Research map / Architectures and efficiency / Attention and transformers
Efficient Lightweight Image Denoising with Triple Attention Transformer
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 LIDFormer, a lightweight Transformer for image denoising that applies double Haar DWT to shrink feature resolution 16x, plus CPFR feature reuse and TMDTA triple-direction attention, reaching ~2.8 GFLOPs with competitive SIDD/DND PSNR.
Problem
Existing Transformer-based image denoisers have large model sizes and high computational cost, making them unsuitable for resource-constrained devices.
Approach
Losslessly compress the input to a low-resolution frequency space via double DWT so that attention and feature processing run at 1/16 resolution; restore shallow detail with complementary periodic feature reuse (learnable channel-attention weighting), and replace full-pixel self-attention with horizontal, vertical and channel transposed attention (TMDTA).
Key results
- LIDFormer reaches 39.62 dB PSNR / 0.9575 SSIM on SIDD and 39.76 dB on DND at 2.83 GFLOPs (256x256) with 2.72M params.
- Double DWT reduces Restormer from 140 GFLOPs to 8.75 GFLOPs, and the full LIDFormer design reaches 2.83 GFLOPs with 39.62 dB PSNR in the ablation.
- ADFNet is slightly better in PSNR on SIDD (39.63 dB) but uses 117.32 GFLOPs, more than 40 times LIDFormer's cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply double DWT feature lightweighting (ablation) | Restormer · SIDD | PSNR: 40.02 -> 39.55 | Hurt |
| apply double DWT feature lightweighting (ablation) | Restormer · SIDD | GFLOPs: 140 -> 8.75 | Helped |
| add CPFR (ablation) | DWT-based Restormer · SIDD | PSNR: 39.55 -> 39.55 | No clear effect |
| add CPFR (ablation) | DWT-based Restormer · SIDD | GFLOPs: 8.75 -> 2.82 | Helped |
| add CACA (ablation) | DWT + CPFR Restormer · SIDD | PSNR: 39.55 -> 39.58 | Helped |
| add TMDTA (ablation) | DWT + CPFR + CACA Restormer · SIDD | PSNR: 39.58 -> 39.62 | Helped |
| use LIDFormer instead of ADFNet | real-world image denoising · SIDD | PSNR: 39.63 vs. 39.62 | Hurt |
| use LIDFormer instead of ADFNet | real-world image denoising · SIDD | GFLOPs: 117.32 vs. 2.83 | Helped |
Surprising findings
- Applying the DWT lightweighting directly to CBDNet collapses performance (27.68 dB on SIDD), showing that DWT compression alone does not guarantee performance.
Limitations
- Performance is slightly below ADFNet on SIDD and the DWT lightweighting alone can severely hurt some architectures such as CBDNet.
- Evaluation is limited to SIDD and DND real-world denoising benchmarks.
Experimental setup
Datasets: SIDD
Benchmarks: SIDD DND
Metrics: PSNR SSIM GFLOPs parameters
Baselines: DnCNN FFDNet CBDNet RIDNet VDN DANet DeamNet InvDN Thunder ADFNet Restormer NAFNet
Builds on: Restormer Multi-Dconv Head Transposed Attention Discrete Wavelet Transform NAFNet
Closest papers by meaning
- Attention-Only Transformers via Unrolled Subspace Denoising (ICML 2025 · similarity 0.84)
- Pseudo-Siamese Blind-spot Transformers for Self-Supervised Real-World Denoising (NeurIPS 2024 · similarity 0.84)
- Xformer: Hybrid X-Shaped Transformer for Image Denoising (ICLR 2024 · similarity 0.83)
- Dual-Scale Transformer for Large-Scale Single-Pixel Imaging (CVPR 2024 · similarity 0.83)
- Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration (CVPR 2024 · similarity 0.82)
- Transformer Meets Twicing: Harnessing Unattended Residual Information (ICLR 2025 · similarity 0.82)
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.