Deep Learning Aided Joint Source-Channel Coding for Wireless Networks

Abstract

Wireless image transmission and estimation techniques have found a variety of applications in 5G Internet of Things (IoT) under the finite block length transmissions. However, the traditional digital transmission scheme is known to suffer from the "cliff effect". By using deep machine learning techniques, we propose an autoencoder-based joint source-channel coding (JSCC) scheme for image transmission and estimation in multi-user wireless sensor networks. At the wireless edge, sensors independently observe a common image and encode the noisy observation to a complex vector. At the receiver side, the decoder receives all the transmitted signals to estimate the common image. We design an end-to-end autoencoder to complete this task, and combine mean squared error (MSE) and structural similarity index matrix (SSIM) as the loss function to explore both pixel-wise and structural features of the images. Our proposed scheme is evaluated in two types of noisy observation scenarios, multi-focus scenario and partial observation scenario, considering both additive white Gaussian noise (AWGN) channel and Rayleigh fading channel, and comparing the performance with the traditional transmission scheme. In both scenarios, our scheme achieves better peak-to-noise ratio (PSNR) performance, especially when the channel bandwidth is limited or the signal-to-noise ratio (SNR) is low.