A Neural Network-Based Compressive LDPC Decoder Design Over Correlated Noise Channel

To combat the correlated noise and utilize the non-equiprobability of the source, we propose a novel Low-density parity-check (LDPC) decoder design based on a priori information and neural network, which includes two neural networks, namely Correlated-Denoising Convolutional Neural Network (Correlated-DnCNN) and Belief Propagation-Deep Neural Network (BP-DNN), respectively. Correlated-DnCNN is used to remove the correlation of channel noise and reduce the noise power. BP-DNN is used to eliminate the harmful effect of short cycles of the Tanner graph and explore the priori information of the source. With the priori information to assist in initializing the input nodes of the BP-DNN network, we introduce compression capability to the LDPC, thus the proposed scheme can improve efficiency significantly when there is a large amount of compressible data in the physical layer. To train the Correlated-DnCNN and BP-DNN jointly, we design a new three-objective joint denoising-decoding loss function, which can not only guarantee the decoding performance, but also constrain the noise to conform to the normal distribution as much as possible, and improve the network convergence performance. Experiments show that the proposed decoding framework can remove the noise correlation, reduce the noise power, and improve the decoding performance effectively.