基于语义信息和注意力机制的低光照图像增强
Low light image enhancement based on semantic information and attention mechanism
李浩彬 1华云松1
作者信息
- 1. 上海理工大学 光电信息与计算机工程学院,上海 200093
- 折叠
摘要
针对低光照条件下拍摄的图片存在低对比度、噪声等问题,提出了一种结合语义信息与注意力机制的低光照增强方法.首先,利用一对联合训练的U-Net网络,通过共享特征提取器,分别得到低光照图像的初步增强结果和语义信息分布概率图;然后,通过注意力机制模块把通过U-Net网络得到的低光照增强特征和语义特征进行信息融合,解决低光照下图片边缘信息丢失和曝光不足导致的图像模糊不清的问题.实验表明,该方法在处理低光照对比度不高和曝光不均匀图片时,可有效消除图像伪影以及提高图像饱和度与不同区域块的对比度.
Abstract
Aiming at the problems of low contrast and noise in low-light images,this paper proposes a low light enhancement method which combines semantic information and attention mechanism.First,a pair of jointly trained U-Net networks were used to obtain the preliminary enhancement results and the distribution probability of semantic information of low-light images by sharing feature extractors.Then,the low-light enhancement features and semantic features obtained by U-Net networks were fused through the attention mechanism module.The problem of image edge information loss under low illumination and image blurring under exposure was addressed.Experiment results show that the proposed method can effectively eliminate artifacts when processing low illumination images with low contrast and uneven exposure,and improve image saturation and contrast of different regions.
关键词
低光照图像增强/注意力机制/语义信息Key words
low light image enhancement/attention mechanism/semantic information引用本文复制引用
出版年
2024
NETL
NSTL
万方数据