Deep Joint Source-Channel Coding for Wireless Image Transmission

Abstract

We propose a joint source and channel coding (JSCC) technique for wireless image transmission that does not rely on explicit codes for either compression or error correction; instead, it directly maps the image pixel values to the complex-valued channel input symbols. We parameterize the encoder and decoder functions by two convolutional neural networks (CNNs), which are trained jointly, and can be considered as an autoencoder with a non-trainable layer in the middle that represents the noisy communication channel. Our results show that the proposed deep JSCC scheme outperforms digital transmission concatenating JPEG or JPEG2000 compression with a capacity achieving channel code at low signal-tonoise ratio (SNR) and channel bandwidth values in the presence of additive white Gaussian noise (AWGN). More strikingly, deep JSCC does not suffer from the "cliff effect", and it provides a graceful performance degradation as the channel SNR varies with respect to the SNR value assumed during training. In the case of a slow Rayleigh fading channel, deep JSCC learns noise resilient coded representations and significantly outperforms separation-based digital communication at all SNR and channel bandwidth values.Index Terms-Joint source-channel coding, deep neural networks, image communications.

DeepJSCC-f: Deep Joint Source-Channel Coding of Images With Feedback

“…We first compare our scheme to the state of the art JSCC scheme for image transmissions, without the use of feedback. For this, we use the deep-JSCC algorithm from [42], further enhancing it by employing the architecture shown in Figure 4. This is also the special case of our scheme with L = 1; that is, a single transmission with a bandwidth ratio k/n = 1/6.…”

“…Most related prior work to the current paper are [39]- [43], which consider the JSCC problem, and propose autoencoder-based solutions for end-to-end optimization, without explicitly focusing on the compression or the channel coding problems. While [39] focuses on text as the information source and binary channels, and [43] deals with lossy data storage, image transmission over an additive white Gaussian noise (AWGN) wireless channel is studied in [41], [42]. In [42] the authors propose a fully convolutional autoencoder architecture, which maps the input images directly to channel symbols, without going through any digital interface, and show that the proposed deepJSCC architecture not only improves upon the concatenation of state-of-the-art compression and channel coding schemes in a separate architecture, but also provides graceful degradation with channel signal-to-noise ratio (SNR).…”

“…While [39] focuses on text as the information source and binary channels, and [43] deals with lossy data storage, image transmission over an additive white Gaussian noise (AWGN) wireless channel is studied in [41], [42]. In [42] the authors propose a fully convolutional autoencoder architecture, which maps the input images directly to channel symbols, without going through any digital interface, and show that the proposed deepJSCC architecture not only improves upon the concatenation of state-of-the-art compression and channel coding schemes in a separate architecture, but also provides graceful degradation with channel signal-to-noise ratio (SNR). This latter property, which is common to analog transmission schemes, provides significant benefits compared to digital schemes, which exhibit threshold behaviour; particularly when broadcasting to multiple receivers, or when transmitting over a time-varying channel.…”

“…We plot in Figure 2 the peak SNR (PSNR) for the transmission of images from the CIFAR10 dataset over an AWGN channel with a compression ratio of 1/3, which refers to the available channel uses per image pixel (please see below for the proper definition). Here we compare the performance of the deepJSCC-f scheme with the deepJSCC architecture from [42] and with the concatenation of BPG codec followed by LDPC channel coding. These two schemes do not exploit the feedback link.…”

Deep Joint Source-Channel Coding for Wireless Image Retrieval

“…Instead, the source measurements are directly mapped to channel symbols. A deep neural network (DNN) based JSCC has recently been shown to outperform state-of-the-art digital schemes for wireless image transmission [3]. Here, we consider both digital and JSCC architectures for image retrieval over wireless channels, which can be considered as task-based JSCC.…”