Scan predictive vector quantization of multispectral images

It has previously been shown that efficient scanning techniques serve to enhance the performance of image compression schemes like vector quantization. However, work in this direction has mostly concentrated on using an alternate scanning technique based on space filling curves. Recently we defined the notion of a scan model that can be viewed as defining a connected scan of an image and gave algorithms to compute optimal scans for a variety of cost functions. Since an optimal scan varies from image to image, an encoding of the scan has to accompany an encoding of an image with respect to the scan. In most applications this turns out to be prohibitively expensive. By predictive scanning, we mean scanning an image along a prediction of the optimal scan. For multispectral images, an optimal scan of one band serves well as the prediction of the optimal scan for a spectrally adjacent band due to the presence of spectral correlations. In this paper we show how predictive scanning techniques can be used along with vector quantization to give an efficient compression technique for multispectral image data. The technique captures spectral as well as spatial correlations and hence performs significantly better than techniques that use spectrally adjacent blocks to form a vector that is subsequently quantized.