Research On Reconstruction Algorithm Of Snapshot Spectral Imaging Without Coded Aperture

Spectral imaging acquires spectral information in addition to two-dimensional spatial information,and the obtained three-dimensional data cube can restore the radiance of the scene at different wavelength,which has significant applications in military reconnaissance,agricultural monitoring,biological science,food safety and many other fields.The traditional spectral imaging technology based on scanning,whose temporal resolution is limited by scanning time,is unable to meet the demand of current applications for high dynamic spectral imaging.However,the snapshot spectral imaging technology takes a snapshot of the encoded compressed observation,and the original data cube is reconstructed by the decoupling algorithm,which can effectively improve the temporal resolution.The common snapshot spectral imaging systems encode spectral data cube with coded masks and dispersion optics,therefore they are bulky in practice due to large size and complex system.In order to seek for a more compact snapshot spectral imaging scheme,this paper focuses on the study of spectral imaging system without coded aperture.The main work and innovations include:1.A prototype of a snapshot spectral imaging system without coded aperture is proposed,which consists of a single dispersion component and a camera,so it can be built as a compact and portable system.The paper deduces the image formation model and introduces calibration in the experiment.The proposed scheme can effectively reduce the system size,system complexity and calibration difficulty.2.Based on the proposed spectral imaging system without coded aperture,an optimization reconstruction algorithm is proposed,which includes two steps: spatial alignment and spectral enhancement.Comparing the results after each step,it is proved that the process of spectral enhancement can effectively improve accuracy.In the paper,block operation and equivalent matrix transformation are also introduced to reduce computational complexity,and limitations of the algorithm are discussed.3.Based on the proposed spectral imaging system without coded aperture,two learning-based reconstruction algorithms are proposed,including deep neural networks combined with optimization method and end-to-end deep neural networks.The former inputs the aligned data cube obtained with optimization method into deep neural networks,and the networks output the enhanced data cube.The latter directly inputs the dispersed observation into deep neural networks,and the networks output the reconstructed data cube.The snapshot spectral imaging scheme without coded aperture proposed in the paper can effectively reduce the system size,system complexity and calibration difficulty.The reconstruction algorithms proposed for the system are validated in the paper,and they perform well on online datasets and real data.