Lossless Region-based Multispectral Image Compression

In this paper we present a lossless coding scheme for multispectral images. The algorithm di ers from classical lossless approaches of multispectral image coding (1, 2, 3) in the fact that it is based on an independent coding of spectrally homogeneous regions. Regions that present a common multispectral signature are segmented. Then, spectral prediction is performed within these regions and nally spatial prediction removes the remaining correlation in the error images. This spatial prediction is also performed inside the regions by a region growing prediction algorithm that exploits the spatial correlation within region boundaries. The motivations of using a region-based approach are twofold: i) to achieve better coding performances by adaptively exploiting spectral redundancies and ii) to introduce region scalability functionality to multispectral coding, where regions of interest are coded di erently according to user preferences. In recent papers (4, 5), we already suggested the use of arbitrarily shaped regions of support in order to compress multispectral images. In a lossy framework, we proposed a region-based KLT resulting in better performances than classical block-based KLT approaches. Lossy algorithms are not applicable in many scienti c elds due to precision constraints. For this reason we present a lossless method for the compression of multispectral images with arbitrarily shaped segments. Typically, lossless multispectral image coding methods are based on linear prediction between spectral bands in order to remove the spectral redundancy. Pixels from a given band are used to predict the pixel values for another band. The prediction coe cients are usually calculated from the statistics of both images by least-square criteria. Such a prediction is optimal in the least-square sense for the whole image but, due to the di erent spectral signatures present in the scene, a considerable error may be produced in some regions. By introducing the notion of region-based spectral prediction, regions with similar spectral signature are taken as support for prediction. Optimality for each region is obtained and the prediction error is substantially reduced. This paper is organized as follows. Section 2 discusses the possible multispectral segmentations. Section 3 proposes an ordering of the spectral bands before spectral prediction. Spectral prediction is explained in Section 4 and a clustering method that optimizes this prediction is presented in Section 5. Finally, the spatial decorrelation step is discussed in Section 6 and simulation results are given in Section 7.