Real-time Hyperspectral Imager with High Spatial-Spectral Resolution Enabled by Massively Parallel Neural Network
Junren Wen, Haiqi Gao, Weiming Shi, Shuaibo Feng, Lingyun Hao, Yujie Liu, Liang Xu, Yuchuan Shao, Yueguang Zhang, Weidong Shen, Chenying Yang
Abstract
One-shot spectral imaging has been a hot research topic recently, with primary challenges in the efficient fabrication techniques of encoded masks and high-speed, high-accuracy algorithms for real-time imaging. We introduce a real-time hyperspectral imager that leverages multilayer thin film microfilters and the Massively Parallel Network (MP-Net). Each curved microfilter uniquely modulates incident light across the underlying 3 × 3 CMOS pixels, thereby rendering each pixel an efficient spectral encoder. MP-Net, specially designed to address transmittance variability and manufacturing errors such as misalignment and nonuniformities in thin film deposition, greatly increase the robustness to fabrication errors. A spectral resolution of 2.19 nm is achieved for monochromatic spectra. Tested in varied environments on both static and moving objects, the imager demonstrates high-fidelity spatial-spectral data reconstruction capabilities with a maximum imaging frame rate exceeding 30 fps. This hyperspectral imager represents a significant advancement in real-time, high-resolution spectral imaging, offering a versatile solution for applications ranging from remote sensing to consumer electronics.