Lossy Compression of Noisy Sparse Sources Based on Syndrome Encoding

Data originating from devices and sensors in Internet of Things scenarios can often be modeled as sparse signals. In this paper, we provide new source compression schemes for noisy sparse and non-strictly sparse sources, based on channel coding theory. Specifically, nonlinear excision filtering by means of model order selection or thresholding is first used to detect the support of the non-zero elements of sparse vectors in noise. Then, the sparse sources are quantized and compressed using syndrome-based encoders. The theoretical performance of the schemes is provided, accounting for the uncertainty in the support estimation. In particular, we derive the operational distortion-rate and operational distortion-energy of the encoders for noisy Bernoulli-uniform and Bernoulli-Gaussian sparse sources. It is found that the performance of the proposed encoders approaches the information-theoretic bounds for sources with low sparsity order. As a case study, the proposed encoders are used to compress signals gathered from a real wireless sensor network for environmental monitoring.