Research map / Architectures and efficiency / CNN design
Frequency-Biased Synergistic Design for Image Compression and Compensation
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 Lo-Hi synergistic framework that redesigns JPEG/HEVC/VVC quantization to coarsely quantize low-frequency DCT components (using their std as step size) and compensates them with a regression network plus basis attention block, with a lightweight GAN for high frequencies.
Problem
Compression artifact removal (CAR) in edge-cloud image transmission treats compression and compensation as separate designs, leaving low-frequency redundancy that CNN compensators can already recover well.
Approach
Exploit a frequency bias where CNN compensators restore low-frequency components better than high ones: replace the default quantization step for low-frequency DCT components with their standard deviation (Lo-Hi sensitive quantization), then compensate with a two-step network (regression for low frequencies via a basis attention block, generative for high frequencies).
Key results
- Lo-Hi Uformer reaches up to 47.91% BD-rate reduction over LPIPS on HEVC, versus at most 35.86% for baselines.
- Lo-Hi sensitive quantization achieves up to 27.55% pure bitrate savings over default JPEG.
- Lo-Hi NAFNet yields up to 54.53% BD-rate reduction on COCO mAP with VVC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Lo-Hi quantization and Lo-Hi Uformer instead of the default codec with Uformer | Uformer with JPEG · JPEG | BD-rate (LPIPS): -26.54% -> -43.68% | Helped |
| use Lo-Hi quantization and Lo-Hi Uformer instead of the default codec with Uformer | Uformer with HEVC · HEVC | BD-rate (LPIPS): -34.99% -> -47.91% | Helped |
| use Lo-Hi quantization and Lo-Hi Uformer instead of the default codec with Uformer | Uformer with VVC · VVC | BD-rate (LPIPS): -35.86% -> -45.62% | Helped |
| use Lo-Hi quantization and Lo-Hi Uformer instead of the default codec with SeeSR | SeeSR with JPEG · JPEG | BD-rate (FID): -1.05% -> -27.55% | Helped |
| use Lo-Hi quantization and Lo-Hi Uformer instead of the default codec with SeeSR | SeeSR with HEVC · HEVC | BD-rate (FID): -26.61% -> -39.53% | Helped |
| use Lo-Hi quantization and Lo-Hi Uformer instead of the default codec with SeeSR | SeeSR with VVC · VVC | BD-rate (FID): -14.61% -> -40.04% | Helped |
| use Lo-Hi quantization and Lo-Hi NAFNet instead of default compression with NAFNet | NAFNet with JPEG · JPEG | BD-rate over mAP: -31.22% -> -48.12% | Helped |
| use Lo-Hi quantization and Lo-Hi NAFNet instead of default compression with NAFNet | NAFNet with HEVC · HEVC | BD-rate over mAP: -24.54% -> -41.63% | Helped |
| use Lo-Hi quantization and Lo-Hi NAFNet instead of default compression with NAFNet | NAFNet with VVC · VVC | BD-rate over mAP: -6.32% -> -54.53% | Helped |
| replace default quantization with Lo-Hi sensitive quantization (ablation) | JPEG compression · JPEG | BD-rate over LPIPS: -11.50% | Helped |
| replace default quantization with Lo-Hi sensitive quantization (ablation) | HEVC compression · HEVC | BD-rate over PSNR: 2.51% | Hurt |
| add BAB to the regression model (ablation) | Lo-Hi Uformer · JPEG | BD-rate over PSNR: -25.02% -> -49.43% | Helped |
Surprising findings
- Lo-Hi frameworks trade some PSNR (up to 6.93% BD-rate increase for Lo-Hi Uformer on VVC/HEVC PSNR) for perceptual gains, and the regression model alone performs worse than uncompensated output on high-frequency-removed data.
Limitations
- Diffusion-based Hi models struggle with text rendering in the framework, and the GAN is used instead.
- PSNR fidelity is reduced relative to default compression, reflecting the rate-distortion-perception trade-off.
- Gains shrink for more advanced codecs (VVC) on PBS and BD-rate.
Experimental setup
Datasets: LSDIR COCO ImageNet
Benchmarks: LSDIR COCO
Metrics: BD-rate PSNR LPIPS NIQE FID mAP pure bitrate savings
Baselines: NAFNet Uformer FBCNN CRESNet BSRGAN SeeSR OSEDiff
Builds on: NAFNet Uformer BSRGAN AOT block
Closest papers by meaning
- One-To-Many Network for Visually Pleasing Compression Artifacts Reduction (CVPR 2017 · similarity 0.87)
- Frequency-Aware Transformer for Learned Image Compression (ICLR 2024 · similarity 0.86)
- Causal Context Adjustment Loss for Learned Image Compression (NeurIPS 2024 · similarity 0.85)
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.85)
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.85)
- When MLLMs Meets Compression Distortion: A Coding Paradigm Tailored to MLLMs (ICLR 2026 · 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.