Improving Wireless Image Transmission Using Machine Learning-Based Joint Source–Channel Coding

Abstract

Transmitting digital images reliably over bandwidth-limited wireless links remains difficult, particularly when signal-to-noise ratios (SNRs) are low. Under unfavorable channel conditions, conventional source–channel separation architectures may exhibit substantial deterioration and abrupt reconstruction failure. Although deep joint source–channel coding (DeepJSCC) provides an alternative, improving reconstruction fidelity and resilience remains an important objective. This study proposes Enhanced Deep Joint Source–Channel Coding (EDeepJSCC), an end-to-end convolutional neural network (CNN) framework for image delivery through additive white Gaussian noise (AWGN) and Rayleigh fading channels. Joint optimization of the coding and reconstruction stages enables the approach to operate without a separate explicit channel-estimation procedure. Experiments use CIFAR-10, comprising 50,000 training images and 10,000 test images, and are implemented in TensorFlow with Adam optimization over multiple SNR settings and bandwidth ratios. Performance is characterized using peak signal-to-noise ratio (PSNR), mean squared error (MSE), and channel SNR. The reported comparisons suggest improved image recovery and more progressive performance degradation than conventional JPEG/BPG compression paired with LDPC coding, particularly when channel quality and available bandwidth are limited. The findings support further investigation of EDeepJSCC for resilient visual communication in future wireless systems.