DehazeSNN: Spiking Neural Network for Dehazing

• The paper introduces a novel SNN-based dehazing model that combines a U-Net–like structure with the innovative Orthogonal Leaky-Integrate-and-Fire Block (OLIFBlock) to fuse multi-scale features efficiently.
• It leverages bio-inspired spiking neuron dynamics to enhance cross-channel and spatial communication, dramatically reducing parameters and computational cost compared to CNN and transformer alternatives.
• Quantitative results on benchmarks show competitive PSNR and SSIM performance, highlighting its suitability for resource-constrained image restoration applications.

DehazeSNN is a spiking neural network (SNN)-based single image dehazing architecture that integrates U-Net-like multi-scale feature extraction with novel bio-inspired modules to efficiently process hazy images. Designed to overcome the limitations of conventional convolutional neural networks (CNNs) and transformer-based models in capturing both local and long-range dependencies while significantly reducing computational cost, DehazeSNN introduces the Orthogonal Leaky-Integrate-and-Fire Block (OLIFBlock) to enhance cross-channel and spatial communication. The model achieves competitive performance on standard benchmarks with a fraction of the parameter and operation counts of state-of-the-art dehazing architectures (Li et al., 30 Dec 2025).

1. Network Architecture

DehazeSNN adopts a five-stage U-Net–like architecture composed of:

• Shallow feature extraction: A single $3 \times 3$ convolution module maps the input image $I \in \mathbb{R}^{3 \times H \times W}$ to an initial feature map $F \in \mathbb{R}^{C \times H \times W}$.
• Encoder–decoder SNNBlocks: Five scales are processed via encoder and decoder ladders, each built from SNNBlocks. These replace self-attention or large kernel convolutions with the OLIFBlock.
• Skip connections: “SKfusion” implements attention-weighted elementwise addition, offering an efficient alternative to conventional U-Net concatenation.
• Reconstruction: A final $3 \times 3$ convolution restores the dehazed image $J \in \mathbb{R}^{3 \times H \times W}$.

The high-level architecture is:

[Input Image] → Conv3×3 → SNNBlock ×5 levels (encoder ↓, decoder ↑, SKfusion across) → Conv3×3 → [Output]

Key innovations include the SNN-inspired, group-wise spatial accumulation in place of computationally demanding attention layers and the tight coupling of residual MLPs to enhance representational diversity.

2. The Orthogonal Leaky-Integrate-and-Fire (OLIF) Block

OLIFBlock implements two parallel, group-wise Leaky-Integrate-and-Fire computations (horizontal and vertical), fusing multi-channel features and capturing anisotropic dependencies. The mathematical formulation per spatial group $g$ is:

$$\begin{aligned} y_{g+1} &= W x \ U_{g+1} &= \tau U_g (1 - o_g) + y_{g+1} \ o_{g+1} &= H(U_{g+1} - V_{th}) \ r_{g+1} &= \max(U_{g+1}, V_{th}) \end{aligned}$$

where:

• $x$ = input feature map,
• $W$ = depth-wise convolution kernel,
• $U_g$ = membrane potential,
• $\tau$ = leak rate,
• $o_g \in \{0,1\}$ = spike (via Heaviside function $H$),
• $V_{th}$ = threshold,
• $r_{g+1}$ = full-precision output for backpropagation.

OLIFBlock's cross-channel fusion emerges by spatially slicing features and applying LIF accumulation across these groups, creating long-range cross-channel communication with minimal computation.

3. Spiking Neuron Model

DehazeSNN employs a continuous-output variant of the standard Leaky-Integrate-and-Fire neuron model. The classical neuron’s membrane dynamics are:

$\tau \frac{du(t)}{dt} = -u(t) + I(t)$

$$o(t) = \begin{cases} 1,&u(t)\ge V_{th},\ 0,&u(t)<V_{th} \end{cases}$$

$u \leftarrow u_{reset}\quad\text{after spike}$

In OLIFBlock, the time dimension is replaced with the spatial group index $g$. The binary output $o_{g+1}$ acts as a gating signal, and the output for learning is the clipped continuous potential $r_{g+1}$.

4. Training Procedure

DehazeSNN training utilizes the AdamW optimizer with $\beta_1=0.9$, $\beta_2=0.999$. Learning rates are cosine-annealed: general weights from $1\times 10^{-4}$ to $1\times 10^{-6}$; LIF parameters $(\tau, V_{th})$ from $5\times10^{-5}$ to $1\times10^{-6}$. The batch comprises $256 \times 256$ image crops (batch size 5 for Medium, 4 for Large models).

The loss function is a weighted sum of L1 and LPIPS (perceptual) losses:

$$L_{total} = \alpha L_{1} + (1-\alpha) L_{\mathrm{LPIPS}}$$

with optimal performance at $\alpha=0.5$.

Datasets include:

• RESIDE (ITS, OTS, 6K): indoor/outdoor synthetics.
• RS-Haze: real, non-uniform remote sensing haze.

Evaluation metrics are PSNR and SSIM; model cost is assessed via parameter count and multiply-accumulate operations (MACs).

5. Quantitative Results and Complexity

Extensive benchmarks demonstrate the architectural efficiency and competitiveness of DehazeSNN:

Method	PSNR (RESIDE-ITS)	SSIM (RESIDE-ITS)	Params	MACs
DehazeSNN-M	40.10	0.995	2.70M	26.3G
DehazeSNN-L	41.26	0.996	4.75M	37.3G
DehazeFormer	40.05	0.996	25.44M	279.7G
MAXIM-2S	38.11	0.991	14.10M	108G

Additional results show comparable or superior performance on synthetic outdoor (RESIDE-OTS), mixed synthetic (RESIDE-6K), and remote-sensing (RS-Haze) benchmarks. DehazeSNN reduces parameters by a factor of 3–6 and MACs by 4–8 compared to CNN/Transformer SOTA for matching or better dehazing metrics.

6. Ablation Study, Strengths, and Limitations

Ablation analyses clarify the critical contributions of each model component:

• OLIFBlock: Augmenting the U-Net with OLIFBlock yields a PSNR increase from 25.70 to 29.88 ($\Delta=+4.18$ on RESIDE-6K).
• SKfusion: Offers marginal improvement in PSNR (25.69 to 25.70), and when combined with OLIFBlock, further raises overall performance.
• Loss blending: Optimal results are produced with $\alpha=0.5$ (both L1 and LPIPS); pure L1 or pure LPIPS underperform.
• Group number $g$ in OLIF: $g=4$ demonstrates best PSNR/SSIM; values above or below are suboptimal.

Strengths:

• Efficient multi-scale, long-range feature integration via OLIFBlock without heavy attention or deep convolutions.
• Small model size (2.7–4.8M parameters) and low MACs (26–37G) support deployment under strict resource constraints.
• Effective on both synthetic and remote sensing haze benchmarks.

Limitations:

• Slightly trails large transformer models on very large/outdoor datasets (RESIDE-OTS).
• The group number $g$ in OLIFBlock requires empirical tuning for optimality.
• Applicability to other restoration tasks (e.g., denoising, deraining) is a topic of future investigation.

7. Positioning and Potential Extensions

DehazeSNN represents a paradigm shift in image restoration: employing biologically motivated SNN principles (spatial group-wise LIF) in a canonical U-Net backbone, it achieves competitive perceptual quality with dramatically reduced computational requirements. The architecture generalizes to both synthetic and real haze, spanning indoor, outdoor, and remote sensing settings. A plausible implication is the suitability of DehazeSNN for embedded and edge vision systems demanding low-resource inference. Extensions of DehazeSNN to additional image restoration domains remain open for further academic exploration (Li et al., 30 Dec 2025).