A competitive learning algorithm for non-zero memory codebook design in encoding of CT sequences

The implementation of artificial neural networks (ANN) in various aspects is increasing day by day. One of the major applications is image compression. Here an algorithm has been developed for use in the encoding of computed tomography (CT) image sequences. The method is based on application of an ANN distributed system which classifies all possible m×m (here 4×4) blocks into a smaller number well distinct classes of vectors. In an extension of the Kohonen self organising net called the frequency sensitive competitive learning (FSCL) algorithm, the required time for obtaining an ignorable error will depend on both the distortion and the number of iterations which, are more or less equal for all units. The application of an ANN to vector quantisation (VQ) stems from the concept that in the usual methods the error between each input pattern and the pattern of the codebook (word), is calculated without regarding the weight of each pixel value in the entire pattern. A proper ANN exploits this concept in an efficient classification of various patterns in an image and/or sequence of images. This significantly decreases the artefact, such as the blocking effect which normally appears in ordinary VQ reconstructed images at a low bit rate. In the case of sequences, interframes correlation is exploited in the provision of a common codebook for highly correlated frames. Further, the redundancy is decreased by optimal decomposition of the sequence into the most correlated subsequences