Research On Hyperspectral Image Reconstruction Algorithm Based On Compression Sensing And Deep Learning

Hyperspectral image is spectral image with spectral resolutions ranging from 1 to 10 nm collected by an imaging spectrometer.Full of rich spectral information,hyperspectral image plays an unique and important role in remote sensing,medical detection,astronomical exploration and other fields.However,despite of the urgent need,the image quality of hyperspectral images has been affected by various reasons,with question about of information loss.If this problem is considered as a degradation model of image quality for hyperspectral images,how to recover the lost information through the reconstruction algorithm has always been the top focus of related researchers.In this paper,on the basis of the research progress at home and abroad,the hyperspectral image reconstruction algorithm is studied by applying compression sensing and deep learning.First of all,this paper analyses the problem of information loss before and after the transmission of hyperspectral images,introduces the compressed sensing theory and deep learning theory,and takes classical reconstruction algorithms of both theories as examples.In addition,the evaluation criteria for the quality of hyperspectral images are discussed.Secondly,in order to reduce information loss in the transmission of hyperspectral images,a two-dimensional stepwise orthogonal matching tracking algorithm based on compressed sensing is proposed.Through theoretical analysis,corresponding threshold is proposed,and by matching atoms with the same weight in one step,the new reconstruction algorithm could reconstruct image quickly under the two-dimensional measurement model,which have better performance with same sampling rate,and solves the problem that traditional compressed sensing may lose image information between rows.Finally,some hyperspectral images have lost spatial information.To solve this problem,a hyperspectral image reconstruction algorithm based on deep learning is proposed.After analyzing the possible problems of the existing learning-based reconstruction algorithm,by taking the spectral domain information of hyperspectral images into the loss function,the algorithm effectively improves the spatial resolution of hyperspectral images and recovers the lost spatial information while suppressing the spectral domain distortion.At the same time,the training method of this network has also been improved.By adding pre-training step and improving the discriminator loss function,the generative adversarial network could run more smoothly.