DecompJSCC separates noise adaptation from rate control
The DecompJSCC framework uses a new module to select image channels based on content.
Researchers developed DecompJSCC to handle image transmission. This framework separates signal-to-noise ratio (SNR) adaptation from rate adaptation.
Deep joint source-channel coding (DeepJSCC) maps images directly to channel symbols. It then reconstructs the images at a receiver. In wireless links, the SNR and available rate change over time. Using multiple models for these changes is costly to train and deploy.
Existing methods have limitations. Some methods tie latent formation to the rate request. Others use static masks after encoding. Static masks cannot adapt channel selection to the image content.
DecompJSCC uses SNR-conditioned block attention residuals in the encoder. This aggregates semantic representations across different stages. At the receiver, an adapter performs SNR-conditioned affine correction.
A post-encoding rate-adaptive channel semantic budgeting (RCSB) module is also used. This module uses a shared gate to select channels. It chooses channels based on image content and the requested count .
The encoder output remains independent of the rate request for a fixed image and SNR. The RCSB module keeps exactly the requested count of channels. It also adjusts their gains.
In tests using Kodak and CLIC2021, DecompJSCC outperformed SwinJSCC and SA-RA-JSCC. It beat SwinJSCC by 0.64 and 1.00 dB in peak signal-to-noise ratio (PSNR). It beat SA-RA-JSCC by 0.97 and 0.96 dB. It used only 7.5% of the Adaptive ModNet parameters used by SwinJSCC.
Key facts
- DecompJSCC outperforms SwinJSCC and SA-RA-JSCC in PSNR.
- The RCSB module selects channels based on image content.
- DecompJSCC uses only 7.5% of the parameters used by SwinJSCC's Adaptive ModNet.