Research On Hyperspectral Image Compression Based On Clustering

Hyperspectral images have more than one hundred spectral bands.Because hyperspectral images can provide rich spectral and spatial information,the applications of hyperspectral images are becoming more and more common,such as ground objects,target detection,and unmixing.Although image analysis can benefit from the rich data of hyperspectral images,large amounts of data can place a heavy burden on the storage and transmission of hyperspectral images.Therefore,how to efficient compress is becoming an important issue in hyperspectral image applications.Compression techniques can be broadly divided into two broad categories: lossless and lossy compression methods,depending on whether the original image can be accurately regenerated into the original image.For lossless compression,the key is to eliminate data redundancy without losing information.In contrast,lossy compression,while losing some information,yields a higher compression ratio than lossless compression,lossy compression is a promising research topic in hyperspectral image compression and it is the focus of this thesis.The main contents of this thesis are:1.Neural networks are mostly used in 2D data images,and they are not fully applied to 3D hyperspectral images.By reading a large number of literatures on neural networks,a hyperspectral image compression algorithm based on adaptive band clustering principal component analysis and backpropagation neural network is proposed.The algorithm compresses the hyperspectral image by using the input layer of the back propagation neural network to the corresponding compression of the hidden layer and the decompression of the hidden layer to the output layer.The algorithm can effectively improve the image signal to noise ratio.2.Hyperspectral imagery is studied and its correlation between spectra is greater than spatial correlation.A hyperspectral image compression algorithm based on prediction and vector quantization is proposed for this feature.Firstly,by using the inter-spectral correlation,by predicting the pixel data of the previous band,the predicted result will continue to predict the next band.By setting reasonable parameters,more than 95% of the bands in the prediction step can predict the band by the prediction result.data.Finally,the prediction data is compressed using vector quantization.3.A hyperspectral image compression algorithm based on sparse representation is proposed.The algorithm is aimed at solving the problem of large amount of hyperspectral data and long compression time.It only compresses non-zero coefficients and dictionary,and the computational complexity is low.At the decoding end,the image can be quickly reconstructed according to the dictionary and non-zero coefficients,which significantly shortens the reconstruction time.