Research On Hyperspectral Image Preprocessing And Visualization

Hyperspectral images(HSIs)contain a large amount of spectral information in numerous contiguous channels.Compared with common single-channel image and multi-channel image,it has higher spectral recognition ability and can provide better detection and classification ability.Due to the influence of instruments and atmosphere in the process of spectral data acquisition,noise and low spatial resolution often occur.Hyperspectral images contain high dimensional data and their adjacent bands have high correlation.Therefore,how to display them in RGB tri-color display is a key problem.Aiming at the above problems,this paper studies the hyperspectral image denoising algorithm,hyperspectral image super-resolution algorithm and hyperspectral image visualization algorithm.The main work and innovations in the field of hyperspectral images in this paper include:1.A neural network architecture based on self encoder is proposed(HIS-UBDAE),improves the traditional denoising self encoder in the ordinary image denoising method,sends the adjacent or all bands of hyperspectral image into the encoder for coding,and finally encodes and outputs a clean denoised image after continuously updating the selected points in the potential space.Since all spectral information is learned in the network in the updating process,the comparison process of each training is only limited.It needs a single noise picture instead of the whole spectral image.This method can train a satisfactory image with only a small amount of data,and can effectively avoid the problem of insufficient data in hyperspectral images.The experimental results show that the denoising method proposed in this paper has better objective indexes and visualization effects in MPSNR,MSSIM and other aspects.2.Aiming at the problem of low spatial resolution of hyperspectral images,a prior super-resolution reconstruction algorithm PR-DCNNs is proposed.Using linear spectral decomposition,the super-resolution reconstruction problem is transformed into abundance mapping problem,which is extended by using a priori information.After finding the spectral characteristics of abundance and end elements,the mapping and spectrum are used through network models such as convolution network and residual network The experimental results show that PR-DCNNs algorithm is an effective hyperspectral resolution reconstruction algorithm.3.The existing linear prediction visualization algorithm is improved to optimize the performance of the algorithm,and an initialization band selection algorithm based on information entropy is proposed.When the initialization band with the largest difference is selected,the differences between different bands are compared through similarity,and the joint differences between multiple bands can be evaluated to select the band with the best display effect,The improved linear prediction visualization algorithm can significantly shorten the visualization time and improve the execution efficiency of the algorithm.