Abstract
Snapshot Compressive Imaging (SCI) enables the capture of high-dimensional data using standard 2D sensors by multiplexing multiple frames into a single exposure. However, sensor saturation—where total intensity exceeds the sensor's dynamic range—introduces non-linearities that degrade reconstruction. This paper provides the first theoretical framework to address this issue, deriving a new recovery bound and a design rule favoring sparser optical masks. We also introduce SAPnet, a saturation-aware reconstruction algorithm that utilizes inequality constraints for saturated pixels. SAPnet demonstrates significant performance gains, outperforming existing methods by up to 30 dB in high-saturation scenarios.
Introduction
Snapshot Compressive Imaging (SCI) is an efficient technique for capturing high-speed video or hyperspectral data. By using a coded mask to multiplex frames into a single snapshot, SCI bypasses the bandwidth limitations of traditional sensors. A persistent challenge in practical SCI systems is sensor saturation. When the combined intensity of multiplexed frames exceeds the sensor's full-well capacity, the signal is clipped. This clipping creates a non-linear distortion that standard linear reconstruction algorithms cannot handle, often resulting in "blown-out" regions or artifacts.
Theoretical Framework and Design Principles
The authors model sensor clipping as an element-wise nonlinearity and derive a finite-sample recovery bound. This mathematical foundation reveals that reconstruction error is intrinsically linked to mask density and saturation severity.
A key finding is the "Sparser is Better" rule. While traditional SCI theory suggests mask densities should be around 0.5, this research proves that as saturation increases, the optimal mask density should decrease. In high-brightness environments, a sparser mask (e.g., density of 0.2 or 0.3) blocks more light, preventing the sensor from reaching its saturation threshold and preserving more reconstructible information.
SAPnet: Saturation-Aware Reconstruction
To address saturation algorithmically, the authors developed SAPnet (Saturation-Aware PnP Net). Unlike standard Plug-and-Play (PnP) frameworks that attempt to match reconstructed pixels exactly to clipped measurements, SAPnet treats saturated pixels as inequality constraints. For any pixel $y$ at the saturation threshold $T$, the algorithm enforces that the reconstructed value must be $\ge T$. This "consistency" approach is far more effective than simply rejecting saturated data. SAPnet integrates a pre-trained deep denoiser (FastDVDnet) within this iterative framework, ensuring structural fidelity without requiring retraining for different saturation levels.
Experimental Results
The approach was evaluated on six grayscale video SCI benchmarks: Kobe, Runner, Drop, Traffic, Aerial, and Crash ($256 \times 256$ resolution, $B=8$ frames). These datasets represent varying levels of brightness and motion.
SAPnet achieved massive PSNR gains over standard PnP-FastDVDnet, particularly in highly saturated scenarios ($T/B = 0.25$). For instance, the 'Drop' dataset saw an improvement of +30.017 dB. Qualitatively, SAPnet recovers fine structural details in bright areas that appear as flat blobs in standard reconstructions.
Discussion and Conclusion
This work establishes that Bernoulli mask density should generally stay below 0.5, with lower densities being optimal under saturation. The success of SAPnet highlights the power of consistency-based constraints over simple data rejection. While the current implementation relies on a fixed denoiser and Bernoulli masks, future work could extend this theory to optimized "blue noise" masks and end-to-end learned models. This research provides a critical step toward making SCI systems robust for real-world hardware operating under unpredictable lighting conditions.
Summary updated