Scalable and perception-aware deep joint source–channel coding for underwater acoustic image transmission

Abstract

Underwater acoustic communication is fundamentally constrained by limited bandwidth, severe multipath propagation, Doppler distortion, and strong time variability, which severely limit reliable image transmission. Conventional separationbased digital transmission schemes are sensitive to channel mismatch and often exhibit abrupt quality degradation. Existing deep joint source–channel coding (Deep JSCC) approaches, although more robust, are typically fixed-rate and offer limited perceptual stability under rapidly varying underwater conditions. This paper proposes a scalable, perception-aware Deep JSCC framework for underwater acoustic image transmission, in which a layered encoding architecture enables progressive reconstruction, allowing image quality to degrade gracefully at low signal-to-noise ratios (SNRs) and improve incrementally as channel conditions become more favorable, without relying on feedback mechanisms. Perceptual robustness is further enhanced through a perception-aware training objective that promotes structural and visual coherence beyond pixel-level fidelity. Extensive experiments under realistic underwater acoustic channel models demonstrate that the proposed method consistently outperforms conventional digital transmission and single-layer Deep JSCC baselines across a wide range of SNRs. At 0 dB, the proposed framework achieves a PSNR gain of approximately 1.7 dB over single-layer Deep JSCC and more than 7 dB over digital transmission, while improving SSIM by up to 0.09 and reducing LPIPS by approximately 28%. At 10 dB, it maintains a performance advantage of 1.7 dB in PSNR and achieves the lowest perceptual distortion among all methods. These results confirm improved structural preservation and perceptual stability, particularly in low-SNR regimes.