| 刘洋,刘思瑞,徐晓淼,王竹筠.基于跳跃连接神经网络的无监督弱光图像增强算法[J].电子测量与仪器学报,2025,39(5):208-216 |
| 基于跳跃连接神经网络的无监督弱光图像增强算法 |
| Low-light image enhancement algorithm of multi-layerneural network based on hopping connection |
| |
| DOI： |
| 中文关键词: 弱光图像增强 深度曲线估计 无参考损失函数 多层卷积神经网络 无监督学习 |
| 英文关键词:low-light image enhancement deep curve estimation no reference loss function multi-layer convolutional neural networks unsupervised learning |
| 基金项目:国家自然科学基金(62003224)、辽宁省教育厅项目(JYT2020042)资助 |
|
|
|
|
| 摘要点击次数: 775 |
| 全文下载次数: 816 |
| 中文摘要: |
| 针对Zero-DCE网络存在细节丢失和不同亮度区域处理结果出现差异等问题,设计了一种基于增强深度曲线估计网络(EnDCE Net)的无监督弱光图像增强算法。通过探索弱光图像与未配对的正常光照图像之间的潜在映射关系,实现了对低光照场景下图像质量的显著改善。首先,提出新的特征提取网络,该网络整合了多个跳跃连接与卷积层,实现低层与高层特征的有效融合,从而学习到弱光图像中的关键特征,增强网络对弱光图像的学习能力。其次, 设计一组联合的无参考损失函数,强调优化过程中与亮度相关的特性,从而更有利于图像增强模型的参数更新,提高图像增强的质量和效果。为了验证所提出算法的有效性,在5个公开数据集上进行了对比实验,与次优算法Zero-DCE相比,有参考数据集SICE上的峰值信噪比（PSNR）和结构相似性（SSIM）分别提升了9.4%、21%。无参考数据集LIME、DICM、MEF、NPE上NIQE分别达到了4.04、3.04、3.35、3.83。实验结果表明,所提出算法表现出色,增强后的图像色彩自然,亮度均衡且细节清晰。无论是主观视觉评价还是客观定量指标,均显著优于对比算法,充分体现了在图像增强效果上的卓越性和先进性。 |
| 英文摘要: |
| To address issues such as detail loss and inconsistent results across different brightness regions in the Zero-DCE network, an unsupervised low-light image enhancement algorithm based on the enhanced depth curve estimation network (EnDCE-Net) is proposed. This algorithm explores the potential mapping relationship between low-light images and unpaired normal-light images to achieve significant improvements in image quality under low-light conditions. First, a novel feature extraction network is introduced, which integrates multiple skip connections and convolutional layers, allowing for the effective fusion of low-level and high-level features. This enables the network to learn the key features of low-light images and enhances its ability to process them. Second, a set of joint no-reference loss functions is designed, emphasizing brightness-related features during the optimization process, which facilitates more efficient parameter updates and enhances the overall quality and effectiveness of the image enhancement. To evaluate the effectiveness of the proposed algorithm, comparative experiments were conducted on five publicly available datasets. Compared to the suboptimal algorithm Zero-DCE, the PSNR and SSIM on the reference dataset SICE were improved by 9.4% and 21%, respectively. On the no-reference datasets LIME, DICM, MEF, and NPE, the NIQE scores reached 4.04, 3.04, 3.35, and 3.83, respectively. The experimental results demonstrate that the proposed algorithm outperforms others, producing enhanced images with natural colors, balanced brightness, and clear details. Both subjective visual assessments and objective quantitative metrics show significant improvements over the competing algorithms, highlighting the excellence and advancement of the proposed method in image enhancement. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
