Signal compression using growing cell structures: A transformational approach

Abstract

We present an adaptive compression system (ACS) that compresses signals using signal primitives obtained by the self organizing neural architecture growing cell structures (GCS) [6]. We determine the length w(max) of the primitive that maximizes the compression. We decompose the signal into w(max)-long segments. Then GCS is trained to adaptively construct categories from segments. A reconstruction of the original signal may be obtained as a sequence of GCS categories with some error. We analyze the performance of ACS using two criteria: CR and PRD. We define CR as the ratio of the memory space required to hold the original signal over that required by the compressed version of the signal. We define PRD as the error between original signal and reconstructed signal from the compressed signal information. CR and PRD counteract providing a trade-off among the compression potential and the reconstruction quality of ACS. We apply ACS to electrocardiogram (ECG) signals.