Spectral Dictionary Learning Based Multispectral Image Compression

Multispectral image encoding/decoding methods using spectral dictionary learning and sparse representation to fully exploit spectral features are proposed. In the scheme, K-SVD is first adopted for training a redundant dictionary from typical similar spectra. Then the sparse representative coefficients of each spectrum are obtained by the dictionary for spectral redundancy removal. Finally the equivalent nonzero sparse coefficients are quantified and stored. Experimental results show the superior spectral reconstructed performance compared with sample principal component analysis (PCA) and classical adaptive PCA at the same or even lower bit rates. Besides, the spectral dictionary learning can also be combined with compressed sensing or spatial decorrelation technologies to further expand its application.