Research map / Architectures and efficiency / CNN design
Another Way to the Top: Exploit Contextual Clustering in Learned Image Coding
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 CLIC, a learned image codec whose transform is built from contextual clustering blocks plus local spatial/channel attention, together with GuidedPQF for compensating quantization error at the start of decoding.
Problem
Learned image compression transforms usually rely on convolutions or window self-attention, which are costly in parameters/MACs and have limited receptive fields; quantization errors also propagate through the decoder and degrade reconstruction.
Approach
Treat image pixels as an unordered point set, cluster them by cosine similarity into a few centers to aggregate features over the whole image, reorder back to spatial positions, and apply local spatial and channel attention; additionally, predict post-quantization error corrections guided by side coefficients sent in the bitstream.
Key results
- With MSE optimization, CLIC achieves -9.83% BD-Rate on Kodak, -11.00% on Tecnick and -10.05% on CLIC 2022 against VVC intra anchor.
- With MS-SSIM optimization, CLIC achieves -51.71% BD-Rate on Kodak, the only method exceeding 50% savings.
- CLIC uses 721K MACs per pixel versus 833K for ELIC and 1781K for Mixed, with 78.8M parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CLIC instead of VVC | CLIC · Kodak, MSE | BD-Rate reduction: -9.83% | Helped |
| use CLIC instead of VVC | CLIC · Tecnick, MSE | BD-Rate reduction: -11.00% | Helped |
| use CLIC instead of VVC | CLIC · CLIC 2022, MSE | BD-Rate reduction: -10.05% | Helped |
| use CLIC instead of VVC | CLIC · Kodak, MS-SSIM | BD-Rate reduction: -51.71% | Helped |
| replace clustering with convolution (ablation) | CLIC · Kodak | coding performance | Hurt |
| replace clustering with convolution (ablation) | CLIC · CLIC | MACs: +25.02% | Hurt |
| replace clustering with W-MHSA (ablation) | CLIC · Kodak | coding performance | No clear effect |
| replace clustering with W-MHSA (ablation) | CLIC · CLIC | encoding and decoding time: 67.35%; 127.49% | Hurt |
| remove attention units (ablation) | CLIC · Kodak | coding performance | Hurt |
| replace channel and spatial attention with convolution (ablation) | CLIC · CLIC | parameters: +101.27% | Hurt |
| remove positional encoding (ablation) | CLIC · Kodak | coding performance | Hurt |
| remove checkerboard pattern in CCB (ablation) | CLIC · Kodak | coding performance | Hurt |
Surprising findings
- Replacing the clustering token mixer with a 3x3 convolution lowers RD performance while raising MACs by 25%, indicating clustering is both more efficient and more effective than convolution here.
- Using Euclidean distance instead of cosine similarity for clustering yields poor results.
Limitations
- MSE-optimized performance trails Mixed on Kodak and Tecnick, attributed partly to smaller training set (21,600 vs 300,000 images).
- Parameter count (78.8M) is higher than several baselines, and encoding time is higher than some channel-conditional methods.
- GuidedPQF requires transmitting filter coefficients in the bitstream (four bits each).
Experimental setup
Datasets: Flickr2W LIU4K
Benchmarks: Kodak Tecnick CLIC 2022
Metrics: BD-Rate PSNR MS-SSIM MACs per pixel encoding/decoding time parameters
Baselines: VVC (VTM-18.0) Cheng2020 Entroformer NeuralSyntax QResVAE GLLMM STF WACNN ELIC Contextformer Mixed
Builds on: Metaformer SCCTX Contextual clustering (Ma et al. 2023b) hyperprior model (Ballé et al. 2018) PQF (Fu et al. 2023a)
Closest papers by meaning
- Linear Attention Modeling for Learned Image Compression (CVPR 2025 · similarity 0.87)
- FLAVC: Learned Video Compression with Feature Level Attention (CVPR 2025 · similarity 0.84)
- Causal Context Adjustment Loss for Learned Image Compression (NeurIPS 2024 · similarity 0.83)
- Conditional Latent Coding with Learnable Synthesized Reference for Deep Image Compression (AAAI 2025 · similarity 0.83)
- MambaIC: State Space Models for High-Performance Learned Image Compression (CVPR 2025 · similarity 0.83)
- Frequency-Aware Transformer for Learned Image Compression (ICLR 2024 · 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.