Noise-Adaptive and Task-Specific Coherent Illuminations with a Programmable-Metasurface Imager
Résumé
We study the impact of noise on the end-to-end optimization of multi-shot single-detector meta-imagers with respect to a specific information-extraction task. Latency constraints and noise can severely limit the total amount of information that can be extracted from the scene. Therefore, the ability to predominantly extract task-relevant information is a strong advantage. Considering dynamic metasurface antenna (DMA) hardware in a prototypical object-recognition task, we observe remarkable performance improvements over conventional meta-imaging with random DMA configurations. Moreover, we analyze the learned sequence of scene illuminations and discover intuitively understandable trends in its dependence on latency constraints and the noise level.