exposure control can be generalized to settings where spatial as well as temporal statistics inform adaptive acquisition strategies.

5. Performance Metrics and Empirical Findings

Empirical evaluation of memory-guided exposure control underscores its effectiveness for high dynamic range applications. SPARCS simulation campaigns using full-frame synthetic images and light curves in both NUV and FUV demonstrated:

• Saturation avoidance exceeding 99%, including during flare-rise intervals.
• Accurate high-cadence light-curve resolution (sub-10 s exposure at flare peak).
• Signal-to-noise ratio (S/N) of 20-50 per maximum quiescent exposure and above 100 during short exposures at flare peak.
• Responsiveness characterized by an update lag of one exposure (9.7 s frame overhead), with setpoint convergence within 10% in a single cycle.
• No occurrences of false aborts in faint-star quiescent states; quiescent integration remains at $\tau_{\max}$, with no false-positive flare triggers.

These metrics confirm robust separation of quiescent and flaring regimes, fine temporal sampling at transient peaks, and resilience to unexpected illumination events (Ramiaramanantsoa et al., 2021). In HDR SLAM, memory-aware exposure selection contributes to improved texture quality and suppression of color inconsistency/artifacts in dense 3D reconstructions (Alexandrov et al., 2018).

6. Adaptability, Extensions, and Future Directions

The architecture and control logic of memory-guided exposure control are broadly adaptable. The SPARCS controller relies solely on PSF peak statistics, enabling computational efficiency and independence from full-photometric or scene-interpretive pipelines. Its proportional-only feedback design is robust for steep, impulsive transients without requiring integral or derivative control layers. Parameters—including setpoint, exposure/min/max, gain/min/max, and control regions—are externally configurable.

Potential extensions include:

• Predictive templates for anticipated events (e.g., flare modeling for feed-forward control).
• Machine learning applied to sliding windows (ring buffers) of parameter histories $(\tau_i, G_i, \widetilde S_i)$ for drift or anomaly identification.
• Generalization to other transient phenomena—gamma-ray bursts, kilonovae, or multi-channel cross-triggering (e.g., Hα vs. white-light, X-ray vs. UV).

Adoption is feasible in ground-based or other space-based imaging systems provided with programmable gain/exposure and simple statistic-driven feedback, with or without coordinated multi-band operations (Ramiaramanantsoa et al., 2021).