image: Figure | Metasurface physical encoding and high-dimensional optical sensing architecture. (a) Schematic of the hyperspectro-polarimetric meta-sensor integrating a metasurface with an InGaAs focal plane array (FPA). (b) Schematic of the metasurface-enabled joint polarization–spectral encoding mechanism. (c) Schematic of the encoded image. (d,e) Schematic illustrations of multidimensional optical information sensing and hyperspectro-polarimetric data-cube imaging. (f) Performance comparison between this work and previously reported HSP/HSPI devices in terms of polarization reconstruction error, spectral resolution capability, and spatial resolution.
Credit: Jing Zhou et al.
A hyperspectro-polarimetric imaging framework based on metasurface physical encoding and AI-driven decoding is proposed. By jointly utilizing polarization multiplexing and structural dispersion, the system enables single-shot reconstruction of hyperspectral full-Stokes information. The device achieves 167 spectral channels across 1150-1650 nm with 3 nm spectral resolution, 0.14° polarization reconstruction error, and 55 fps real-time imaging. Embedded deployment validation further demonstrates the potential for compact intelligent high-dimensional optical sensing systems.
In a new paper published in 《Light: Science & Applications》, a team of scientists led by Professor Wei Lu, Professor Xiaoshuang Chen, and Professor Jing Zhou from the Shanghai Institute of Technical Physics, Chinese Academy of Sciences, developed a novel high-dimensional optical sensing framework based on “metasurface physical encoding + data-driven decoding.” By jointly utilizing the polarization multiplexing capability and structural dispersion characteristics of the metasurface as high-dimensional information encoding degrees of freedom, and further combining deep neural networks to learn the complex physical mapping relationships of the real optical system, the team achieved single-shot high-accuracy hyperspectral full-Stokes real-time imaging. The proposed system simultaneously reconstructs 167 spectral channels and full-Stokes polarization information across the 1150–1650 nm broadband range, achieving a minimum distinguishable wavelength interval of 3 nm, an average polarization reconstruction error as low as 0.14°, and a real-time imaging speed of 55 fps. Moreover, embedded-platform deployment validation further demonstrates the potential for compact intelligent high-dimensional optical sensing systems.
The proposed hyperspectro-polarimetric imaging system is centered around a metasurface-integrated infrared meta-sensor empowered by deep neural networks. Unlike conventional approaches that attempt to suppress wavelength-dependent dispersion effects, the proposed framework actively utilizes the polarization–dispersion coupling of the metasurface as a multidimensional encoding mechanism. Different polarization states are spatially multiplexed into different diffraction channels, while different wavelengths are continuously dispersed along the radial direction through structural dispersion. The encoded high-dimensional optical information is then directly reconstructed from a single grayscale image using the HSPI-net deep neural network.
These scientists summarize the operational principle of their system:
“We utilize the polarization modulation capability and structural dispersion characteristics of the metasurface simultaneously as high-dimensional optical encoding degrees of freedom. By combining metasurface physical encoding with AI-driven decoding, the system can directly reconstruct hyperspectral full-Stokes information from a single captured frame, enabling real-time high-dimensional optical sensing.”
The scientists further highlighted the significance of integrating metasurface optics with artificial intelligence:
“Instead of treating wavelength-dependent dispersion as a limitation, our work transforms the intrinsic polarization–dispersion coupling of metasurfaces into a usable information encoding mechanism. This physical encoding and data-driven decoding paradigm provides a new route toward compact, intelligent, and real-time high-dimensional optical sensing systems.”
“The presented technique could open new avenues for future intelligent optical sensing systems in autonomous driving, infrared target recognition, remote sensing, biomedical imaging, machine vision, and high-dimensional optical communications,” the scientists forecast.
Journal
Light: Science & Applications
Article Title
High-accuracy hyperspectro-polarimetric real-time imaging via a deep learning empowered infrared meta-sensor