Deepjscc: The Future Of Wireless Video Transmission

We propose a demonstration of a joint source-channel coding (JSCC) scheme, called DeepJSCC, for wireless video transmission. Unlike conventional digital communication systems, which rely on separate source and channel coding, DeepJSCC is a purely data-driven end-to-end approach inspired by recent developments in deep learning. The encoder and decoder of DeepJSCC consist of a pair of neural networks that are optimized jointly through training. DeepJSCC can learn to transmit over the air without any explicit code design and achieve similar or better performance than conventional transmission schemes. Moreover, DeepJSCC is able to avoid the “cliff effect”, which the conventional digital designs suffer from. The cliff effect refers to the complete failure of digital transmission when the channel condition deteriorates. In contrast, DeepJSCC provides graceful degradation of video quality with the channel quality (as in analog transmission). Thus, DeepJSCC does not suffer from drop out of video as conventional digital communication systems do. The demonstration consists of a pair of software defined radios that act as the transmitter and receiver. We will first show the cliff effect that the conventional digital communication suffers from with live plots of channel quality vs video quality as well as the video itself. Next, we will implement DeepJSCC on the same hardware showing the smooth adjustment of the video quality with varying channel quality. Interactivity will consist of attendants moving the radio to introduce channel distortions and they will be able to see the effects of distortions they introduce on a live video. We believe this demonstration will inspire audiences on the power of deep learning for solving the challenges of wireless video transmission, and how it can revolutionize the future of communications, which has so far depended on a strictly modular design consisting of many complex expert-designed components. Sample images and results: https://ipc-lab.github.io/deepJSCC-demo/