Compressive Hyperspectral Computational Imaging Via Spatio-Spectral Coding
Xu Chang,Xu Tingfa,Shi Guokai,Wang Xi,Fan Axin,Zhang Yuhan,Li Jianan
DOI: https://doi.org/10.3788/aos230748
2023-01-01
Acta Optica Sinica
Abstract:Significance Hyperspectral images are made up of tens or even hundreds of contiguous spectral bands for each spatial position of the target scene. Consequently, each pixel in a hyperspectral image contains the complete spectral profile of that specific position. With the superiority of high spectral resolution and image-spectrum merging, hyperspectral imaging has emerged as a powerful tool to obtain multi-dimensional and multi-scale information and has important applications in precision agriculture, mineral identification, water quality monitoring, gas detection, food safety, medical diagnosis, and other fields. Due to the limitations of existing devices, materials, and craftsmanship, traditional hyperspectral imaging technology still suffers from the contradiction between high spatial resolution and high spectral resolution, as well as large data volume and high redundancy in practical applications. The emergence of computational imaging technology has brought new ideas to traditional hyperspectral imaging, and thus a new research field, namely hyperspectral computational imaging has been bred. Hyperspectral computational imaging uses system-level imaging methods to establish the relationship between target scenes and observation results in a more flexible sampling form and jointly optimizes the front-end optical system and back-end processing system, thus fundamentally breaking through the limitations of traditional hyperspectral imaging technology to achieve high-dimensional and high-resolution acquisition of hyperspectral information. Currently, there are numerous hyperspectral computational imaging systems based on various theories and methods, and hyperspectral computational imaging systems based on compressive sensing theory are key branches. The compressive sensing (CS) theory can acquire the signal at much lower than the Shannon-Nyquist sampling rate, solve the underdetermined problem based on the sparse a priori of the signal, and finally recover the original high-dimensional signal with high accuracy. Compressive hyperspectral computational imaging obtains spectral images of the target scene by computing the compressive projections acquired on the detector through reconstruction algorithms, thus significantly improving the system performance while keeping the characteristics of the system components unchanged. For compressive hyperspectral computational imaging, how to design the computational model is a crucial scientific challenge. The coded aperture snapshot spectral imager (CASSI) is a classical model, in which the scene information is projected onto the detector through coded apertures and dispersive elements, and the original data cube is subsequently recovered by the reconstruction algorithm. However, the CASSI system can only obtain a limited number of spectral bands due to the performance of dispersive elements and the detector, which makes it difficult to achieve high spectral resolution detection. Moreover, the reconstruction quality still has much room for improvement because the reconstruction solution problem is too underdetermined. To address the above problems, our team proposes the compressive hyperspectral computational imaging technique via spatio-spectral coding, which achieves super-resolution in both spatial and spectral dimensions and effectively solves the contradiction between high spatial resolution and high spectral resolution. Furthermore, our team has carried out a series of work on improving the quality of system reconstruction and expanding the dimensionality of acquired information, so as to achieve high quality acquisition of high-dimensional and high-resolution hyperspectral data cubes. The research on compressive hyperspectral computational imaging via spatio-spectral coding has laid a solid foundation for the hyperspectral computational imaging technology towards practical applications. Hence, it is important and necessary to summarize the background knowledge of compressive hyperspectral computational imaging and the research work of compressive hyperspectral computational imaging via spatio-spectral coding, which can bring new ideas for researchers to explore the new architecture of compressive hyperspectral computational imaging and promote the development of hyperspectral computational imaging technology. Progress First, the research background and basic concepts of hyperspectral computational imaging are outlined. Then, the current development status of compressive hyperspectral computational imaging systems is summarized, and two classical forms and subsequently improved designs are detailed: one is the coded aperture snapshot spectral imager and the improved systems derived from it, and the other is the hyperspectral computational imaging system based on liquid crystal and the improved systems derived from it. Subsequently, the compressive hyperspectral computational imaging technique via spatio-spectral coding proposed by our team is highlighted, and the system composition, mathematical and theoretical models, and the latest progress are presented. Our team has worked on the coded aperture design and reconstruction algorithm optimization (Fig. 13) to improve the reconstruction quality of the system. The study on the acquisition of polarization dimension information (Fig. 14) is carried out to expand the information acquisition dimension of the proposed system. Finally, the future research trends of compressive hyperspectral computational imaging via spatio-spectral coding are discussed. Conclusions and Prospects Compressive hyperspectral computational imaging technology has a wide range of application prospects. We review compressive hyperspectral computational imaging, including its basic principles, representative systems, and key technologies, so as to provide background knowledge for scholars to engage in related research. Compressive spectral computational imaging via spatio-spectral coding can overcome the contradiction between high spatial resolution and high spectral resolution, and it has made progress in improving the reconstruction quality and expanding the information dimension, which is expected to solve more scientific and engineering challenges. In the future, in-depth research will continue in optimizing the optical design of the system, applying deep learning algorithms for reconstruction, using adaptive compressive sensing theory to improve the imaging quality, and increasing the dimensions of time and depth, so as to promote the practical and industrial development of hyperspectral computational imaging systems.