Overview
Research has introduced TS-JSCC, a single-model adaptive framework designed for learned joint source-channel coding (JSCC) in wireless image transmission. This framework addresses the challenges of bandwidth-limited and time-varying visual links by aiming to support user-adjustable transmission rates and adapt to changing wireless channel conditions. Furthermore, it seeks to dynamically allocate resources based on spatial content within a simple encoder-decoder structure.
Research Context
Learned JSCC facilitates robust wireless image transmission through the joint optimization of transmitters and receivers over differentiable channel models. Existing approaches for content-adaptive or dynamic allocation in wireless image transmission often rely on various mechanisms, including entropy coding, context/probability prediction, explicit rate maps or masks, or auxiliary allocation networks. These methods can introduce complexities into the encoder-decoder pipeline and increase side-information overhead.
Approach
The TS-JSCC framework incorporates two primary mechanisms:
- L1-based Tail-Structured Sparsification: This objective is designed to encourage each feature token to maintain an active feature-channel prefix while simultaneously suppressing trailing feature-channels. This mechanism facilitates content-adaptive feature-channel allocation. It enables the transmission of compact side information through the transmission of the active prefix.
- Lightweight Stage-wise Neural Regulating Modules: These modules employ a normalized sparsity-control coefficient and the channel signal-to-noise ratio (SNR) to rescale intermediate features. This rescaling enables both single-model transmission rate adaptation and SNR adaptation.
The overall design aims to create a one-shot encoder-decoder architecture that operates without requiring extra structures or computations beyond the core JSCC model.
Findings
Experiments were conducted on the CIFAR-10, Kodak, and CLIC2021 datasets. The evaluations were performed under two distinct channel conditions: additive white Gaussian noise (AWGN) and Rayleigh fading. The findings indicate that TS-JSCC:
- Achieves strong rate-distortion performance when compared against current learned-JSCC baselines.
- Maintains competitive performance relative to idealized separation baselines considered in the study.
- Operates with a simple one-shot encoder-decoder, avoiding additional structures or computations.
Why This Matters
This research addresses the need for adaptive wireless image transmission solutions in dynamic environments. By developing a single model that can adapt to varying transmission rates, channel conditions, and spatial content, it offers a method to manage bandwidth and deliver robust visual data over wireless links.