Enhancement of Image Transmission Through Deep Joint Source- Channel Coding under Convolutional Neural networks

Abstract

Recent influential research shows that the optimality of separation is valid only under the assumptions of unlimited delay and complexity. Moreover, it does not hold in multi-user scenarios or with non-ergodic source or channel distributions. Addressing this, our EDeepJSCC scheme outperforms conventional digital transmission, which combines LDPC and JPEG/BPG compression with a capacity-achieving channel code, particularly at low signal-to-noise ratios (SNR) and restrained bandwidth within the presence of (AWGN).Unlike conventional methods, deep JSCC circumvents the "cliff effect" and demonstrates a gradual performance degradation as the signal-to-noise ratio (SNR) decreases. Notably, EDeepJSCC can communicate without explicit pilot signals or channel estimation, significantly surpassing autoencoder-based solutions in generating robust and compact codes from image pixels. It performs comparably to or better than state-of-the-art (SOA) solutions. Our solution, EdeepJSCC, utilizes convolution autoencoders and introduces three architectures of differing complexity. We implemented a multiple-description DeepJSCC scheme and tested it with real-world sources and channels. Our numerical results indicate that EdeepJSCC is capable of progressively transmitting images with minimal end-to-end performance degradation compared to single transmission techniques. Additionally, it performs comparably to SOA digital progressive transmission schemes in low SNR and small bandwidth conditions, with the added benefit of graceful degradation as channel SNR decreases.