Research map / Image generation / Image restoration
Gyro-based Neural Single Image 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 GyroDeblurNet, a gyro-based single-image deblurring network with gyro refinement and gyro deblurring blocks, a camera motion field gyro embedding, and curriculum learning, plus GyroBlur-Synth and GyroBlur-Real datasets.
Problem
Single-image deblurring is ill-posed, and real-world smartphone gyro data contain errors (noise, unknown rotation center, missing translation, moving objects), so existing gyro-based deblurring methods underperform on real blurred images.
Approach
Refine erroneous gyro data using blur cues from the input image via gyro refinement blocks, deblur with a gyro deblurring block that uses deformable convolution offsets, represent complex shakes with a multi-vector camera motion field, and train with a curriculum that gradually adds gyro noise.
Key results
- On GyroBlur-Synth, GyroDeblurNet reaches 27.28 PSNR / 0.7803 SSIM, versus 26.01 PSNR for the best single-image baseline (FFTformer) and 25.49 for the best gyro baseline (EggNet).
- On GyroBlur-Real-S, it achieves NIQE 4.47 and TOPIQ 0.548, the best among compared methods.
- Ablation shows that increasing the camera motion field hyperparameter M from 2 to 8 raises PSNR from 25.71 to 27.28, with little gain at M=16.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add gyro data to deblurring (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 24.90 -> 27.28 | Helped |
| train with erroneous gyro data instead of error-free gyro data (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 24.94 -> 27.28 | Helped |
| add image-feature refinement of gyro features (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 25.47 -> 27.28 | Helped |
| include image features in gyro refinement (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 26.17 -> 27.28 | Helped |
| use image and gyro features, rather than only gyro features, for deformable convolution (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 26.32 -> 27.28 | Helped |
| add curriculum learning (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 26.94 -> 27.28 | Helped |
| increase camera motion field M from 2 to 4 (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 25.71 -> 26.80 | Helped |
| increase camera motion field M from 4 to 8 (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 26.80 -> 27.28 | Helped |
| increase camera motion field M from 8 to 16 (ablation) | GyroDeblurNet · GyroBlur-Synth test | PSNR: 27.28 -> 27.32 | No clear effect |
| increase camera motion field M from 8 to 16 (ablation) | GyroDeblurNet · GyroBlur-Synth test | SSIM: 0.7803 -> 0.7811 | No clear effect |
| use GyroDeblurNet rather than FFTformer | GyroDeblurNet · GyroBlur-Synth test | PSNR: 26.01 -> 27.28 | Helped |
| use GyroDeblurNet rather than EggNet | GyroDeblurNet · GyroBlur-Real-S | NIQE: 5.18 -> 4.47 | Helped |
Surprising findings
- Training with error-free gyro data yields almost no gain over training without any gyro data (24.94 vs 24.90 PSNR), indicating that modeling real gyro errors during training is essential.
Limitations
- Uses a fixed hyperparameter M, which may need to grow with exposure time.
- Does not use the accelerometer, which could provide additional motion information.
- Relatively simple image deblurring backbone compared with recent non-gyro methods may limit deblurring quality.
Experimental setup
Datasets: GyroBlur-Synth 4KRD COCO GyroBlur-Real-S GyroBlur-Real-H
Benchmarks: GyroBlur-Synth GyroBlur-Real-S GyroBlur-Real-H
Metrics: PSNR SSIM NIQE TOPIQ
Baselines: MPRNet NAFNet AdaRevD Uformer Stripformer FFTformer Vasu et al. non-blind deblurring Nan et al. non-blind deblurring DeepGyro EggNet INformer
Builds on: NAFNet RSBlur EggNet DeepGyro U-Net
Closest papers by meaning
- GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring (NeurIPS 2024 · similarity 0.84)
- Spike-guided Motion Deblurring with Unknown Modal Spatiotemporal Alignment (CVPR 2024 · similarity 0.84)
- DynaMoDe-NeRF: Motion-aware Deblurring Neural Radiance Field for Dynamic Scenes (CVPR 2025 · similarity 0.84)
- Exploiting Deblurring Networks for Radiance Fields (CVPR 2025 · similarity 0.84)
- MoE-Gyro: Self-Supervised Over-Range Reconstruction and Denoising for MEMS Gyroscopes (NeurIPS 2025 · similarity 0.83)
- Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.83)
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.