| Hyperspectral remote sensing images are obtained by the reflection of different wavelength electromagnetic waves on the same object. It is generally composed of hundreds of bands and containing a large number of fine spectral information. Hyperspectral remote sensing images are an important ways of ground monitoring. It plays a very important role in water-quality assessment, resource prospection and military reconnaissance. However, hyperspectral images are kind of large amounts data, which makes it big problem in transmission and storage in the limited bandwidth and storage resource constraints. Therefore, the compression method is required so that it can be widely used.Based on the spatial and spectral correlations of hyperspectral images, this thesis proposes two effective compression methods by multi-scale dictionary learning and classified dictionary learning. It also explore the feasibility of online dictionary learning on hyperspectral images. The detailed work is concluded as follow.(1) The statistical properties of spatial and spectral correlations has been analyzed, so that the compression algorithm can be designed specifically.(2) we propose an efficient compression algorithm based on wavelets and dictionary learning for hyperspectral images. The multi-scale training samples are obtained by wavelets decomposition and followed by dictionary learning algorithm. In the process of sparse coding, we using the defined frequency selection factor to simplify the matching pursuit process. The results suggest that this algorithm can improve the SNR of hyperspectral images reconstruction, and it can obtain sparse solutions more quickly than using the traditional methods.(3) we propose another compression algorithm based on the classified dictionary learning. Quadtree decomposition is used to slice the image into nonoverlapping patches so that we can train the dictionary in spatial domain and transform domain separately. In the process of sparse coding, using the pseudo inverse approach instead of OMP to reduce the calculation. The experiments results show that the proposed algorithm provides a much better compression performance compared with others, the pseudo inverse method also improves the speed of sparse coding greatly. (4) The performance of traditional dictionary learning method and the online dictionary learning and fast orthogonal dictionary learning approach has been compared.In conclusion, this thesis propose two efficient compression algorithm for hyperspectral images based on the distinct properties of hyperspctral images and sparse coding. The result show that these two compression algorithms both have impressive performance. They can be used as an effective way of hyperspectral images compression. |
