SPIDEN: deep Spiking Neural Networks for efficient image denoising

Area of Science:

• Artificial Intelligence
• Computer Vision
• Computational Neuroscience

Background:

• Deep Convolutional Neural Networks (DCNNs) excel at image restoration but lack computational efficiency.
• Image denoising is a challenging regression task requiring high precision pixel value prediction.
• Training deep networks in the spiking domain for complex tasks like denoising is difficult.

Purpose of the Study:

• Investigate Spiking Neural Networks (SNNs) for image denoising.
• Achieve DCNN-level performance with reduced computational cost.
• Analyze the trade-off between conversion error and activation sparsity in SNNs for denoising.

Main Methods:

• Formal analysis of Integrate and Fire (IF) neuron information processing.
• Development of the first SNN-based image denoising solution.
• Direct training of SNNs in the spike domain using surrogate gradient learning and backpropagation through time.

Main Results:

• Proposed SNN achieves 30.18 dB SNR on Set12, only 0.25 dB below DCNNs.
• Competitive denoising performance is achieved with low latency (few timesteps) and high sparsity.
• SNNs show potential for energy efficiency, with a 20% reduction possible by increasing network size.

Conclusions:

• SNNs offer a computationally efficient alternative for image denoising.
• The proposed SNN training method enables high performance in the spiking domain.
• SNNs present a promising direction for energy-efficient AI in image restoration tasks.