Research map / Vision-language and multimodal / Cross-modal learning
Asymmetric Hierarchical Difference-aware Interaction Network for Event-guided Motion 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 AHDINet, an asymmetric bidirectional hierarchical network that fuses event and image features differently at low and high encoder levels for event-guided motion deblurring.
Problem
Existing event-based motion deblurring methods fuse image and event features with symmetric or unidirectional interactions at all encoder levels, ignoring that low-level and high-level cross-modal features have different complementary roles.
Approach
Uses an event-assisted edge complement module (Gaussian-blurred event confidence mask plus channel attention) at low encoder levels to inject edges from events into image features, and an image-assisted semantic complement module (channel-spatial attention plus channel-wise injection) at high levels to transfer image semantics to the event branch.
Key results
- On GoPro, AHDINet reaches 37.09 dB PSNR / 0.9820 SSIM, versus 36.70 dB for the best prior event-based method (FAEVD) in the table.
- On cross-dataset DVD (trained on GoPro), AHDINet achieves 34.34 dB PSNR, about 0.4 dB above the best prior event-based method.
- On real-world REBlur, AHDINet achieves 40.85 dB PSNR, 1.87 dB above the second-best method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AHDINet instead of STCNet | Event-guided motion deblurring · GoPro | PSNR: 36.45 -> 37.09 | Helped |
| Use AHDINet instead of STCNet | Event-guided motion deblurring · DVD | PSNR: 33.94 -> 34.34 | Helped |
| Use AHDINet instead of STCNet | Event-guided motion deblurring · REBlur | PSNR: 38.98 -> 40.85 | Helped |
| Add EEC to the baseline without EEC or ISC (ablation) | AHDINet · GoPro | PSNR: 33.40 -> 36.50 | Helped |
| Add ISC to the baseline without EEC or ISC (ablation) | AHDINet · GoPro | PSNR: 33.40 -> 36.08 | Helped |
| Add EEC and ISC to the baseline (ablation) | AHDINet · GoPro | PSNR: 33.40 -> 37.09 | Helped |
| Add DMPC to EEC (ablation) | AHDINet · GoPro | PSNR: 36.14 -> 36.50 | Helped |
| Add GSE to ISC (ablation) | AHDINet · GoPro | PSNR: 35.76 -> 36.08 | Helped |
| Use asymmetric bidirectional multi-level interaction instead of unidirectional interaction (ablation) | AHDINet · GoPro | PSNR: 35.95 -> 37.09 | Helped |
| Use asymmetric instead of symmetric bidirectional multi-level interaction (ablation) | AHDINet · GoPro | PSNR: 36.61 -> 37.09 | Helped |
Limitations
- Real-world evaluation is limited to the REBlur dataset, which is indoor and relatively small.
- Event noise handling relies on a fixed Gaussian blur controller rather than a learned one.
Experimental setup
Datasets: GoPro DVD REBlur
Benchmarks: GoPro DVD REBlur
Metrics: PSNR SSIM
Baselines: RED eSL-Net D2Nets LEBMD EVDI DS-Deblur MADANET+ ERDNet EFNet REFID EIFNet STCNet
Builds on: Charbonnier loss ESIM event simulator MPRNet
Closest papers by meaning
- Motion Deblurring via Spatial-Temporal Collaboration of Frames and Events (AAAI 2024 · similarity 0.89)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · similarity 0.88)
- Frequency-aware Event-based Video Deblurring for Real-World Motion Blur (CVPR 2024 · similarity 0.88)
- Learning Scale-Aware Spatio-temporal Implicit Representation for Event-based Motion Deblurring (ICML 2024 · similarity 0.86)
- LaSe-E2V: Towards Language-guided Semantic-aware Event-to-Video Reconstruction (NeurIPS 2024 · similarity 0.86)
- Diffusion-based Event Generation for High-Quality Image Deblurring (CVPR 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.