| 摘 要: 针对低照度图像增强中特征学习不足与噪声干扰问题,提出一种融合注意力与高级特征学习的增强网络。首先,设计融合注意力模块,通过多尺度通道感知与空间特征融合抑制噪声并增强关键区域;同时,引入自注意蒸馏机制,利用高层特征引导底层特征优化,实现注意力特征的高效传递;其次,结合深度监督与对比学习策略,通过多层级监督信号与特征区分性约束提升高级特征表示能力。实验表明,所提方法在LOL-v1和LOL-v2-real以及LOL-v2-synthetic数据集上的SSIM指标达到了0.904、0.922、0.929。均优于其他方法。 |
| 关键词: 低照度 注意力机制 高级特征 自注意蒸馏 深度监督 对比学习 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(32360437);甘肃省高等学校创新基金项目(2021A-056) |
|
| Low-Light Image Enhancement Based on Fused Attention and High-Level Feature Learning |
|
LI Wan1,2,WANG Hangfei1,2,HUANG Sha1,2, CHANG Chenyang1,2
|
(1.Shaanxi University of Science and Technology Joint Artificial Intelligence Laboratory, Xi’an 710021, China;
2.School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi’an 710021, China)
liwan@sust.edu.cn; 3467431562@qq.com; 11627200235@qq.com; 1950694830@qq.com
|
| Abstract: To address the challenges of insufficient feature learning and noise interference in low-light image enhancement, we propose an enhancement network that integrates attention mechanisms with high-level feature learning. First, a fused attention module is designed to suppress noise and highlight key regions through mult-i scale channel perception and spatial feature fusion. Simultaneously, a sel-f attention distillation mechanism is introduced, leveraging high-level features to guide the optimization of low-level features, thereby enabling efficient transfer of attentive features. Second, deep supervision and contrastive learning strategies are combined to enhance high-level feature representation through mult-i level supervisory signals and feature discriminability constraints. Experiments
demonstrate that the proposed method achieves SSIM scores of 0.904, 0.922, and 0.929 on the LOL-v1, LOL-v2-real, and LOL-v2-synthetic datasets, respectively, consistently outperforming existing approaches. |
| Keywords: low-light enhancement attention mechanism high-level features sel-f attention distillation deep
supervision contrastive learning |
