Fast multispectral imaging via hybrid-encoded LED illumination and a lightweight deep-learning model.

Active LED-based spectral imaging systems provide flexibility and cost-efficiency but suffer from poor temporal resolution due to the need to individually activate LEDs with different light-emitting wavelengths. This work presents a fast spectral imaging scheme leveraging hybrid-encoded LED illumination and a lightweight deep-learning model, LiteSpectralNet (LSNet). It simultaneously activates multiple LEDs in each measurement, significantly enhancing the encoding efficiency compared to traditional sequential methods. LSNet, a one-dimensional convolutional neural network, effectively reconstructs spectra from these compressed measurements. Experimental results demonstrate an 8.2-fold reduction in total exposure time and a 54% reduction in data storage. This method offers 180.5-fold acceleration in reconstruction speed over traditional approaches, with comparable spectral imaging performance, providing an efficient solution for active multispectral imaging.