Research map / Video / Video segmentation and tracking
FMA-Net: Flow-Guided Dynamic Filtering and Iterative Feature Refinement with Multi-Attention for Joint Video Super-Resolution and Deblurring
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 FMA-Net, a joint video super-resolution and deblurring (VSRDB) network using flow-guided dynamic filtering and iterative multi-attention feature refinement trained with a temporal anchor loss.
Problem
Restoring clean high-resolution videos from blurry low-resolution inputs, jointly solving video super-resolution and deblurring, where sequential cascades propagate errors and standard dynamic filters struggle with large spatio-temporally variant motion.
Approach
Flow-guided dynamic filtering whose kernels are placed along motion trajectories (warped positions) rather than fixed grid neighborhoods, combined with stacked FRMA blocks that iteratively learn multiple self-induced optical flows with occlusion masks, center-oriented and degradation-aware attention, and a temporal anchor loss.
Key results
- On REDS4 for x4 VSRDB, FMA-Net reaches 28.83 dB PSNR / 0.8315 SSIM, 1.03 dB above retrained RVRT (27.80 dB) with 9.6M params.
- On GoPro and YouTube, FMA-Net averages 2.08 dB above retrained RVRT and 1.93 dB above retrained GShiftNet.
- Replacing FGDF with conventional dynamic filtering drops performance substantially at large motion, with FGDF giving up to 1.93 dB gain for motion magnitude >=40.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FMA-Net instead of RVRT* | VSRDB · REDS4 ×4 | PSNR (dB): 27.80 -> 28.83 | Helped |
| use FMA-Net instead of RVRT* | VSRDB · GoPro ×4 | PSNR (dB): 25.99 -> 27.65 | Helped |
| use FMA-Net instead of GShiftNet* | VSRDB · YouTube ×4 | PSNR (dB): 24.44 -> 26.02 | Helped |
| replace conventional dynamic filtering with FGDF (ablation) | FMA-Net, REDS4 ×4 VSRDB, total motion ranges, kd=20 · REDS4 | PSNR (dB): 27.86 -> 28.83 | Helped |
| increase multi-flow-mask pairs from n=1 to n=5 (ablation) | FMA-Net, REDS4 ×4 · REDS4 | PSNR (dB): 28.24 -> 28.60 | Helped |
| replace multi-flow-mask pairs with DCN (#offset=9) (ablation) | FMA-Net, REDS4 ×4 · REDS4 | PSNR (dB): 28.83 -> 28.52 | Hurt |
| add RAFT and TA losses (ablation) | FMA-Net, REDS4 ×4 · REDS4 | PSNR (dB): 28.68 -> 28.83 | Helped |
| add TA loss (ablation) | FMA-Net, REDS4 ×4 · REDS4 | PSNR (dB): 28.73 -> 28.83; 0.1 dB | Helped |
| use two-stage training instead of end-to-end learning (ablation) | FMA-Net, REDS4 ×4 · REDS4 | PSNR (dB): 28.39 -> 28.83; 0.44 dB | Helped |
| replace self-attention + SFT with CO + DA attention (ablation) | FMA-Net, REDS4 ×4 · REDS4 | PSNR (dB): 28.50 -> 28.83; 0.33 dB | Helped |
| replace self-attention + SFT with CO attention + SFT (ablation) | FMA-Net, REDS4 ×4 · REDS4 | PSNR (dB): 28.58 -> 28.83 | Helped |
| replace self-attention + DA attention with CO + DA attention (ablation) | FMA-Net, REDS4 ×4 · REDS4 | PSNR (dB): 28.80 -> 28.83 | Helped |
Surprising findings
- Retrained single-image Restormer* achieves higher REDS4 PSNR (27.29 dB) than retrained BasicVSR++* (27.06 dB), and retrained GShiftNet* is weakest (25.77 dB).
- Sequential cascades of SR and deblurring models perform notably worse than joint VSRDB methods, supporting error propagation as a problem.
Limitations
- Conventional dynamic filtering with fixed neighborhoods struggles with large motion, which FGDF addresses but at added design complexity.
- Runtime (0.427 s) is higher than lightweight baselines such as BasicVSR++ (0.072 s) and Restormer (0.081 s).
Experimental setup
Datasets: REDS GoPro YouTube (40 240fps videos, downsampled)
Benchmarks: REDS4 GoPro YouTube
Metrics: PSNR SSIM tOF parameters runtime
Baselines: HOFFR Restormer* GShiftNet* BasicVSR++* RVRT* SwinIR + Restormer HAT + FFTformer BasicVSR++ + RVRT FTVSR + GShiftNet
Builds on: Dynamic Filter Network Restormer RVRT BasicVSR++ Deformable convolution
Closest papers by meaning
- Exploiting Blurry Representations for Event-guided Video Super-Resolution (AAAI 2026 · similarity 0.88)
- Event-Enhanced Blurry Video Super-Resolution (AAAI 2025 · similarity 0.87)
- Hazy Low-Quality Satellite Video Restoration Via Learning Optimal Joint Degradation Patterns and Continuous-Scale Super-Resolution Reconstruction (CVPR 2025 · similarity 0.86)
- Parameterized Blur Kernel Prior Learning for Local Motion Deblurring (CVPR 2025 · similarity 0.86)
- Blur-aware Spatio-temporal Sparse Transformer for Video Deblurring (CVPR 2024 · similarity 0.85)
- One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 2025 · 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.