Research map / Architectures and efficiency / Attention and transformers
Learning a Sparse Transformer Network for Effective Image Deraining
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace dense self-attention with learnable top-k sparse attention (ablation) | DRSformer · deraining benchmarks | PSNR | Helped |
| use a learnable interval for k rather than a fixed k value (ablation) | DRSformer TKSA · Rain200H | PSNR: 32.18 dB | Helped |
| make k too small (ablation) | DRSformer TKSA · Rain200H | PSNR | Hurt |
| increase k enough to include irrelevant attention features (ablation) | DRSformer TKSA · Rain200H | PSNR | Hurt |
| replace conventional feed-forward network with MSFN (ablation) | DRSformer · Rain200H | PSNR: 31.84 -> 32.18 | Helped |
| replace Dconv feed-forward network with MSFN (ablation) | DRSformer · Rain200H | PSNR: 31.88 -> 32.18 | Helped |
| replace gated-Dconv feed-forward network with MSFN (ablation) | DRSformer · Rain200H | PSNR: +0.21 dB | Helped |
| add MEFC-1 only with 8 experts (ablation) | DRSformer · Rain200H | PSNR: 32.03 -> 32.01 | Hurt |
| add MEFC-2 only with 8 experts (ablation) | DRSformer · Rain200H | PSNR: 32.03 -> 32.07 | Helped |
| use both MEFC modules with 8 experts instead of one expert (ablation) | DRSformer · Rain200H | PSNR: 32.06 -> 32.18 | Helped |
| increase the experts in both MEFC modules from 4 to 8 (ablation) | DRSformer · Rain200H | PSNR: 32.14 -> 32.18 | Helped |
| use DRSformer instead of Restormer | DRSformer · Rain200L | PSNR: 40.99 -> 41.23 | Helped |
Closest papers by meaning
- Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration (CVPR 2024 · similarity 0.89)
- Rethinking Multi-Scale Representations in Deep Deraining Transformer (AAAI 2024 · similarity 0.88)
- Attention-Only Transformers via Unrolled Subspace Denoising (ICML 2025 · similarity 0.86)
- CLEAR: Conv-Like Linearization Revs Pre-Trained Diffusion Transformers Up (NeurIPS 2025 · similarity 0.86)
- Sharing Key Semantics in Transformer Makes Efficient Image Restoration (NeurIPS 2024 · similarity 0.86)
- Dual-Scale Transformer for Large-Scale Single-Pixel Imaging (CVPR 2024 · 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.