高低频特征融合的低照度图像增强方法

PDF下载HTML

Title:

Low light image enhancement based on high and low frequency feature fusion

作者:

王德文^1,2, 胡旺盛¹, 张润磊¹, 赵文清^1,3

1. 华北电力大学 计算机系, 河北 保定 071003;
2. 河北省能源电力知识计算重点实验室, 河北 保定 071003;
3. 复杂能源系统智能计算教育部工程研究中心, 河北 保定 071003

Author(s):

WANG Dewen^1,2, HU Wangsheng¹, ZHANG Runlei¹, ZHAO Wenqing^1,3

1. Department of Computer, North China Electric Power University, Baoding 071003, China;
2. Hebei Key Laboratory of Knowledge Computing for Energy & Power, Baoding 071003, China;
3. Engineering Research Center of the Intelligent Computing for Complex Energy System, Ministry of Education, Baoding 071003, China

关键词:

低照度; 图像增强; 高频特征; 低频特征; 特征融合; 注意力; 多尺度; 残差网络; 密集连接

Keywords:

low light; image enhancement; high frequency feature; low frequency feature; feature fusion; attention; multi-scale; residual net; dense connection

分类号:

TP391.41

DOI:

10.11992/tis.202405026

摘要:

针对现有低照度图像增强方法中性能与开销不平衡的问题，本文提出一种高低频特征融合的低照度图像增强方法。该方法在多尺度上提取几何特征丰富的低频特征与语义特征丰富的高频特征，经过高低频特征融合得到增强图像，在保证良好图像质量的同时降低开销。为提升低照度环境下的特征提取能力，构建残差混合注意力模块，从像素与通道两方面对重要的局部区域给予更多关注。针对下采样导致的信息丢失问题，提出一种特征合并模块对下采样后的特征进行特征补充。此外，通过多级残差密集连接模块增强特征复用能力。在SID(see-in-the-dark)数据集上的实验表明，该方法峰值信噪比和结构相似度分别达到29.67和0.792，模型参数量仅为1.5×10⁶。

Abstract:

To address the imbalance between performance and cost in existing low light image enhancement, a low light image enhancement method is proposed based on high and low frequency feature fusion. By integrating multi-scale data, this fusion combines geometrically rich low frequency features with semantically rich high frequency features to obtain enhanced images, thereby reducing cost while guaranteeing good image quality. To enhance the feature extraction ability in a low light environment, the residual mix-attention module is designed to focus more on important local regions from the pixel and channel perspectives. To address the information loss due to downsampling, the feature merging module is used to supplement the features after downsampling. Additionally, a multi-residual dense block module is designed to strengthen the feature-reuse capability. Furthermore, the see-in-the-dark dataset was subjected to experiments. Overall, this method achieved peak signal-to-noise ratio and structural similarity of 29.67 and 0.792, respectively, with only 1.5×10⁶ parameters.