Scholarlink

<p>Underwater images often suffer from severe color distortion, low contrast, and visibility loss due to light scattering and wavelength dependent attenuation, posing challenges for autonomous underwater vehicle (AUV) vision systems. While deep learning methods improve image quality, their computational cost hinders real time deployment in resource constrained environments. We propose an underwater image enhancement with energy-efficient potential framework using a Spiking Generative Adversarial Network (Spiking GAN), built upon a temporally adapted Cycle-SNSPGAN architecture. The model undergoes conventional training using PyTorch and the EUVP dataset, followed by conversion to a spiking domain using the Sinabs framework with rate-based encoding over 50 time steps. Key innovations include a spike compatible adversarial loss, temporal batch normalization for color consistency, and multi channel spike encoding to preserve RGB structure. The generator is fully transformed into a spiking network using Leaky Integrate-and-Fire (LIF) neurons, while maintaining weight fidelity through direct transfer and activation scaling. The discriminator is intentionally maintained as an analog network to ensure stable adversarial training. To bridge the spiking analog interface, we introduce a spike compatible adversarial loss, enabling effective gradient flow while preserving the temporal and sparse characteristics of the spiking generator. Experimental results show our model achieves a PSNR of 28.73&#xa0;dB and SSIM of 0.802, approaching or surpassing leading methods. Notably, it achieves a 0.07&#xa0;s inference time per image significantly faster than conventional models highlighting its suitability for real time, low power underwater vision systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Spiking GAN for underwater image enhancement

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Ouafa Benaida,
  • Abdelhamid Loukil

摘要

Underwater images often suffer from severe color distortion, low contrast, and visibility loss due to light scattering and wavelength dependent attenuation, posing challenges for autonomous underwater vehicle (AUV) vision systems. While deep learning methods improve image quality, their computational cost hinders real time deployment in resource constrained environments. We propose an underwater image enhancement with energy-efficient potential framework using a Spiking Generative Adversarial Network (Spiking GAN), built upon a temporally adapted Cycle-SNSPGAN architecture. The model undergoes conventional training using PyTorch and the EUVP dataset, followed by conversion to a spiking domain using the Sinabs framework with rate-based encoding over 50 time steps. Key innovations include a spike compatible adversarial loss, temporal batch normalization for color consistency, and multi channel spike encoding to preserve RGB structure. The generator is fully transformed into a spiking network using Leaky Integrate-and-Fire (LIF) neurons, while maintaining weight fidelity through direct transfer and activation scaling. The discriminator is intentionally maintained as an analog network to ensure stable adversarial training. To bridge the spiking analog interface, we introduce a spike compatible adversarial loss, enabling effective gradient flow while preserving the temporal and sparse characteristics of the spiking generator. Experimental results show our model achieves a PSNR of 28.73 dB and SSIM of 0.802, approaching or surpassing leading methods. Notably, it achieves a 0.07 s inference time per image significantly faster than conventional models highlighting its suitability for real time, low power underwater vision systems.