Research map / Image generation / Image restoration
Learned Image Transmission with Hierarchical Variational Autoencoder
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HJSCC, a hierarchical-VAE-based joint source-channel coder for image transmission whose rate adapts per image via learned priors, extended to feedback links with a generative formulation.
Problem
Deep JSCC for image transmission uses fixed-rate coding, struggles on high-resolution images, and lacks rate adaptation, especially when a feedback link is present.
Approach
Uses a hierarchical VAE (bottom-up/top-down paths) to produce multiple latent representations, maps each to channel symbols via Swin-based JSCC encoders whose lengths are masked by learned entropy priors; treats noisy transmission with feedback as a sampling process, and adds a rate attention module and spatial grouping to reduce length-signaling overhead.
Key results
- On Kodak and CLIC2022 at SNR 10 dB, HJSCC significantly outperforms DeepJSCC and SwinJSCC and achieves comparable performance to NTSCC, with the gap widening at higher resolution and CBR.
- With feedback, HJSCC gives about 1 dB PSNR gain over JSCCformer-f and DeepJSCC-f at high CBR.
- Spatial grouping reduces the length-information CBR overhead from 0.0200 to 0.0025 at lambda=64, at a cost of about 0.2 dB PSNR (32.24 to 32.03).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spatial grouping (ablation) | HJSCC · ImageNet, λ=64 | CBR: 0.0590 -> 0.0403 | Helped |
| add spatial grouping (ablation) | HJSCC · ImageNet, λ=64 | PSNR: 32.24 -> 32.03 | Hurt |
| add spatial grouping (ablation) | HJSCC · ImageNet, λ=16 | CBR: 0.0430 -> 0.0249 | Helped |
| add spatial grouping (ablation) | HJSCC · ImageNet, λ=16 | PSNR: 30.13 -> 29.96 | Hurt |
| add the rate attention module (ablation) | HJSCC · Kodak and CLIC2022, AWGN at 10 dB | PSNR | Helped |
| add a feedback link | HJSCC · Image transmission with feedback, high CBR | PSNR: about 1 dB | Helped |
| use a noisy feedback link | HJSCC with feedback · Image transmission with noisy feedback | PSNR | Hurt |
Limitations
- Length information must be sent over an error-free side link, with overhead of N_q bits per length group.
- Performance is slightly below NTSCC on some settings (described as comparable).
- Spatial grouping causes a slight PSNR degradation.
Experimental setup
Datasets: ImageNet
Benchmarks: Kodak CLIC2022
Metrics: PSNR CBR rate-distortion
Baselines: DeepJSCC SwinJSCC NTSCC BPG + LDPC JPEG + LDPC JPEG2000 + LDPC JSCCformer-f DeepJSCC-f
Builds on: hierarchical VAE (ResNet VAE) DeepJSCC learned image compression with hyperprior SwinJSCC NTSCC
Closest papers by meaning
- Bit Prioritization in Variational Autoencoders via Progressive Coding (ICML 2022 · similarity 0.88)
- Generative Latent Coding for Ultra-Low Bitrate Image Compression (CVPR 2024 · similarity 0.84)
- Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation (ICLR 2025 · similarity 0.84)
- Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words (ICLR 2024 · similarity 0.83)
- Deep Hierarchical Video Compression (AAAI 2024 · similarity 0.83)
- Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.