Abstract
The perception of multidimensional information (e.g., spatial, temporal, and spectral domains) plays a vital role in fields like remote sensing that require high optical resolution and precision. The current approach typically relies on hyperspectral imaging, a band-by-band image acquisition modewith subsequent feature learning and fusion through post-processing algorithms. Such an asynchronous workflow introduces substantial data redundancy, transmission latency, and high energy consumption, limiting its practical deployment. Here, we propose a novel spectral-spatial associated vision sensor that enables the synchronous acquisition and feature fusion of spectral and spatial information at the hardware level. Specifically, scalable highly oriented 2D Bi2Te3 thin films with broadband response are employed for the fabrication of highly uniform device arrays, thus achieving simultaneous capture of spectral-spatial information. The arrays perform enhanced synaptic behavior under multi-wavelength stimuli, with a maximum enhanced ratio of more than 20, facilitating feature discriminability and recognition efficiency. By leveraging such a synergistic enhancement characteristic, an increased recognition accuracy of 91.12% is achieved for topography recognition on the Indian Pines dataset. These findings demonstrate that the proposed vision sensor streamlines hardware-level data acquisition while improving processing efficiency, thereby establishing a new paradigm formultidimensional information fusion, particularly in scenarios with massive data streams.
NETL
NSTL
Wiley