Hyperspectral Image Compression And Reconstruction Algorithm

Hyperspectral images are superior to traditional images in terms of spectral resolution,information volume,and information acquisition capability.However,excessive redundancy and excessive data volume can result in a significant reduction in image processing efficiency.Image compression is a good solution for this problem.By compression,not only the image dimension is highly reduced,but also the most discriminant information will be extracted for reducing the redundant as much as possible.Conventional compression method are often proposed for gray level image and RGB image.When dealing with hyperspectral images,they cannot compress image effectively because of the many bands and great redundant.Consider of this problem,we proposed two tensor based compression method in this paper.Tensor learning is an important theory of machine learning.A tensor is a multidimensional array,which can represent data of any orders.Hyperspectral image in this paper can be represented as three order tensor.A tensor can preserve both the spatial structure and spectral continuity,which helps to take full advantage of hyperspectral images.The first method is multiway tensor projection.It can be treated as tensor expansion over the conventional PCA algorithm.The algorithm obtains the transformation matrix satisfying the maximum divergence condition by each order of the tensor image.Eventually a low dimensional coefficient tensor and transformation matrices along each order are obtained.The algorithm can be seen as a process of a high dimensional tensor decomposition into low dimensional tensor and low dimensional matrices.Since the decomposed data are of low dimension and far less than the dimension of the original hyperspectral image,efficient compression can be achieved.The second method is called patch based low rank tensor decomposition.The method comprehensive utilized tensor patch,cluster,tensor dictionary learning,and low rank decomposition,the original problem is finally transferred into many sub problem of low rank decomposition.Since the sparse constraint,the final tensor and matrix are all of very low dimension.Compared with the first method,the second one can compress only one hyperspectral image,and has better compression performance.To testify the effectiveness of the proposed methods,we conduct compression and reconstruction experience on this two methods.Experience results on different hyperspectral image dataset have shown that,the first method performs better than conventional compression method,and the second method performs better than the first one and conventional ones.