Research map / Image generation / Image restoration
Neural Sparse Representation for Image Restoration
Yuchen Fan, Jiahui Yu, Yiqun Mei, Yulun Zhang, Yun Fu, Ding Liu, Thomas S. Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add softmax-based sparse representation to EDSR (S) | EDSR (S) · Super-resolution benchmarks, ×2, ×3, and ×4 | PSNR: 37.99 -> 38.02; 34.37 -> 34.43; 32.09 -> 32.25 (Set5) | Improved |
| Add two sparsity groups with a narrower width while preserving model size | Denoising baseline · Set12, BSD68, Urban100 | PSNR: 32.97 -> 33.00; 30.59 -> 30.63; 30.33 -> 30.36 (noise 15, 25, 25) | Improved |
| Add two sparsity groups and expand model capacity | Denoising baseline · Set12 noise 25 | PSNR: 30.59 -> 30.68 | Improved |
| Add two sparsity groups to the compression-artifact model | Compression-removal baseline · LIVE1 and Classic5, JPEG qualities 10–40 | PSNR: 29.36 -> 29.39; 31.73 -> 31.79; 33.17 -> 33.21; 34.18 -> 34.23 (LIVE1) | Improved |
| Replace softmax sparse constraints with sigmoid gates (ablation) | WDSR · DIV2K validation ×2 | PSNR: 34.87 -> 34.81 | Hurt |
| Replace soft sparsity with Gumbel hard selection (ablation) | WDSR · DIV2K validation ×2 | PSNR: 34.87 -> 34.45 | Hurt |
| Set softmax temperature to 10 rather than 1 (ablation) | WDSR · DIV2K validation ×2 | PSNR: 34.87 -> 34.86 | No clear effect |
| Set softmax temperature to 0.1 rather than 1 (ablation) | WDSR · DIV2K validation ×2 | PSNR: 34.87 -> 34.83 | Hurt |
| Increase cardinality (ablation) | Neural sparse representation · DIV2K validation ×2 | PSNR | Improved |
| Increase sparsity group size at matched model size (ablation) | WDSR · DIV2K validation ×2, 16 blocks | PSNR: 34.76 -> 34.77 -> 34.68 | Hurt |
| Increase sparsity group size at matched FLOPs (ablation) | WDSR · DIV2K validation ×2, 16 blocks | PSNR: 34.76 -> 34.83 -> 34.87 -> 34.89 -> 34.91 | Improved |
Closest papers by meaning
- Fully Trainable and Interpretable Non-Local Sparse Models for Image Restoration (ECCV 2020 · similarity 0.89)
- Revisiting Sparse Convolutional Model for Visual Recognition (NeurIPS 2022 · similarity 0.89)
- Learning Distortion Invariant Representation for Image Restoration From a Causality Perspective (CVPR 2023 · similarity 0.88)
- Inference, Learning and Attention Mechanisms that Exploit and Preserve Sparsity in CNNs (IJCV 2020 · similarity 0.87)
- A General Efficient Hyperparameter-Free Algorithm for Convolutional Sparse Learning (AAAI 2017 · similarity 0.87)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.87)
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.
