Gaussian Source Coding Based on P-LDPC Code

A lossy source coding system based on the protograph low-density parity-check (P-LDPC) code is proposed for Gaussian source compression. In the proposed system, the conventional belief propagation (BP) algorithm is modified to be a concatenated BP-inverse BP (BP-iBP) for encoding and decoding, where the iBP is constructed by a fully-connected layer of a neural network. Compared to the existing approximate message passing algorithm, the proposed BP-iBP realizes a float-to-bit compression with low complexity for arbitrary Gaussian sources. The BP-iBP is implemented based on the linking relation of the protograph; therefore, it is necessary to optimally design the protograph to obtain better rate-distortion function (RDF) performance. Regarding the coding optimal procedure, a mutual information iteration convergence (MIIC) algorithm is designed as the optimal criterion to determine the source P-LDPC code with minimum distortion. Inspired by the plane construction of quantum stabilizer code, a lattice topological splicing (LTS) algorithm is proposed for regularly building the protograph to reduce the code searching complexity. By using the MIIC and the LTS algorithms, the BP-iBP based on the designed P-LDPC code maintains good distortion performance close to the RDF limit. © 2023 IEEE.