PaperlayerView exact revision on arXiv

Source-linked AI summary

U-Net-Like Spiking Neural Networks for Single Image Dehazing

Huibin Li, Haoran Liu, Mingzhe Liu, Yulong Xiao, Peng Li, Guibin Zan

arXiv:2512.23950v1cs.CV

TL;DR

Image dehazing seeks to restore image clarity obscured by atmospheric haze, while CNNs and Transformers present different challenges in handling local and long-range information efficiently. DehazeSNN combines a U-Net-like architecture with spiking neural networks and OLIFBlock; the paper reports competitive benchmark performance with a small model and reduced computational cost.

  • Problem

    Image dehazing must recover visibility and image quality from haze, while addressing the limitations of CNNs in long-range dependency modeling and the computational demands of Transformers.

    Source: §I · pp. 1–2

  • Method

    DehazeSNN combines a U-Net-like multi-scale architecture with spiking neural networks and an OLIFBlock for cross-channel communication.

    Source: §I · p. 2

  • Results

    DehazeSNN performs favorably against state-of-the-art methods on three benchmarks while maintaining a small model size and fewer MACs.

    Source: §I · p. 2

  • Takeaways & Limitations

    The reported performance and model size indicate potential for DehazeSNN in computationally sensitive dehazing scenarios.

    Source: §V · p. 7

Abstract

from arXiv · show

Image dehazing is a critical challenge in computer vision, essential for enhancing image clarity in hazy conditions. Traditional methods often rely on atmospheric scattering models, while recent deep learning techniques, specifically Convolutional Neural Networks (CNNs) and Transformers, have improved performance by effectively analyzing image features. However, CNNs struggle with long-range dependencies, and Transformers demand significant computational resources. To address these limitations, we propose DehazeSNN, an innovative architecture that integrates a U-Net-like design with Spiking Neural Networks (SNNs). DehazeSNN captures multi-scale image features while efficiently managing local and long-range dependencies. The introduction of the Orthogonal Leaky-Integrate-and-Fire Block (OLIFBlock) enhances cross-channel communication, resulting in superior dehazing performance with reduced computational burden. Our extensive experiments show that DehazeSNN is highly competitive to state-of-the-art methods on benchmark datasets, delivering high-quality haze-free images with a smaller model size and less multiply-accumulate operations. The proposed dehazing method is publicly available at https://github.com/HaoranLiu507/DehazeSNN.

I. INTRODUCTION

DehazeSNN combines a U-Net-like design with spiking neural networks to address local- and long-range-feature processing in image dehazing. The paper reports competitive benchmark performance with a small model and reduced computational cost.

I. INTRODUCTION

  • The paper motivates dehazing as a way to restore visibility and image quality degraded by atmospheric haze, including for remote sensing, photography, and computer vision.Haze can obscure details and distort color accuracy.Source: §I · p. 1
  • CNNs capture local image features and patterns, while Transformers model global dependencies via self-attention; their long-range capability entails computationally intensive cross-attention and typically more parameters and MACs.Source: §I · pp. 1–2
  • DehazeSNN combines a U-Net-like architecture with SNNs to process multi-scale features and manage local patterns and long-range dependencies.The U-Net-like design handles features at different scales, while OLIFBlock facilitates cross-channel communication.Source: §I · p. 2
  • The authors present OLIFBlock as an SNN-based approach to long-term dependencies that reduces computational burden and is the first SNN application to image dehazing.The introduction also attributes computational efficiency to SNNs.Source: §I · p. 2
  • DehazeSNN performs favorably against state-of-the-art methods on three benchmarks while maintaining a small model size and fewer MACs.Fig. 1 compares methods on RESIDE-ITS, with circle size proportional to parameter count; the paper also reports smaller model size and MACs than competitors.Source: Figure 1 · p. 1, §I · p. 2

A. Image Dehazing

Image dehazing has moved from physical-model and prior-based approaches toward deep learning architectures. The section also describes efforts to improve dehazing capability and learning while noting the associated computational or design costs.

A. Image Dehazing

  • Traditional single-image methods use physical priors such as Dark Channel Prior, whereas end-to-end approaches can outperform physical-model methods.Source: A. Image Dehazing · p. 2
  • Dehazing backbones expanded from CNNs and GANs to Transformers, with ViT and shifted-window Swin models improving reported performance.Source: A. Image Dehazing · p. 2
  • MixDehazeNet uses multi-scale parallel large kernels for uneven haze, while attention mechanisms or large kernels can increase model size, parameters, and MACs.Source: A. Image Dehazing · p. 2
  • Orthogonal decoupling, semi-supervision, and joint depth estimation and dehazing address learning challenges but add architectural or process complexity.Source: A. Image Dehazing · p. 2

B. Spiking Neural Networks

Spiking neural networks encode information with discrete spikes and can reduce computational costs relative to conventional artificial neural networks. The methodology section introduces DehazeSNN and outlines its architecture and core components.

B. Spiking Neural Networks

  • SNNs encode information through spike timing rather than continuous activations, and accumulating incoming spike weights can reduce computational costs compared with conventional ANNs.The passage also connects SNNs to neuromorphic hardware and the von Neumann bottleneck.Source: B. Spiking Neural Networks · p. 3
  • Training SNNs is challenging because spike generators are non-continuous, motivating approaches such as ANN-to-SNN conversion, modified backpropagation, and continuous spike approximations.The passage notes applications including image classification, segmentation, and robotic control.Source: B. Spiking Neural Networks · p. 3

III. METHODOLOGY

  • The methodology presents DehazeSNN through an overview of its architecture, the SNNBlock and OLIFBlock, and SKfusion-based skip connections.Source: §III · p. 3

A. DehazeSNN Architecture

DehazeSNN is a five-stage U-Net with shallow feature extraction, a two-scale SNN encoder-decoder, and image reconstruction. Its SNNBlock and OLIFBlock use iterative spiking and directional spatial processing to handle image features.

A. DehazeSNN Architecture

  • A five-stage U-Net processes shallow features through a two-scale encoder-decoder with five SNNBlocks, then reconstructs the image.SKfusion transfers information between matching scales.Source: A. DehazeSNN Architecture · p. 3

B. Spiking Neural Network Block

  • Each SNNBlock combines residual OLIFBlock and MLP sub-blocks, using iterative spiking instead of traditional attention to extract local information.Drop path connects the modules for regularization.Source: B. Spiking Neural Network Block · p. 3

C. Orthogonal Leaky-Integrate-and-Fire Module

  • A LIF neuron integrates input as its potential decays and emits a pulse when potential exceeds its threshold, after which the potential resets.The OLIF design uses this neuron behavior in spatial branches.Source: C. Orthogonal Leaky-Integrate-and-Fire… · p. 4
  • OLIFBlock accumulates potential spatially in feature-map groups rather than over time and replaces binary LIF pulses with continuous full-precision outputs.The output at each group step is retained as an intermediate state for the next iteration.Source: C. Orthogonal Leaky-Integrate-and-Fire… · pp. 4–5
  • OLIFBlock applies depth-wise convolution, then processes features in parallel horizontal and vertical branches whose outputs are combined at the original feature-map size.The branches accumulate spatial information through LIF calculations.Source: C. Orthogonal Leaky-Integrate-and-Fire… · p. 5

D. Residual Connection with SKfusion

DehazeSNN uses SKfusion to combine features from the two branches of a skip connection, rather than relying on direct concatenation.

D. Residual Connection with SKfusion

IV. EXPERIMENTS

Experiments used two DehazeSNN variants, trained with AdamW on a single RTX 4090 using combined L1 and perceptual losses.

IV. EXPERIMENTS

  • DehazeSNN-M and DehazeSNN-L were trained on one NVIDIA GeForce RTX 4090 GPU using AdamW.The passage refers to Table IV for the variants’ specific designs.Source: §IV · p. 5
  • Training used 256 × 256 patches, with batch sizes of 5 for DehazeSNN-M and 4 for DehazeSNN-L.Source: §IV · p. 5
  • The combined training loss used L1 loss and perceptual loss LPIPS.The initial learning rates were 5 × 10^-5 for OLIFBlock parameters Vth and τ and 1 × 10^-4 for the rest of the network, decaying to 1 × 10^-6.Source: §IV · p. 5

A. Datasets and Evaluation Criteria

The evaluation covers RESIDE synthetic haze subsets and RS-Haze remote-sensing images, using image-quality and computational-cost metrics.

A. Datasets and Evaluation Criteria

B. Photography Dehazing

DehazeSNN-L delivers strong photography-dehazing results across RESIDE benchmarks, with a lightweight computational profile and visual improvements in indoor and outdoor scenes.

B. Photography Dehazing

  • DehazeSNN-L raises RESIDE-ITS PSNR above 41 while using fewer parameters and less computation than models with PSNR above 40.On RESIDE-6K, DehazeSNN also surpasses the other algorithms.Source: B. Photography Dehazing · p. 6
  • On the larger RESIDE-OTS dataset, DehazeSNN ranks among the top performers, though its compact model struggles to capture all dataset features.The authors still highlight its parameter and computational efficiency.Source: B. Photography Dehazing · p. 6
  • Visual comparisons show fewer edge-shadow artifacts indoors and better haze and texture recovery at outdoor water-mist boundaries.The authors report that competing methods leave varying amounts of residual haze at those outdoor boundaries.Source: B. Photography Dehazing · p. 6

C. Remote Sensing Dehazing

DehazeSNN performs well on remote-sensing images with non-uniform haze, and ablations identify OLIFBlock and specific loss and neuron-group settings as important design choices.

C. Remote Sensing Dehazing

  • On RS-Haze, DehazeSNN significantly outperforms advanced dehazing methods while using minimal parameters and model size.Visual results are reported as closer to haze-free real images, with better color and contrast.Source: C. Remote Sensing Dehazing · p. 7

D. Ablation Study

  • Ablations identify OLIFBlock as pivotal to overall performance, while SKfusion offers a lower-cost alternative to concatenation without substantially boosting performance.The analysis uses models trained on RESIDE-6K and evaluated on mixed SOTS indoor-outdoor images.Source: D. Ablation Study · p. 7
  • The loss-function ablation finds the best performance with α1 = 0.5 for the weighted combination of L1 and LPIPS losses.The study evaluates the weighting between the two losses.Source: D. Ablation Study · p. 7
  • The LIF-group ablation achieves its highest accuracy at g = 4, which the experiments adopt uniformly.Here, g denotes the number of LIF neuron groups.Source: D. Ablation Study · p. 7

V. CONCLUSION

DehazeSNN is a U-Net-like spiking network for image dehazing, with OLIFBlock supporting multiscale feature extraction. Experiments report first-tier performance with a minimal model size, and future work will explore other image-restoration tasks.

V. CONCLUSION

  • DehazeSNN combines a U-Net-like design with spiking neural networks to dehaze images.The conclusion describes it as effective at detailed local analysis and managing long-term dependencies while reducing computational overhead.Source: §V · p. 7
  • OLIFBlock uses orthogonal image grouping and leaky-integrate-and-fire activity to extract and generalize features across scales.The conclusion attributes DehazeSNN's stated advantages to this block.Source: §V · p. 7
  • Across photography and remote-sensing dehazing datasets, DehazeSNN achieves first-tier performance with minimal model size and few parameters.Future research will explore applying the architecture to additional image-restoration tasks.Source: §V · p. 7
Loading 2512.23950v1…