|The contemporary research in different application areas sets the task of the efficient archiving of multispectral images. For this, in most cases is necessary to process several images of the same object(s). Multispectral images are characterized by very high spatial, spectral, and radiometric resolution and, hence, by ever-increasing demands of communication and storage resources. Such demands often exceed the system capacity like, for example, in the downlink from satellite to Earth stations, where the channel bandwidth is often much inferior to the intrinsic data rate of the images, some of which must be discarded altogether. In this situation the high-fidelity image compression is a very appealing alternative. As a matter of fact, there has been intense research activity on this topic , focusing, particularly, on transform-coding techniques, due to their good performance and limited computational complexity. Linear transform coding, however, does not take into account the nonlinear dependences existing among different bands, due to the fact that multiple land covers, each with its own interband statistics, are present in a single image. Based on this observation, a class-based coder was proposed in that addresses the problem of interband dependences by segmenting the image into several classes, corresponding as much as possible to the different land covers of the scene. As a consequence, within each class, pixels share the same statistics and exhibit only linear interband dependences, which can be efficiently exploited by conventional transform coding.Satellite-borne sensors have ever higher spatial, spectral and radiometric resolution. With this huge amount of information comes the problem of dealing with large volumes of data. The most critical phase is on-board the satellite, where acquired data easily exceed the capacity of the downlink transmission channel, and often large parts of images must be simply discarded, but similar issues arise in the ground segment, where image archival and dissemination are seriously undermined by the sheer amount of data to be managed. The reasonable approach is to resort to data compression, which allows reducing the data volume by one and even two orders of magnitude without serious effects on the image quality and on their diagnostic value for subsequent automatic processing. To this end, however, is not possible to use the general purpose techniques as they do not exploit the peculiar features of multispectral remote-sensing images, which is why several ad hoc coding schemes have been proposed in recent years.The transform coding is one of the most popular approaches for several reasons. First, transform coding techniques are well established and deeply understood; they provide excellent performances in the compression of images, video and other sources, have a reasonable complexity and besides, are at the core of the famous standards JPEG and JPEG2000, implemented in widely used and easily available coders. The common approach for coding multispectral images is to use some decorrelating transforms along the spectral dimension followed by JPEG2000 on the transform bands with a suitable rate allocation among the bands.Less attention has been devoted to techniques based on vector quantization (VQ) because, despite its theoretical optimality, VQ is too computationally demanding to be of any practical use. Nonetheless, when dealing with multiband images, VQ is a natural candidate, because the elementary semantic unit in such images is the spectral response vector (or spectrum, for short) which collects the image intensities for a given location at all spectral bands. The values of a spectrum at different bands are not simply correlated but strongly dependent, because they are completely determined (but for the noise) by the land covers of the imaged cell. This observation has motivated the search for constrained VQ techniques, which are suboptimal but simpler than full-search VQ, and show promising performances.
Multispectral images require large amounts of storage space, and therefore a lot of attention has recently been focused to compress these images. Multispectral images include both spatial and spectral redundancies. Usually we can use vector quantization, prediction and transform coding to reduce redundancies. For example, hybrids transform/vector quantization (VQ) coding scheme is proposed. Instead, Karhunen-Loeve transform (KLT) is used to reduce the spectral redundancies, followed by a two-dimensional (2D) discrete cosine transform (DCT) to reduce the spatial redundancies. A quad-tree technique for determining the transform block size and the quantizer for encoding the transform coefficients was applied across KLT-DCT method. A frequently used approach is the use of a wavelet transform (WT) to reduce the spatial redundancies and KLT to reduce the spectral redundancies, and then encoded using the 3-dimensional (3D) SPIHT algorithm.
Similar problems exist in processing and archiving of multi-view images.
The state-of-the-art analysis shows that despite of the vast investigations and various techniques used for the efficient compression of multispectral images, a recognized general method able to solve the main problems is still not created.
The R&D result offers one new method for compression of multispectral and multi-view images, based on the modified Inverse Difference Pyramid (IDP) Decomposition. Here is supposed to process a group of images together, using the similarity between these images.