Compressive spectral imaging with color-coded illumination

Compressive spectral imaging has emerged as a sensing tool where three-dimensional (3D) data can be obtained using two-dimensional (2D) projection measurements. The 3D data cube is reconstructed computationally. However, compressive spectral imagers suffer from system complexity, sophisticated calibration and expensive hardware, which impedes its wide applications. In this work, we propose a simple and low-budget color-coded illumination compressive spectral imaging system (CCISI), that allows for the acquisition of spectral images from a single exposure. CCISI employs color-coded illumination strategy to modulate the spatial and spectral information of the target simultaneously. A grayscale detector then captures the encoded scattered light from the target. To reconstruct the 3D spectral data cube from the single encoded compressive measurement, we propose a regularized end-to-end deep-learning-based architecture, where a set of layers of the deep neural network models the color-coded pattern and the spectrum curves of color-coded illuminations are optimized in the end-to-end network. In contrast, the rest of the network performs the reconstruction task. The effectiveness and accuracy of the proposed CCISI system are verified on both synthetic and real captured images. Extensive experiments show that our method significantly outperforms state-of-the-art (SOTA) methods.