New Neural JSCC Codec Boosts Video Robustness Against Erasures

A team of researchers from Nokia Bell Labs and academic institutions has introduced a novel semantic-aware neural joint source-channel coding (JSCC) framework designed to maintain video quality over block erasure channels — a common impairment in wireless and packet-switched networks. The paper, accepted for IEEE VTC FALL 2026, proposes two complementary designs: spatial-domain JSCC and feature-domain JSCC. In the spatial domain, video frames are partitioned into blocks, allowing fine-grained erasure handling with either uniform or semantic-guided non-uniform erasure strategies. In the feature domain, latent representations are split into chunks, so missing features can be semantically recovered while preserving overall spatial consistency.

Comprehensive experiments under varying erasure probabilities reveal clear trade-offs. Spatial-domain JSCC significantly outperforms when dealing with random localized losses, maintaining visual continuity and limiting error propagation. Feature-domain JSCC, on the other hand, provides superior robustness to distributed erasures across the frame and achieves higher peak signal-to-noise ratio (PSNR) in low-erasure scenarios. The analysis highlights how spatial continuity can be exchanged for semantic redundancy, offering a principled way to design task-aware video communication systems. This work paves the way for next-generation video codecs that prioritize semantic meaning over raw pixel fidelity.