scientific article; zbMATH DE number 2154095

From MaRDI portal
Publication:4664644





The authors' summary: A usual technique for data compression is vector quantization. There are some well-known algorithms for vector quantization such as the Linde-Buzo-Gray algorithm and the self-organizing feature map (SOFM). This paper presents some results related to a SOFM-based approach for constructing the mapping between the codebook of the quantizer and the channel signal set, the classical vector quantization LBG algorithm, with its modified variant for the noise protection (LBG with splitting) in the case of a binary communication channel. This approach was introduced by Leung and Chan. The modified learning rule for SOFM proposed for this mapping is as follows: given a modulation system, a Delaunay neighborhood structure on the channel set can be defined. This neighborhood structure can be used as a neighborhood structure for training SOFM. Let \(x(t)\) be the \(t\)th training vector. The updating rules for the codevectors, the weights of the SOFM hypercube, are \(c_i(t+1)=c_i(t)+\alpha _t(x(t)-c_i(t))\) for \(c_j\in N_{i*}(v_t)\); \(c_j(t+1)=c_j(t)\) for \(c_j\in N_{j*}(v_t),\) where \(v_t\) monotonically decreases from the half of \(N\) matrix order to zero during the learning process, and the learning rate \(\alpha _t\) \((0<\alpha _t<1)\) decreases to zero monotonically. This approach uses a modified learning scheme of the SOFM neural network, so that SOFM provides automatically a mapping between the channel signal set and the codebook. We have accomplished a set of computer simulations for the previously discussed algorithm. We have used grey-scale images for training three different types of vector quantizers: the classical LBG quantizer, the LBG with splitting method, and the proposed vector quantizer, and we have made a comparison of the performances in the presence of the channel noise between these types. The structure of the SOFM hypercube matches that of the binary channel signal set for the Leung-Chan approach. In the LBG case, a random mapping is made between the channel signal set and the codebook. For the entire collection see [Zbl 0944.00096].











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4664644)