Abstract
Recent works have shown that the task of wireless transmission of images can be learned with the use of machine learning techniques. Very promising results in end-to-end image quality, superior to popular digital schemes that utilize source and channel coding separation, have been demonstrated through the training of an autoencoder, with a non-trainable channel layer in the middle. However, these methods assume that any complex value can be transmitted over the channel, which can prevent the application of the algorithm in scenarios where the hardware or protocol can only admit certain sets of channel inputs, such as the use of a digital constellation. Herein, we propose DeepJSCC-Q, an end-to-end optimized joint source-channel coding scheme for wireless image transmission, which is able to operate with a fixed channel input alphabet. We show that DeepJSCC-Q can achieve similar performance to models that use continuous-valued channel input. Importantly, it preserves the graceful degradation of image quality observed in prior work when channel conditions worsen, making DeepJSCC-Q much more attractive for deployment in practical systems.
Introduction
Traditional wireless communication systems rely on the separation principle, decoupling source coding (data compression) from channel coding (error correction). While theoretically optimal in the limit of infinite block length, practical implementations often suffer from the cliff effect: a sudden, catastrophic failure in decoding when the channel signal-to-noise ratio (SNR) drops below a specific threshold.
Recent advances in Deep Joint Source-Channel Coding (DeepJSCC) have demonstrated "graceful degradation," where image quality declines smoothly with increasing noise rather than failing abruptly. However, standard DeepJSCC models typically assume the channel can accept any complex-valued input, which is incompatible with the fixed digital constellations (e.g., QPSK, QAM) used in real-world radio hardware. DeepJSCC-Q addresses this by constraining neural network outputs to these standard digital constellations.
Proposed Architecture: DeepJSCC-Q
The core innovation of DeepJSCC-Q is a soft-to-hard quantization strategy that makes discrete constellation mapping differentiable for end-to-end training. During training, a "soft" approximation is used where outputs are represented as softmax-weighted sums of constellation points based on Euclidean distance. For inference, the model snaps to the nearest "hard" constellation point.
The architecture incorporates Attention Feature (AF) modules to enable a single model to adapt to varying channel conditions dynamically. Furthermore, KL Divergence Regularization is employed to ensure uniform usage of constellation points, maximizing power efficiency. The model utilizes a fully convolutional encoder and decoder with Pixel Shuffle for upsampling and Generalized Divisive Normalization (GDN) for effective density modeling.
Experimental Results
The model was evaluated on the CIFAR-10 dataset, consisting of 60,000 32x32 RGB images. Results indicate that constraining the output to digital constellations does not significantly degrade performance compared to unconstrained models. Notably, DeepJSCC-Q outperforms traditional separation-based schemes, such as BPG (Better Portable Graphics) compression combined with LDPC (Low-Density Parity-Check) codes, across a wide range of SNRs.
As shown in the performance comparison above, DeepJSCC-Q maintains a steady, gradual decline in quality even when the actual channel SNR is lower than the estimated SNR ($SNR_{Est}$), effectively defeating the cliff effect that plagues BPG+LDPC systems. Furthermore, as the constellation size ($M$) increases (e.g., moving to 4096-QAM), the performance of DeepJSCC-Q converges to that of the unconstrained continuous-valued DeepJSCC.
Conclusion and Future Work
DeepJSCC-Q demonstrates that AI-driven image transmission can be made compatible with standard radio hardware without sacrificing the benefits of joint source-channel coding. While promising, future work is needed to scale the architecture to high-definition imagery and evaluate performance over complex fading channels typical of mobile environments. Additionally, the computational latency and power consumption of DNN-based encoders compared to hardware-optimized traditional decoders remain areas for further investigation.
Summary updated