Research map / Architectures and efficiency / Spiking networks
Spiking Meets Attention: Efficient Remote Sensing Image Super-Resolution with Attention Spiking Neural Networks
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 SpikeSR, a spiking neural network with attention spiking blocks (hybrid dimension attention plus deformable similarity attention) for efficient remote sensing image super-resolution, reporting state-of-the-art PSNR on AID, DOTA and DIOR.
Problem
SNNs are underexplored for pixel-level regression such as remote sensing image super-resolution, where they suffer from limited capacity from binary spike information loss and suboptimal membrane potential dynamics, while ANN SR models are computationally heavy.
Approach
A spiking attention block that combines CNN-SNN layers, a hybrid temporal-channel attention to recalibrate membrane potentials/spiking responses, and a deformable similarity attention that exploits global self-similarity of RSIs via patch matching with Gumbel-Softmax and deformable convolution, plus a fusion block for converting spike sequences to continuous values.
Key results
- SpikeSR achieves average PSNR 31.95 dB and SSIM 0.8339 across AID, DOTA and DIOR for x4 SR, the best among compared methods.
- On AID, SpikeSR improves PSNR by 0.08 dB over SwinIR-light; the small SpikeSR-S uses 472K parameters and 15.21G FLOPs versus 897K and 23.56G for SwinIR.
- Ablation on AID-tiny: HDA adds 0.12 dB over the baseline and DSA adds 0.21 dB; combining both yields 28.19 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SpikeSR instead of Omni-SR | Remote sensing image SR · AID | PSNR: 30.89 -> 30.91 | Helped |
| Use SpikeSR instead of Omni-SR | Remote sensing image SR · DOTA | PSNR: 33.94 -> 33.98 | Helped |
| Use SpikeSR instead of Omni-SR | Remote sensing image SR · DIOR | PSNR: 30.89 -> 30.95 | Helped |
| Add HDA in place of separate TA and CA attention (ablation) | SpikeSR on AID-tiny · AID-tiny | PSNR (dB): +0.12 dB | Helped |
| Add DSA (ablation) | SpikeSR on AID-tiny · AID-tiny | PSNR (dB): +0.21 dB | Helped |
| Combine HDA and DSA (ablation) | SpikeSR on AID-tiny · AID-tiny | PSNR (dB): +0.08 dB | Helped |
| Add a multi-level feature pyramid (ablation) | SpikeSR DSA · AID-tiny | PSNR (dB): 0.05 dB | Helped |
| Remove deformable convolution (ablation) | SpikeSR DSA · AID-tiny | PSNR (dB): 0.12 dB | Hurt |
| Set the number of SAGs to 3 rather than the final model's 4 (ablation) | SpikeSR · AID-tiny | SR performance | Helped |
Limitations
- Representational capacity remains improvable.
- Attention mechanisms add computational overhead, limiting efficiency.
Experimental setup
Datasets: AID
Benchmarks: AID DOTA DIOR
Metrics: PSNR SSIM FLOPs #Param
Baselines: SRCNN VDSR EDSR CARN IMDN RFDN LatticeNet SwinIR-light Omni-SR HiT-SR SPIN ESRT
Builds on: leaky integrate-and-fire neuron temporal-channel joint attention Gumbel-Softmax deformable convolution TTST
Closest papers by meaning
- STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks (CVPR 2025 · similarity 0.86)
- ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks (CVPR 2026 · similarity 0.86)
- Finding Visual Saliency in Continuous Spike Stream (AAAI 2024 · similarity 0.85)
- Spatio-Temporal Interactive Learning for Efficient Image Reconstruction of Spiking Cameras (NeurIPS 2024 · similarity 0.85)
- Spiking Vision Transformer with Saccadic Attention (ICLR 2025 · similarity 0.85)
- Advancing Spiking Neural Networks Towards Multiscale Spatiotemporal Interaction Learning (AAAI 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.