Research map / Image generation / Image restoration
Make Lossy Compression Meaningful for Low-Light Images
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 joint end-to-end network that compresses and enhances low-light images in one pass, using an SNR-aware branch to guide enhancement in the compressed latent domain, outperforming sequential compress-then-enhance and enhance-then-compress pipelines.
Problem
Lossy compression of low-light images with enhancement is usually done sequentially (compress before enhance or enhance before compress), which causes error accumulation, loss of texture detail after compression, and high computational cost.
Approach
A two-branch end-to-end architecture: a main enhancement branch operating on compressed-domain latents, fused with local/non-local features from an SNR-aware branch (SNR map from a no-learning denoising operation), so the decoder reconstructs the enhanced image directly from enhanced latents.
Key results
- The joint solution achieves rate-distortion gains in PSNR and MS-SSIM over six sequential CbE/EbC combinations of Xu2022, TCM, VTM and Cheng on SID, SDSD-indoor, SDSD-outdoor and SMID.
- The joint solution has much lower computational cost and fewer parameters than sequential pipelines, since enhancement and compression share feature extractors/decoder.
- Removing the SNR-aware branch ('Joint Guidance without SNR-aware') causes a large RD performance drop.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| jointly perform low-light image compression and enhancement instead of sequential CbE or EbC processing | proposed joint solution · SID | PSNR | Helped |
| jointly perform low-light image compression and enhancement instead of sequential CbE or EbC processing | proposed joint solution · SDSD-indoor, SDSD-outdoor, and SMID | PSNR | Helped |
| jointly perform low-light image compression and enhancement instead of sequential CbE or EbC processing | proposed joint solution · SID, SDSD-indoor, SDSD-outdoor, and SMID | MS-SSIM | Helped |
| share the feature extractor and decoder for compression and low-light enhancement | proposed joint solution · RGB image at 960 × 512 resolution | computational cost | Helped |
| share the feature extractor and decoder for compression and low-light enhancement | proposed joint solution · model size | number of model parameters | Helped |
| add an SNR-aware branch to extract local and non-local information (ablation) | joint low-light image compression and enhancement · SID | rate-distortion performance | Helped |
| add an SNR-aware branch to extract local and non-local information (ablation) | joint low-light image compression and enhancement · SDSD-indoor and SDSD-outdoor | rate-distortion performance | Helped |
| add a teacher guidance branch during training (ablation) | joint low-light image compression and enhancement · SID, SDSD-indoor, and SDSD-outdoor | rate-distortion performance | Hurt |
| add a teacher guidance branch during training (ablation) | joint low-light image compression and enhancement · training | computational cost | Hurt |
Surprising findings
- Jointly training compression and enhancement from the start caused convergence problems, requiring two-stage training.
Limitations
- Two-stage training required; joint training from scratch failed to converge.
- Relies on a no-learning SNR map computed from simple local averaging and grayscale conversion.
- Comparison with only one or two enhancement methods (Xu2022, SMG) and no existing joint baseline.
Experimental setup
Datasets: Flickr2W SID SDSD-indoor SDSD-outdoor SMID
Benchmarks: SID SDSD-indoor SDSD-outdoor SMID
Metrics: PSNR MS-SSIM BPP computational cost (FLOPs) model parameters
Baselines: Cheng2020-anchor TCM VVC (VTM 12.1) Xu2022 SMG sequential CbE/EbC pipelines
Builds on: Cheng2020 anchor compression model CompressAI hyperprior entropy model
Closest papers by meaning
- Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach (ICML 2024 · similarity 0.86)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.85)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.85)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.85)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.85)
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.84)
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.