Research On Mid-wave Infrared Computational Spectral Imaging Technology

Mid-wave infrared spectral imaging technology is able to simultaneously get the radiation spectrum information of the object which is hard to obtain by visible light imaging,and obtain the three-dimensional data cube,which is the combination of infrared spectral technology and infrared imaging technology.Because the radiation or reflection spectra of different substances are different,the characteristic spectra of different substances are also different.Therefore,mid-wave infrared spectral imaging can recognize and identify the target with high confidence.Thus,mid-wave infrared spectral imaging has important application value in planetary exploration,geological exploration,remote sensing,agriculture and medical diagnosis.However,most of the traditional spectral imaging techniques use dispersive elements(such as prisms or diffraction grating),filter elements or Fourier transform infrared spectroscopy technology to expand the spectrum along the band,and then use the detector array to scan along the spatial dimension or spectral dimension respectively to record.This means that a large amount of data acquisition is required,which produces heavy burden to the transmission and processing of spectral images.At the same time,in most imaging systems,the balance between cost and spatial-spectral and temporal resolution is challenging.However,compressed sensing theory breaks this trade-off by effectively reconstructing complete target information by gathering a few of target data,which is far smaller than the data required by Nyquist sampling.The spectral imaging technology based on this theory involves two processes,namely compressed sampling and data reconstruction,that is,spatial encoding and spectral dispersion of the target scene are carried out by encoding aperture and spectral element respectively,and then the original 3D data cube is reconstructed from the collected measurement signal by spectral image reconstruction algorithm.In this thesis,spectral image reconstruction algorithms are deeply studied,and single-pixel spectral imaging and snapshot coded spectral imaging systems are established to carry out experiment research and algorithms validation.Firstly,a spectral image reconstruction algorithm based on multi-band weighted kernel norm minimization is proposed.Restoring original high-dimensional data of target from lowdimensional measurement is an under-qualitative problem.In order to solve this problem,sparse priors are used to lead to poor reconstruction quality.In this thesis,a side information assisted reconstruction model is constructed by taking advantage of spectral image sparsity and low-rank characteristics,and full sampled information is taken as side information to enhance spectral image reconstruction quality.In this thesis,a spectral image reconstruction algorithm based on multi-band weighted kernel norm minimization is proposed.Firstly,the dictionary is learned from infrared image data set,and then the initial spectral image is reconstructed by improving the sparsity of the signal.Finally,the low-rank characteristics are used to minimize the kernel norm of non-locally similar image blocks to achieve optimal reconstruction.Secondly,as traditional optimization algorithms based on manual prior information are not robust and efficient,this thesis presents a deep spatial-spectral reconstruction method based on adaptive dual attention.Based on spatial correlation and spectral similarity,a global network,local multi-branch network and integrated network are designed.Thus,the global and local prior information of the spectrum can be extracted from the shallow and deep images collaboratively and efficiently,thus improving the reconstruction efficiency and fidelity.Among them,the parameter sharing strategy in local multi-branch network can greatly reduce network parameters and avoid repeated feature extraction in different bands.At the same time,based on adaptive weighted feature statistics,we design a sequence adaptive dual attention block including adaptive spatial attention and spectral attention mechanism,and use adaptive to distinguish image features.Then,in view of the poor interpretability of data-driven deep learning algorithm and the limited training data of infrared spectral image,it is difficult to give full play to the advantages of data,this thesis proposes a spectral image reconstruction algorithm based on deep infrared denoising prior.Therefore,combining the traditional optimization method and deep learning method,the infrared image denoising network is designed as the deep denoising prior plug-in to replace the denoising module in the traditional optimization algorithm,so as to have the advantages of both and realize efficient and flexible spectral image reconstruction.Finally,in view of the traditional infrared spectral imaging based on Nyquist sampling theorem,which brings serious computational burden for three-dimensional data acquisition,storage,transmission and processing in both software and hardware,this thesis constructs a mid-wave infrared single-pixel spectral imaging based on the compressive sensing.Infrared spatial light modulator and infrared spectral radiometer are used to obtain spatial modulation information of different bands.Since single-pixel spectral imaging requires multiple sampling by spatial light modulation coding,which leads to a long collection time,a midwave infrared snapshot compressive spectral imaging system is constructed.This spatialspectral coded design provides a high degree of randomness in measurement projection,so as to make better use of spectral image information reconstruction.