Research map / Image generation / Image restoration
Learned Image Compression with Dictionary-based Entropy Model
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 DCAE, a learned image compression entropy model that uses a learnable shared dictionary queried by cross-attention (with multi-scale feature aggregation) to inject training-data priors, achieving state-of-the-art BD-rate over VVC with lower latency than MLIC++.
Problem
Entropy models in learned image compression mainly exploit internal dependencies of latent representations (hyper-prior, autoregressive context) and ignore priors extractable from the external training dataset, limiting rate-distortion efficiency and speed trade-offs.
Approach
Learn a shared dictionary of typical natural-image structures as network parameters (no bit cost, shared by encoder and decoder), and let decoded latent features query it via cross-attention, with multi-scale feature aggregation to make queries more precise.
Key results
- BD-rate over VVC of -17.0% on Kodak, -21.1% on Tecnick, and -19.7% on CLIC, outperforming MLIC++ (-15.1%, -18.6%, -16.9%).
- Total latency of 193 ms on Kodak versus about 772 ms for MLIC++, with GFLOPs comparable to other methods.
- Ablation: DCA alone improves BD-rate from -4.20% to -7.28% with 10 ms extra latency; MSFA further improves to -8.50%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the dictionary-based cross-attention entropy module (DCAE) (ablation) | Baseline learned image compression model · Kodak | BD-rate: -4.20% -> -7.28% | Helped |
| Add multi-scale feature aggregation (MSFA) to DCA (ablation) | DCAE model · Kodak | BD-rate: -7.28% -> -8.50% | Helped |
| Increase dictionary size from 64 to 128 entries (ablation) | DCA model · Kodak | BD-rate: -6.84% -> -7.28% | Helped |
| Increase dictionary size from 128 to 192 or 256 entries (ablation) | DCA model · Kodak | BD-rate: -7.28% -> -7.26%; -7.28% -> -6.92% | Hurt |
| Increase MSFA convolutional-layer count from m=0 to m=3 (ablation) | DCA model · Kodak | BD-rate: -7.28% -> -7.62% -> -8.04% -> -8.50% | Helped |
| Increase MSFA convolutional-layer count from m=3 to m=4 (ablation) | DCA model · Kodak | BD-rate: -8.50% -> -8.36% | Hurt |
| Replace DCA with a global token (ablation) | Baseline learned image compression model · Kodak | BD-rate: -7.28% -> -6.59% | Hurt |
| Use the proposed full model instead of MLIC++ | Learned image compression · Kodak | BD-rate over VVC: -15.1% -> -17.0% | Helped |
| Use the proposed full model instead of MLIC++ | Learned image compression · Tecnick | BD-rate over VVC: -18.6% -> -21.1% | Helped |
| Use the proposed full model instead of MLIC++ | Learned image compression · CLIC | BD-rate over VVC: -16.9% -> -19.7% | Helped |
| Use the proposed full model instead of MLIC++ | Learned image compression · Kodak | Latency: 772 -> 193 | Helped |
Surprising findings
- A global-token approach from prior work gives smaller gains (-6.59%) than the dictionary at similar latency, and the dictionary is shared without extra bitrate.
- Performance saturates beyond 128 dictionary entries; more entries do not help.
Limitations
- Parameter count (119.2M) is larger than several baselines such as CCA (64.9M) and ELIC (41.9M).
- Gains are evaluated only on image compression benchmarks with MSE distortion; generalization beyond these datasets is not shown.
- Dictionary size and multi-scale depth were tuned on a smaller ablation model rather than the final model.
Experimental setup
Datasets: OpenImages
Benchmarks: Kodak Tecnick CLIC professional validation
Metrics: BD-rate over VVC PSNR latency (ms) GFLOPs parameter count
Baselines: STF WACNN ELIC M2T MT TCM MLIC+ MLIC++ FTIC WeConvene CCA VVC (VTM-12.1)
Builds on: hyper-prior channel-wise autoregressive entropy model Swin transformer ELIC latent residual prediction
Closest papers by meaning
- Towards Efficient Image Compression Without Autoregressive Models (NeurIPS 2023 · similarity 0.90)
- GLIC: General Format Learned Image Compression (AAAI 2025 · similarity 0.87)
- Linear Attention Modeling for Learned Image Compression (CVPR 2025 · similarity 0.86)
- Diversify, Contextualize, and Adapt: Efficient Entropy Modeling for Neural Image Codec (NeurIPS 2024 · similarity 0.86)
- Good, Cheap, and Fast: Overfitted Image Compression with Wasserstein Distortion (CVPR 2025 · similarity 0.85)
- Conditional Latent Coding with Learnable Synthesized Reference for Deep Image Compression (AAAI 2025 · similarity 0.85)
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.