摘要
零样本低照度图像增强无须配对数据,但光照噪声干扰先验提取易产生伪影。为了解决这一问题,本文提出自注意力生成对抗网络(SA-GAN),集成亮度、空间、通道三重注意力强化暗部建模。设计梯度范数自适应损失平衡算法动态协调多目标约束。多基准集实验表明,该方法性能已媲美主流方法,提升了图像的自然度和可见性。研究结论表明,本文提出的自注意力生成对抗网络(SA-GAN)在低照度图像增强领域展现出显著优势。通过集成亮度、空间、通道三重注意力机制,SA-GAN有效强化了暗部区域的建模能力,提升了图像的亮度和细节重建质量。同时,引入的梯度范数自适应损失平衡算法,动态协调了多目标约束,增强了训练的稳定性和适应性。来研究可探索更轻量的注意力结构,以拓展至视频低光增强场景。
Low-Light Enhancement via Self-Attention and Multi-Loss Optimization
Zhang Min-hang 1, Song Xiao-qi 2, Sun Yu 2
Author Information
- 1 University of Electronic Science and Technology of China, Chengdu 611731, Sichuan Province, China
- 2 China Railway Lanzhou Group Co., Ltd., Lanzhou 730000, Gansu Province, China
Abstract
Zero-shot low-light image enhancement avoids paired data, yet prior extraction disturbed by illumination noise easily produces artifacts. To solve this problem, we propose a Self-Attention Generative Adversarial Network (SA-GAN) that integrates brightness, spatial, and channel attention to strengthen dim-region modeling. Agradient-norm adaptive loss-balancing algorithm is designed to dynamically coordinate multi-objective constraints. Experiments on multiple benchmarks show that the proposed method rivals state-of-the-art approaches and improves image naturalness and visibility. The study concludes that SA-GAN exhibits signifcant advantages in low-light enhancement. The triple-attention mechanism efectively reinforces modeling of dark regions, enhancing brightness and detail reconstruction, while the adaptive loss-balancing algorithm stabilizes training and improves adaptability. Future work can explore lighter attention structures for extension to video low-light enhancement.
图文摘要

SA-GAN低光增强图片摘要:
左栏呈现零样本增强的输入痛点;中栏概括迭代曲线映射、亮度/空间/通道三重注意力、相对判别器与梯度范数自适应多损平衡;右栏汇总LIME、DICM、NPE、MEF数据集上的增强效果与验证结论。
亮点
提出零样本SA-GAN:生成器以迭代曲线映射为骨架,叠加亮度、空间、通道三重自注意力,使网络优先聚焦暗部并重建结构细节;相对判别器从纹理与光照分布上提升真实感。创新点是梯度范数驱动的自适应损失平衡,按各损失梯度幅值动态分配Lgen、LSPF、Lexp、Lcol权重,缓解固定权重导致的训练失衡与跨场景调参。在EnlightenGAN无配对集训练、LIME/DICM/NPE/MEF测试,文中报告主观与客观表现优于主流方法,能提亮暗部、抑制噪声伪影及局部过曝,并保持自然度;未来可转向轻量结构与视频低光增强。
