Introducing supervised classification into spectral VQ for multi-channel image compression

In the last years, on-board data compression has become an urgent need and a lot of study has been directed toward the development of efficient techniques, able to achieve a good trade-off between bit rate reduction and quality degradation. For the particular case of multispectral and hyperspectral images, an appropriate quality measure would take into account the impact on classification, besides the introduced visual distortion. A new approach to vector quantization has been recently proposed, based on a sort of clustering performed in the features domain. This represents a first step toward the combination of compression and classification into a single operation but does not ensure sufficiently precise and reliable classification results. They propose to combine supervised classification and spectral vector quantization (SVQ) into a new compression technique in which visual distortion and classification accuracy can be a-priori determined according to the particular application which data are addressed to. Experimental results demonstrate that the proposed approach can be successfully applied both in “visual” and in “classification” mode