This website is best viewed with the firefox browser.


ipi logo
Image Processing and Interpretation Research Group

Second Generation Image Compression Techniques

The storage and transmission of digital images requires large amounts of data. For instance, a typical uncompressed METEOSAT image consists of approximately 18 megabytes of data. Since it produces an image every 30 minutes, a single satellite generates about 860 megabytes each day. Medical scanners, which are increasingly used in diagnosis, produce a large set of digital images for each examination: in the case of a magnetic resonance (MR) scanner, typically 128 images are recorded, corresponding to a total of 8 megabytes of data.

The aim of data compression techniques is to reduce the amount of data needed to accurately represent an image, such that this image can be economically transmitted or archived. Recent research has focused on image coding using image dependant base functions. Two techniques have been developed:

segmented image coding logo Segmented Image Coding (SIC)

SIC is a region-oriented coding, which considers images to be composed of regions of slowly varying image intensity, separted by image edges. The texture inside each region is approximated by a bi-variate polynomial. The coefficients of these polynomials, and the location of the contours separating the image regions constitute the contour-texture image model.

The visually extremely important image contours are always well retained by this technique, which guarantees a reasonable image quality, even at very low bit rates. In contrast to some conventional compression techniques, region oriented techniques are very well suited for progressive image transmission: in the reconstruction phase, the image is gradually built up. This is important in, e.g., database applications, where the gradual image build-up allows the user to quickly decide whether or not the transmitted image is the correct one required.

Our research concerning region oriented coding concentrates on the representation of the image texture as a weighted sum of orthogonal base functions. Such orthogonal polynomials necessarily depend on the shape of the region on which they are orthogonal and therefore must be generated for every single region. The generation of these base functions and the representation of the image texture requires a very large number of computations. We have developed a collection of algorithms to reduce the complexity of these tasks.

warp logo Warped polynomial Image Coding (WIC)

In this technique, the full image is approximated by a weighted sum of orthonormal image dependant base functions. These base functions are related to the time-warped orthogonal polynomials and are (orthonormal) pulses of varying width and height. In fact, the pulse width and height are inversely proportional to the local horizontal and vertical spatial image bandwidths, respectively. They are described by a two sequences of warping functions.

The image consisting of the coefficients of the base functions in the orthonormal expansion resembled a subsampled version of the original image. Therefore, the WIC image coding method can be interpreted as a form of adaptive subsampling.

Monday, 29-Dec-2003 17:56:09 CET