Competitive learning algorithms for low bit-rate image compression.

摘要

Highly efficient quantization is an integral part of any effective image compression scheme. As such, it is worth investigating the possible performance improvements that can be attained by using alternative methods rather than the standard techniques. In this work, the efficacy of competitive learning algorithms for quantization is investigated and its performance compared with that of the standard Lloyd algorithm. In specific, a competitive learning quantizer is developed featuring the winner-take-all method and bias learning.{09}This scheme is shown to consistently outperform the Lloyd algorithm in the low bit rate case, and the reasons for this are examined in detail. An explanation for this superior performance is presented, and a number of experimental results a shown to support this hypothesis. Finally, the performance and computational complexity of this scheme relative to previously developed competitive learning methods are discussed.