Hyperspectral Imaging Reconstruction Algorithm Based On Broadband Spectral Coding

Hyperspectral imaging technology is a combination of optical imaging and spectral detection of three-dimensional information(two-dimensional spatial information,one-dimensional spectral information)acquisition technology,widely used in Earth remote sensing,medical diagnosis,environmental monitoring,life sciences,color information processing,precision agriculture,and other fields.The narrowband light splitting and Nyquist sampling strategies of traditional hyperspectral imaging technologies result in a large system volume and low light energy utilization.However,hyperspectral imaging technology based on broadband spectral coding utilizes spectral encoders integrated with detectors to achieve spectral information compression,which has the advantages of high altitude spectral resolution,high signalto-noise ratio,small amount of data storage and transmission,simple structure,and compact size,Is a new development direction of hyperspectral imaging technology.In the early days,broadband spectral encoding was implemented using random filters,but due to limitations in the computing power of the processor at that time,it was not applied to spectral imaging.With the development of micro/nano technology and the improvement of machine computing power,a large number of studies have emerged on the use of various micro/nano devices to achieve spectral encoding.Among them,the most representative is the reconfigurable broadband metasurface array,which can achieve single chip hyperspectral imaging.However,due to the independence of the encoder device and reconstruction algorithm design,the encoding efficiency is low.Researchers have proposed a collaborative design method based on neural networks,which achieves encoding efficiency,reconstruction accuracy,and reconstruction speed far exceeding the performance of previous spectral encoding and decoding systems.However,due to the inability to fully optimize the key parameters of encoding and decoding,collaborative design methods can only design locally optimal spectral encoding and decoding systems,and cannot explain the quantitative relationship between filter types and structural parameters,spectral reconstruction accuracy,and resolution,thus unable to further guide the design of optimized spectral encoding and decoding systems.In order to achieve the global optimization design of spectral coding and decoding systems,the object information meaning of broadband spectral coding and decoding is studied from the perspective of sparse coding theory.Through variable parameter training based on sparse coding global optimization networks,the quantitative relationship between filter coding ability and spectral reconstruction accuracy and resolution is explored.In contrast to the prominent problems of compressed sensing spectral reconstruction algorithms such as slow iterative convergence,long reconstruction time,and more sensitive to noise,neural networks combined with software and hardware global optimization methods enable the use of fewer spectral coding units to achieve hyperspectral imaging with higher signal-to-noise ratio and lower delay.Based on this,the following is proposed:(1)Pixel mapping variable resolution spectral imaging.By transforming pixel mapping relationships and using the automatic feature extraction mechanism of a deep learning incomplete self encoder,spectral restoration at different resolutions in different coding dimensions is completed,enabling the resolution of hyperspectral imaging based on broadband spectral coding to have the ability to elastically transform with application scenarios,and achieving high-precision,low-latency spatial spectral resolution dynamic transformation of hyperspectral imaging;(2)Dynamic wide area multi-resolution spectral imaging.Combining the bionic principle of biological central concave imaging,using the distribution of the resolution of a large field of view optical system to adjust the coding pixel combination,achieving progressive changes in spatial spectral resolution with the field of view,enabling the spectral imaging system to observe objects of interest as keenly as the human eye while having the ability to detect background objects,and achieving high dynamic,large field of view multi-resolution spectral imaging.Therefore,hyperspectral imaging technology based on broadband spectral coding can,on the one hand,reconstruct network parameters with different resolutions and have the ability to dynamically transform resolution under different task modes through overloading;On the other hand,it has the ability of high-resolution imaging of local targets and hyperspectral detection of background targets through multi resolution spectral imaging under different fields of view.Through random filter spectral imaging experiments with a wide spectral range(400-700nm),the actual performance of the spectral sparse coding algorithm is verified and RGB images are output,demonstrating good color restoration.Through pixel mapping variable resolution simulation experiments,the output of multi resolution spectral images with 128 x 128 image size,signal to noise ratio of 53 d B,delay of 0.425 s,120 spectrum and 256 x 256 image size,signal to noise ratio of 45 d B,delay of 0.669 s,and 40 spectrum was completed,achieving high signal to noise ratio and low delay variable resolution spectral imaging.