摘要
针对低照度图像质量较差、噪声多、纹理模糊等问题,提出一种基于双频域特征聚合的低照度增强网络(dual frequency-domain feature aggregation network,DF-DFANet)。首先,构建频谱光照估计模块(frequency domain illumination estimation module,FDIEM)实现跨域特征提取,通过共轭对称约束调整频域特征图抑制噪声信号,并采用逐层融合方式提高多尺度融合效率以扩大特征图感受野范围。其次,设计多谱双注意力模块(multiple spectral attention module,MSAM)聚焦图像局部频率特征,通过小波域空间、通道注意力机制关注图像细节信息。最后,提出双域特征聚合模块(dual domain feature aggregation module,DDFAM)融合傅里叶域和小波域特征信息,利用激活函数计算自适应调整权重实现像素级图像增强,并结合傅里叶域全局信息提高融合效果。实验结果表明,在LOL数据集上所提网络的PSNR达到24.3714,SSIM达到0.8937。与对比网络相比,所提网络增强效果更具自然性。
Aiming at the problems of poor low-light image quality,noise,and blurred texture,a low-light enhancement network(DF-DFANet)based on dual-frequency domain feature aggregation is proposed.Firstly,a spectral illumination estimation module(FDIEM)is constructed to realize cross-domain feature extraction,which can adjust the frequency domain feature map to suppress noise signals through conjugate symmetric constraints and improve the multi-scale fusion efficiency by layer-by-layer fusion to expand the range of the feature map.Secondly,the multispectral dual attention module(MSAM)is designed to focus on the local frequency characteristics of the image,and pay attention to the detailed information of the image through the wavelet domain space and channel attention mechanism.Finally,the dual-domain feature aggregation module(DDFAM)is proposed to fuse the feature information of the Fourier domain and the wavelet domain,and use the activation function to calculate the adaptive adjustment weight to achieve pixel-level image enhancement and combine the Fourier domain global information to improve the fusion effect.The experimental results show that the PSNR of the proposed network on the LOL dataset reaches 24.3714 and the SSIM reaches 0.8937.Compared with the comparison network,the proposed network enhancement effect is more natural.
作者
徐胜军
杨华
李明海
刘光辉
孟月波
韩九强
Xu Shengjun;Yang Hua;Li Minghai;Liu Guanghui;Meng Yuebo;Han Jiuqiang(College of Information and Control Engineering,Xi′an University of Architecture and Technology,Xi′an,Shaanxi 710055,China;Xi′an Key Laboratory of Building Manufacturing Intelligent&Automation Technology,Xi′an,Shaanxi 710055,China)
出处
《光电工程》
CAS
CSCD
北大核心
2023年第12期28-44,共17页
Opto-Electronic Engineering
基金
国家自然科学基金面上项目(52278125)
陕西省重点研发计划(2020GY-186,2021SF-429)
陕西省自然科学基础研究计划(2023-JC-YB-532,2022JQ-681)。
关键词
深度学习
图像增强
傅里叶变换
小波变换
双域融合
注意力机制
deep learning
image enhancement
fourier transform
wavelet transform
dual-domain convergence
attention mechanism
