Digital Wireless Image Transmission via Distribution Matching

TL;DR

This paper introduces novel digital modulation schemes for deep learning-based joint source-channel coding in image transmission, effectively converting continuous channel symbols into discrete digital signals and outperforming existing systems.

Contribution

It proposes uniform and non-uniform modulation schemes based on constellation design and feature clustering, enabling practical digital transmission in deep JSCC systems.

Findings

Significant performance improvements over existing digital image transmission methods.

Effective modulation schemes that match the non-uniform distribution of channel symbols.

Use of feature clustering to generate optimized constellations for digital modulation.

Abstract

Deep learning-based joint source-channel coding (JSCC) is emerging as a potential technology to meet the demand for effective data transmission, particularly for image transmission. Nevertheless, most existing advancements only consider analog transmission, where the channel symbols are continuous, making them incompatible with practical digital communication systems. In this work, we address this by involving the modulation process and consider mapping the continuous channel symbols into discrete space. Recognizing the non-uniform distribution of the output channel symbols in existing methods, we propose two effective methods to improve the performance. Firstly, we introduce a uniform modulation scheme, where the distance between two constellations is adjustable to match the non-uniform nature of the distribution. In addition, we further design a non-uniform modulation scheme according to the output distribution. To this end, we first generate the constellations by performing feature clustering on an analog image transmission system, then the generated constellations are employed to modulate the continuous channel symbols. For both schemes, we fine-tune the digital system to alleviate the performance loss caused by modulation. Here, the straight-through estimator (STE) is considered to overcome the non-differentiable nature. Our experimental results demonstrate that the proposed schemes significantly outperform existing digital image transmission systems.