| name | current-injection-spiking-neural-network-image-fusion |
| version | 1.0.0 |
| description | Current Injection Spiking Neural Network (CIS-Fuse) for energy-efficient infrared and visible image fusion using membrane-potential level cross-modal integration |
| tags | ["spiking-neural-networks","image-fusion","neuromorphic-computing","energy-efficiency","computer-vision"] |
| trigger_words | ["CIS-Fuse","current injection spiking","infrared visible fusion","membrane potential fusion","SNN image fusion"] |
Current Injection Spiking Neural Network for Infrared and Visible Image Fusion
Overview
This skill implements the Current Injection Spiking (CIS) Neural Network architecture, specifically CIS-Fuse, for efficient infrared and visible image fusion (IVIF). The key innovation is performing cross-modal fusion directly at the membrane-potential level rather than at the spike level, preserving subthreshold responses that would otherwise be lost in binary spike communication.
Key Concepts
Problem Statement
- Traditional SNNs communicate through sparse binary spikes, which can discard complementary cues that remain below the firing threshold
- Cross-modal fusion requires fine-grained responses from both modalities (infrared and visible)
- Direct application of SNNs to IVIF creates tension between energy efficiency and fusion quality
Core Innovation: Current Injection Spiking (CIS) Operator
- Injects one modality as a gated auxiliary current into the driving neuron of the other modality
- Integration occurs at the membrane-potential level before spike firing
- Per-channel learnable injection strength adaptively regulates modulation magnitude
- Preserves subthreshold responses that contain complementary information
Architecture Components
- Dual-Branch Architecture: Asymmetric stacking depths with clear functional specialization
- Bidirectional Cross-Modal Fusion (BCMF) Module: Built on CIS operators for bidirectional information flow
- Membrane Potential Integration: Cross-modal fusion at pre-spike integration stage
Performance Benefits
- Energy Efficiency: ~10x lower inference energy compared to similarly-sized ANN-based methods (e.g., DCEvo)
- Fusion Quality: Achieves state-of-the-art results on par with ANN-based methods
- Parameter Efficiency: Reduced model complexity while maintaining performance
- Downstream Performance: Improved results on detection and segmentation tasks
Implementation Guidelines
When to Use
- Energy-constrained edge devices requiring real-time image fusion
- Applications needing both high fusion quality and low power consumption
- Multi-modal sensing scenarios where complementary information must be preserved
- Neuromorphic computing platforms
Architecture Design
class CurrentInjectionSpiking(nn.Module):
def __init__(self, channels):
super().__init__()
self.injection_strength = nn.Parameter(torch.ones(channels))
def forward(self, driving_neuron_potential, auxiliary_modality):
gated_auxiliary = self.injection_strength * auxiliary_modality
integrated_potential = driving_neuron_potential + gated_auxiliary
spike = generate_spike(integrated_potential)
return spike, integrated_potential
Training Considerations
- Use contrastive learning or supervised fusion objectives
- Optimize injection strength parameters end-to-end
- Consider asymmetric branch depths for functional specialization
- Validate on multiple IVIF benchmarks (TNO, RoadScene, etc.)
Evaluation Metrics
Primary Metrics
- Fusion Quality: PSNR, SSIM, MS-SSIM, VIF
- Energy Consumption: Total inference energy (pJ or nJ)
- Parameter Count: Model size comparison
- Downstream Tasks: Detection mAP, segmentation IoU
Benchmarks
- IVIF Datasets: TNO, RoadScene, FLIR, M3FD
- Comparison Baselines: DCEvo, FusionDN, U2Fusion, RFN-Nest
- Hardware Platforms: Loihi, TrueNorth, SpiNNaker
Research Impact
This methodology bridges the gap between energy-efficient neuromorphic computing and high-quality multi-modal fusion, demonstrating that SNNs can achieve competitive performance while maintaining their inherent energy advantages. The membrane-potential level integration principle can be extended to other multi-modal fusion scenarios beyond infrared-visible pairs.
References
- arXiv: 2607.19879
- Authors: Rui Zhao, Zhuoyuan Li, Wenrui Li, Yanchen Dong, Yajing Zheng, Giuseppe Valenzise, Weisi Lin
- Keywords: Spiking Neural Networks, Image Fusion, Infrared-Visible Fusion, Membrane Potential, Energy Efficiency, Neuromorphic Computing
Activation Examples
- "Implement CIS-Fuse for thermal and visible image fusion"
- "Design energy-efficient SNN for multi-modal fusion using current injection"
- "Apply membrane potential integration for cross-modal fusion"
- "Compare CIS-Fuse with traditional ANN-based fusion methods"