Research On New-style Hyperspectral Imaging Systems And Corresponding Spectral Reconstruction Algorithms

Spectrum is a high-dimensional signal with two spatial dimensions and one spec-tral dimension composed of monochrome map at different wavelength.Hyperspectral images can provide rich clues beyond human eyes and commercial RGB cameras for various computer vision tasks such as object tracking,combustion field analysis,scene segmentation and so on.However,the low light flux,the high complexity and high cost of existing spectral imaging devices limit the wide application of spectral data.Computational spectral imaging is a branch in the field of computational photog-raphy.This paper briefly introduces the development history of the spectral imager,and summarizes some challenging problems which still need to be overcome.Aiming at maximizing the light throughput of spectral imager and reducing the cost of spectrum acquisition system,this paper mainly introduces two new-type spectral acquisition systems:1.By exploring the DoB constraints of blurred images caused by dispersion,and constituting the corresponding graph model,the basic theory is proposed that the rank of the customized edge matrix A of a dispersive blurred image is N-1.It can be inferred that multispectral information can be successfully decoupled from a single dispersive-blurred image and an additional spectrum of an arbitrary point in the scene.2.The side-blocked mask is designed to ensure the spectral imager approaching full light throughput based on DoB constraints of dispersive blurred images,and provide the additional spectrum of the blocked areas.The hybrid acquisition system is used to capture grey sharp images for one arm,and dispersive blurred images for the other arm.Moreover,the corresponding pixel-level reconstruction algorithm is proposed,and the prototype is built to verify the feasibility of this scheme.3.A compact coded spectral acquisition scheme with commercial-level printers as auxiliary is proposed to reduce system complexity and cost.The size of proposed system is just limited by sensors.Moreover,spectra can be reconstructed with high fidelity.This scheme only uses commercial-level color printers as auxiliary to obtain colorful coding modules to be embedded on the sensor imaging surface,greatly reducing the cost and volume of the spectral acquisition system.The scheme is inspired by the randomness characteristics of droplet distribution and irrelevancy of mixed modulation curve.By analyzing the factors affecting the number of uncorrelated observations of the print mask,the optimal parameters to print are selected according to the experimental results.4.The CNN-based method for spectral reconstruction is proposed.The optimiza-tion objective function is designed and the mixed loss function for spectral dimension is creatively proposed in this paper.Convolutional neural network has strong learning ability and can fully mine the prior characteristics of spectral data.By simplifying the network structure,the reconstruction time can be greatly reduced than optimization-based algorithm.The simplicity of the scheme and the high fidelity of spectral details are verified both on the simulated data and the real data.