Research map / Image generation / Image restoration
Rethinking Reconstruction and Denoising in the Dark: New Perspective, General Architecture and Beyond
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 CANS/CANS++, a shared-backbone "backbone-head" network with chromaticity-and-noise perception and a dual-head cross-attention interaction module that jointly performs low-light RAW denoising and reconstruction.
Problem
Low-light RAW image processing typically handles denoising and reconstruction in cascaded or separate pipelines, which neglects task associativity and yields suboptimal joint performance and high computational cost.
Approach
Treat denoising and reconstruction as tasks sharing a U-shaped backbone with lightweight task-specific heads (denoising and reconstruction), trained jointly with task-specific losses; the backbone includes a Global Chromatic Perceptor and Refined Detail Extractor, and CANS++ adds a cross-attention dual-head interaction module, with heads discarded at test time for the denoising/reconstruction bran
Key results
- CANS++ reaches 30.856 dB PSNR on the Sony reconstruction benchmark versus 30.616 dB for DNF, and 36.5590 dB on MCR versus the prior best reported in Table 2 (about 0.972 dB gain claimed on MCR).
- On Sony denoising, CANS++ averages 40.141 dB PSNR across exposure ratios versus 39.531 dB for DNF.
- CANS++ uses 40.66 GFLOPs and 4.93M parameters for reconstruction, lower FLOPs than DNF's 59.96 G at slightly more parameters than DNF's 2.83M is not the case; it has fewer FLOPs than SID (48.39 G) and DNF (59.96 G).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the dual-head interaction module (CANS++ instead of CANS) (ablation) | CANS · Sony reconstruction | PSNR: 30.658 -> 30.856 | Helped |
| Remove the GCP module (ablation) | CANS++ reconstruction backbone · Sony reconstruction | PSNR: -0.977 dB | Hurt |
| Remove the RDE module (ablation) | CANS++ reconstruction backbone · Sony reconstruction | PSNR: 1.847 dB drop | Hurt |
| Replace the denoising module with the CID Block (ablation) | CANS++ reconstruction backbone · Sony reconstruction | PSNR: -0.253 dB PSNR | Hurt |
| Replace the color-correction module with Restormer Attention (ablation) | CANS++ reconstruction backbone · Sony reconstruction | PSNR: -0.142 dB PSNR | Hurt |
| Set adaptive pooling size to 1 instead of 2 (ablation) | CANS++ · Sony reconstruction | PSNR: 30.8559 -> 30.6851 | Hurt |
| Set adaptive pooling size to 4 instead of 2 (ablation) | CANS++ · Sony reconstruction | PSNR: 30.8559 -> 30.6164 | Hurt |
| Set adaptive pooling size to 8 instead of 2 (ablation) | CANS++ · Sony reconstruction | PSNR: 30.8559 -> 30.5766 | Hurt |
| Set adaptive pooling size to 16 instead of 2 (ablation) | CANS++ · Sony reconstruction | PSNR: 30.8559 -> 30.5617 | Hurt |
| Use the original pooling setting instead of adaptive size 2 (ablation) | CANS++ · Sony reconstruction | PSNR: 30.8559 -> 29.8582 | Hurt |
| Use the backbone-head multi-task strategy instead of the single-task strategy (ablation) | CANS · Sony reconstruction | PSNR: 30.460 -> 30.658 | Helped |
| Add cross-attention interaction between the task heads (ablation) | CANS · Sony reconstruction | PSNR: 30.658 -> 30.856 | Helped |
Surprising findings
- Adding the GCP module to the encoder decreased performance, which the authors attribute to the encoder focusing on denoising and the decoder on color correction.
Limitations
- Denoising-head parameters and reconstruction-head parameters are discarded at test time, but the interaction module is only described for CANS++ and adds extra parameters during training.
- Evaluation relies on existing benchmarks (SID, MCR, LRD) and a limited set of smartphone scenes without quantitative metrics.
Experimental setup
Datasets: SID Sony subset SID Fuji subset MCR LRD
Benchmarks: SID Sony SID Fuji MCR LRD
Metrics: PSNR SSIM LPIPS DISTS FSIM FLOPs parameters
Baselines: SID DID SGN LLP RRT EEM MCR BGL DNF ELD LRD
Builds on: Restormer CID block (DNF)
Closest papers by meaning
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.87)
- Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model (CVPR 2025 · similarity 0.86)
- How Learning by Reconstruction Produces Uninformative Features For Perception (ICML 2024 · similarity 0.86)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.86)
- Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable (CVPR 2025 · similarity 0.86)
- Denoising Task Routing for Diffusion Models (ICLR 2024 · similarity 0.86)
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.