Analysis Of Coded Spectroscopy Measurement Performance And Research On Fast Coded Spectroscopy Imaging Syste

Through the spectral dimension information in the hyperspectral image,the physical and chemical characteristics of the substance can be better detected.Hadamard coding measurement has the advantages of high signal-to-noise ratio(SNR)and high luminous flux.As the quality requirements for hyperspectral imaging images increase,the combination of convolutional neural networks and spectral imaging technology can improve the quality of hyperspectral images while effectively increasing the efficiency of image reconstruction.The main research contents of this topic are as follows:(1)For the change of the SNR with the Hadamard coding measurement of the spectral signal,the traditional analysis method is generally verified through experiments.To solve this problem,a method which denoising analysis of Hadamard coding based on matrix eigenvalue distribution is proposed.According to the definition of the signal-to-noise ratio(SNR)and the matrix generalized Rayleigh quotient,we establish the equivalent relationship between the Hadamard coding transform spectrometer’s SNR and the measurement matrix generalized Rayleigh quotient.Based on noise source during spectral detection,we analyze the changes of SNR with different types of spectrometers when detector noise or photon noise is dominant respectively.The theory shows that when detector noise is dominant the SNR comparison is SNRH>SNRS>SNRSlit.And the difference between SNRH and SNRS is about 3.01dB.The difference between SNRS and SNRSlit is about 10log10(((?))/2)dB,where n is the matrix order.When photon noise is dominant the SNR comparison is SNRH≈SNRSlit>SNRs.And the difference between them is about 1.51dB.The accuracy of the theory is verified by simulation experiments and actual system measurement experiments.(2)Aiming at the contradiction between image resolution and imaging speed in Hadamard coding spectral imaging,a fast imaging method of Hadamard coding transformation based on convolutional neural network is proposed.Enlarging the pixels of the encoding matrix at the image acquisition,and realizes the acquisition of low spatial resolution and high spectral data(LrHS)with more and richer spatial information with the low-order encoding matrix.The network uses the feature extraction module based on the DeepLabV3+network to extract features from LrHS to form a deep-level feature ma.Through the feature super-resolution fusion module combining sub-pixel convolution and feature fusion module,deep-level image features are fused with image-level features in high spatial resolution multispectral data(HrLS).Finally the high spatial resolution hyperspectral data(HrHS)is obtained.Through experiments,the method does not exceed 50ms for super-dividing a single group of hyperspectral images,and the image indicators perform well.That indicates that the system can achieve rapid imaging of high-spatial-high-spectral resolution images.The verification through simulation experiments of public data sets and self-built systems.The experimental results show that this method takes no more than 50ms to reconstruct a single group of hyperspectral images,and the image indicators perform well,which can achieve fast high-spatial-high-spectral resolution imaging.