(A) new vector quantization algorithm using the self-organizing map of multiple hypercube structure다중 하이퍼큐브 구조를 갖는 자기조직형 신경회로망 모델을 이용한 새로운 벡터양자화 알고리즘에 관한 연구
As digital communication has become important, the theory and practice of data compression have received increasing attention. Image compression and speech compression (or image coding and speech coding) are probably currently most important applications of data compression. A fundamental result of Shannon``s rate-distortion theory is that better performance is always achievable in theory by coding vectors instead of scalars, even if the scalars have been produced by preprocessing the original input data so as to make them uncorrelated or independent. Vector quantization for image compression suffers from edge degradation in the reproduced images that provides undesirable artifacts near sharp edges or in the texture areas and requires the tremendous computational complexity to search the whole codebook to find the closest codevector for an input vector.
A more fundamental benefit of formulating vector quantization using self-organizing feature map (SOFM) training algorithms that have been developed can be adapted to the problem of training vector quantizers. It is well known that the SOFM algorithm can yield near optimal by approximating the pdf of input signals within finite iteration sequence and it was also reported that the SOFM technique generates codebooks of more uniform utilization than those produced by K-means clustering.
A new learning algorithm is proposed based on the utilization of an image block (or vector) feature measure to minimize the edge degradation in reproduced images. To achieve this minimization, the incoming block is tested by investigating its activity, and a corresponding weighting factor is assigned. The assigned weighting factor is used to determine the learning rate adaptively. The proposed algorithm can provide the edge preserving image without an additional vector classification procedure, which is not applicable to the VQ blocks larger than 4x4.
As a new neural network scheme embedding the edge preserving characteristic, a m...