Scholarlink

<p>Low-light image enhancement remains a critical challenge in computer vision, affecting tasks such as monitoring and recognition systems. Traditional deep learning models, while effective, often suffer from high computational complexity and inference latency, limiting their deployment on edge devices. To address this, we introduce SelfRefineGAN, a novel Generative Adversarial Network (GAN) that integrates a specialized Refinement Network (RFN) within a vanilla GAN framework. The RFN, operating as a curriculum network during training, guides the Extreme Lightweight Residual Gated Network (XLRGN) to iteratively refine image quality. Notably, the RFN is decoupled during inference, allowing the XLRGN to operate independently with minimal parameters and rapid runtime. Our method achieves state-of-the-art results across five challenging benchmarks, including LOLv1, LOLv2-Real, LOLv2-Synthetic, SICE, and MIT Adobe FiveK, with superior pixel-level accuracy (MAE as low as 0.017 on SICE) and perceptual quality (PSNR of 25.89 dB and SSIM of 0.904 on LOLv1). With only 0.025M parameters and 3.69 GFLOPs, SelfRefineGAN enables robust, high-fidelity perception for real-time applications. By overcoming the computational barriers associated with high-fidelity image restoration at inference, the proposed framework enables efficient, high-quality perception across diverse low-visibility environments, bridging the gap between sophisticated AI models and their practical deployment on resource-constrained devices. Source code is available at <a href="https://github.com/MuhammadAtif-251999/SelfRefineGAN.git">https://github.com/MuhammadAtif-251999/SelfRefineGAN.git</a></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

Enhancing low-light image quality with SelfRefineGAN: a lightweight GAN approach

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Muhammad Atif,
  • Yudong Zhang,
  • Saqib Mamoon

摘要

Low-light image enhancement remains a critical challenge in computer vision, affecting tasks such as monitoring and recognition systems. Traditional deep learning models, while effective, often suffer from high computational complexity and inference latency, limiting their deployment on edge devices. To address this, we introduce SelfRefineGAN, a novel Generative Adversarial Network (GAN) that integrates a specialized Refinement Network (RFN) within a vanilla GAN framework. The RFN, operating as a curriculum network during training, guides the Extreme Lightweight Residual Gated Network (XLRGN) to iteratively refine image quality. Notably, the RFN is decoupled during inference, allowing the XLRGN to operate independently with minimal parameters and rapid runtime. Our method achieves state-of-the-art results across five challenging benchmarks, including LOLv1, LOLv2-Real, LOLv2-Synthetic, SICE, and MIT Adobe FiveK, with superior pixel-level accuracy (MAE as low as 0.017 on SICE) and perceptual quality (PSNR of 25.89 dB and SSIM of 0.904 on LOLv1). With only 0.025M parameters and 3.69 GFLOPs, SelfRefineGAN enables robust, high-fidelity perception for real-time applications. By overcoming the computational barriers associated with high-fidelity image restoration at inference, the proposed framework enables efficient, high-quality perception across diverse low-visibility environments, bridging the gap between sophisticated AI models and their practical deployment on resource-constrained devices. Source code is available at https://github.com/MuhammadAtif-251999/SelfRefineGAN.git