Research map / Image generation / Image restoration
Towards Loss-Resilient Image Coding for Unstable Satellite Networks
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 a loss-resilient learned image compression method for GEO satellite progressive transmission, using spatial-channel rearrangement, mask conditional aggregation, and Gilbert-Elliot packet loss simulation during training.
Problem
Image transmission over unstable GEO satellite networks with frequent packet loss corrupts progressive learned image compression, causing pixel errors, color deviation, and unstable reconstruction quality.
Approach
Encoder-side spatial-channel rearrangement spreads each channel's information across several channels so losses are recoverable from correlated channels; decoder-side mask conditional aggregation feeds the loss mask as a condition; training uses Gilbert-Elliot bursty loss simulation instead of uniform random masks.
Key results
- Under 5% uniform packet loss on Kodak, the method reaches mean PSNR 26.850 dB with variance 0.015 at 0.131 bpp, versus 26.551 dB / 0.392 for the baseline mbt-mean and 26.657 dB / 0.064 for Random masking.
- At 20% uniform packet loss on Kodak (0.131 bpp), mean PSNR is 25.578 dB versus 23.257 dB for baseline and 25.217 dB for Random.
- Model complexity rises from 26.80 MB to 28.53 MB, with decoding time going from 39.30 ms to 50.17 ms per Kodak image.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SCR and MCA to the progressive LIC model | mbt-mean on Kodak · uniform packet loss, 5%, 10%, and 20% | mean PSNR (dB): 26.657 -> 26.850; 25.858 -> 26.342; 25.217 -> 25.578 (low bitrate) | Helped |
| add SCR and MCA to the progressive LIC model | mbt-mean on Kodak · uniform packet loss, 5%, 10%, and 20% | mean PSNR (dB): 27.754 -> 28.247; 27.320 -> 27.605; 25.886 -> 26.617 (middle bitrate) | Helped |
| add SCR and MCA to the progressive LIC model | mbt-mean on Kodak · uniform packet loss, 5%, 10%, and 20% | mean PSNR (dB): 29.256 -> 29.561; 28.569 -> 28.801; 27.620 -> 27.966 (high bitrate) | Helped |
| add SCR individually (ablation) | progressive LIC model · 10% packet loss | reconstruction performance | Helped |
| add MCA individually (ablation) | progressive LIC model · 10% packet loss | reconstruction performance | Helped |
| combine SCR and MCA (ablation) | progressive LIC model · 10% packet loss | loss resilience | Helped |
| add SCR (ablation) | progressive LIC model · coding performance | coding performance | No clear effect |
| replace regular convolution with ConvNeXt (ablation) | loss-resilient LIC on Kodak · uniform packet loss, 5%, 10%, and 20% | mean PSNR (dB): 26.850 -> 27.302; 26.342 -> 26.747; 25.578 -> 26.177 (low bitrate) | Helped |
| replace regular convolution with ConvNeXt (ablation) | loss-resilient LIC on Kodak · uniform packet loss, 5%, 10%, and 20% | mean PSNR (dB): 28.247 -> 28.843; 27.605 -> 28.062; 26.617 -> 27.314 (middle bitrate) | Helped |
| replace regular convolution with ConvNeXt (ablation) | loss-resilient LIC on Kodak · uniform packet loss, 5%, 10%, and 20% | mean PSNR (dB): 29.561 -> 30.088; 28.801 -> 29.509; 27.966 -> 28.329 (high bitrate) | Helped |
| train with the Gilbert-Elliot loss model instead of uniform random loss | loss-resilient LIC · simulated 15% packet loss and real GEO satellite network | average PSNR | Helped |
Surprising findings
- Spatial-channel rearrangement reduces spatial correlation yet has minimal impact on coding performance, which the authors attribute to the i.i.d. assumption of the hyperprior.
Limitations
- Hyperprior bitstream is assumed to be delivered losslessly via redundancy or ARQ, so loss protection for it is not modeled.
- Does not handle packet loss under autoregressive context models, since loss affects subsequent probability estimation.
Experimental setup
Datasets: Flickr2W Kodak CLIC professional validation
Benchmarks: Kodak CLIC professional validation
Metrics: PSNR bpp PSNR variance
Baselines: JPEG2000 Baseline (mbt-mean) ProgDTD Random mask training
Builds on: ProgDTD mbt-mean (Minnen joint hyperprior with autoregressive context) double-tail drop
Closest papers by meaning
- COSMIC: Compress Satellite Image Efficiently via Diffusion Compensation (NeurIPS 2024 · similarity 0.86)
- Hazy Low-Quality Satellite Video Restoration Via Learning Optimal Joint Degradation Patterns and Continuous-Scale Super-Resolution Reconstruction (CVPR 2025 · similarity 0.82)
- GLIC: General Format Learned Image Compression (AAAI 2025 · similarity 0.81)
- Coded Computing for Resilient Distributed Computing: A Learning-Theoretic Framework (NeurIPS 2024 · similarity 0.80)
- AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data (ICLR 2025 · similarity 0.80)
- Balanced Rate-Distortion Optimization in Learned Image Compression (CVPR 2025 · similarity 0.80)
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.