| Spectral imaging technology is a technique that integrates two-dimensional image information with one-dimensional spectral information to obtain three-dimensional data cubes.The acquired multispectral data can characterize the intensity of electromagnetic radiation on the object’s surface and can be applied in various fields,including atmospheric turbulence inversion,meteorological forecasting,disaster mitigation and prevention,resource exploration,military target remote sensing reconnaissance,dynamic perception of battlefield situations,and real-time monitoring of sensitive areas.Traditional multispectral imaging systems typically employ push-broom,whisk-broom,filter wheel,or mechanical interference methods to collect spectral dimension data by sacrificing either temporal or spatial resolution.This approach,which compromises one dimension of information to enhance the resolution of another,cannot achieve a balanced consideration of spatial,temporal,and spectral resolutions.Moreover,the complex calibration and multispectral reconstruction algorithms cannot realize rapid and efficient real-time reconstruction.Additionally,traditional spectral imaging equipment is generally bulky,heavy,powerconsuming,and most devices contain mechanical moving parts,leading to generally low stability and reliability.These factors limit the speed of acquiring single-frame spectral images and reduce the light utilization rate.To address these issues,we propose a complementary full-light-throughput multispectral video imaging system based on a Digital Micro-mirror Device(DMD),which uses the DMD to perform both light beam splitting and spatial encoding operations,achieving full-light-throughput imaging.The system models and corrects coupling errors introduced by the system and uses an improved bilateral filtering algorithm to accomplish spectral reconstruction.Furthermore,to meet the requirements of real-time imaging,a progressive lightweight multispectral reconstruction network based on color image priors is designed and deployed on the computational spectral imaging system to acquire multispectral data of real scenes.The realized multispectral video imaging system,with no moving parts,miniaturization,and lightweight design,meets the requirements of computational spectral imaging.Addressing the issues of low optical throughput,challenging coupling error modeling,and reconstruction algorithms that fail to meet real-time requirements in existing computational spectral imaging systems,this dissertation conducts research on a computational spectral imaging system and method based on complementary full-light-throughput.The proposed system can achieve full light utilization and is capable of self-calibration of the system,as well as real-time reconstruction of multispectral videos.The main research work is as follows:1.In response to the issue of low light utilization in existing computational spectral imaging systems,this dissertation,grounded in the principles of spectral imaging and computational imaging technology,develops an overall design for a complementary full-light-throughput multispectral video imaging system.It proposes the use of a Digital Micro-mirror Device(DMD)to simultaneously perform spectral splitting and spatial encoding operations,achieving full light utilization.The integration of optical and mechanical components is carried out to ensure the stability of the optical path and the accuracy of the mechanical system.In addition,criteria for selecting key system components,such as the DMD and dispersive elements,are established.A non-aliasing sampling encoding template is designed for scene encoding,and a combination of the DMD with a dual-channel camera is used to achieve hard-triggered driving and control.Finally,a dual-channel registration technology based on the same reference plane is proposed to address the registration of the two cameras.A high-resolution,low-power,miniaturized,and lightweight computational spectral video imaging system is developed using existing conventional optical detectors,capable of capturing high-resolution data in time,space,and spectrum.2.To address the issue of coupling errors in complementary full-light-throughput multispectral systems,this dissertation proposes a coupling error modeling method and a self-calibration scheme.Inherent aberrations and chromatic errors in computational spectral imaging systems inevitably cause the actual extracted observations to deviate significantly from the theoretical design model,resulting in substantial differences in spectral accuracy and image quality compared to theoretical simulation results.This work investigates many theoretical and engineering issues closely related to the design of complementary full-lightthroughput spectral imaging systems,considering the noise-level errors introduced by the actual system’s manufacturing,assembly,and optical design processes,and specifically models the errors in the computational imaging system.For imaging errors caused by diffraction effects of the DMD,a dual telecentric optical path structure equipped with a cemented mirror is proposed for parallax compensation;for optical path differences caused by the rotation of DMD mirrors around their axes,tilting the image plane is proposed to compensate for the optical path difference in off-axis fields of view.Furthermore,a co-axial design of prior information and encoding information is proposed to accurately model the point spread function of each pixel,complete the radiometric and spectral calibration of the computational spectral camera system,reduce problems such as aberrations and chromatic errors caused by aliasing dispersion,and achieve geometric self-calibration.In terms of reconstruction algorithms,a bilateral filtering reconstruction algorithm based on compensated prior information points is proposed to achieve real-time multispectral video acquisition.Finally,a comparison with existing computational spectral imaging systems is made to demonstrate the superiority of the proposed system.3.To address the issue that traditional fusion algorithms and existing deep learning-based methods cannot meet the real-time requirements for hyperspectral image fusion,this dissertation proposes a progressive lightweight multispectral reconstruction network based on color image priors.The network is designed as a lightweight feature extraction and fusion tool to meet the reconstruction requirements under video frame rate conditions.Firstly,a residual consistency discriminator based on the fusion problem is introduced to evaluate the performance of the residual module,thereby minimizing the number of parameters and runtime.Based on this criterion,a band-by-band progressive fusion strategy is adopted to gradually improve the quality of features extracted by the network.Finally,comparisons with other methods on simulated datasets and real-world hyperspectral datasets verify the significant advantages of the proposed method in terms of reconstruction performance and speed.Additionally,the method is tested on real scenes captured by a spectral imaging system,and compared with algorithms using trilateral filtering.The results demonstrate that the proposed method achieves a high level of reconstruction quality and speed. |
