Enhancement of Image Transmission through ML Based Joint Source Channel Coding Beza Mezmur Hawaz

Abstract

Reliable wireless image transmission under low signal-to-noise ratio (SNR) and limited bandwidth remains a significant challenge because conventional communication systems based on Shannon's separation theorem often suffer from performance degradation and the cliff effect under adverse channel conditions. Although Deep Joint Source–Channel Coding (DeepJSCC) has

shown promising performance, further improvements are needed to enhance image reconstruction quality and transmission robustness. This study proposes an Enhanced Deep Joint Source–Channel Coding (EDeepJSCC) framework based on convolutional neural networks (CNNs) for end-to-end wireless image transmission over Additive White Gaussian Noise (AWGN) and Rayleigh fading channels. The proposed framework jointly optimizes source and channel coding to improve reconstruction quality without requiring explicit channel estimation. Experiments were conducted using the CIFAR-10 dataset, consisting of 50,000 training images and 10,000 testing images, and implemented in TensorFlow using the Adam optimizer under multiple SNR conditions and bandwidth ratios. Performance was evaluated using Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), and SNR.

Experimental results demonstrate that the proposed EDeepJSCC framework achieves improved image reconstruction quality and more graceful degradation than conventional JPEG/BPG with LDPC-based transmission, particularly under low-SNR and bandwidth-constrained conditions. These findings indicate that EDeepJSCC is a promising solution for robust wireless image transmission in next-generation communication systems