Self-supervised scene-adaptive spectral imaging via implicit mask evolution
Year: 2026
Type: article
Abstract: The performance of snapshot coded aperture spectral imaging (CASSI) is fundamentally constrained by the reliance on scene-agnostic static masks or computationally heavy data-driven optimizations that require extensive spectral ground truth. We propose a self-supervised, scene-adaptive CASSI framework that refines the physical coded aperture without requiring labeled spectral data. By integrating a dual-branch architecture with an efficient, implicit model-driven reconstruction engine, the system establishes a self-supervised consistency loop that tailors the mask morphology to the spatial-frequency characteristics of the incident scene. Numerical benchmarks and physical prototype validation demonstrate the framework’s ability to adapt to structural boundaries, effectively mitigating high-frequency aliasing artifacts while demonstrating robust generalization across disparate scene distributions without offline training. Operating at an algorithmic latency of ∼55ms per iteration and enabling low-latency closed-loop modulation, this approach demonstrates the feasibility of hardware-integrated physical adaptation for static scenes and establishes the theoretical potential for continuous spatiotemporal evolution in high-throughput sensing pipelines.
Source: Optics Express
DOI: 10.1364/oe.608192
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,270
OpenAlex ID: W7223262364