Hyper-spectral imaging resolves and analyzes optical spectra into object-wise inference results, which is useful for applications with complex spectral information outside of the visual wavelength range such as pollution control, precision agriculture, and biomedical sensing. However, collecting the spectral-spatial hypercubes requires substantial energy, data bandwidth, and hardware form factors, which limit its usage on mobile, high-speed, and battery-based platforms such as robots, unmanned aerial vehicles, and satellites.
To solve the data-bandwidth problem, recent works process the spectral information within sensors, sending only the ultimate machine learning results to the digital domain, encoded as photocurrents in each pixel. Sensors designed for this purpose often require material and architecture innovation. Due to their experimental nature, they often suffer from significant device-to-device variation, fragile structures, and limitations on array scale.
This work aims to address these concerns and prove the viability of spectral kernel machines as a platform by repurposing a commercial camera. The mature imaging hardware allows for megapixel snapshot scene inference, when paired with modulated multicolor illumination control, far beyond the limits of single microdevices.
Project currently funded by: Federal