结合密集残差块和注意力的真实图像去噪网络
Real-world images denoising network combining dense residual blocks and attention
余卓璞 1周冬明 1周联敏 1赵倩 1尹稳1
作者信息
- 1. 云南大学信息学院,云南昆明 650000
- 折叠
摘要
为有效去除真实图像噪声并保留图像边缘信息,提出一种结合密集网络思想和并行极化自注意力机制的真实去噪算法.使用3条并行结构处理不同尺度的特征信息,其中每条分支由两个密集注意力块串联而成,形成残差结构.使用选择性核融合机制,获取不同深度下的特征信息,将其融合并使用注意力机制去除冗余信息,获取干净图像.实验结果表明,该算法在SIDD、DND、PolyU测试集上的峰值信噪比分别为39.32 dB、39.52 dB和37.36 dB,结构相似性分别为0.908、0.951和0.952,在SIDD和PolyU测试集上的图像通用质量指标值为0.992和0.982,在去噪任务上可以达到较好的性能,提高了图像视觉的质量.
Abstract
To effectively remove real-world image noise and preserve image edge information,a real-world denoising algorithm was proposed that combined the idea of dense networks and parallel polarized self-attention mechanism.Three parallel structures were used to process feature information at different scales,where each branch was composed of two dense attention blocks in se-ries,forming a residual structure.A selective kernel fusion mechanism was used to obtain feature information at different depths to fuse them,and an attention mechanism was used to remove redundant information to finally obtain a clean image.It is experi-mentally demonstrated that the algorithm can achieve better performance on the denoising task and improve the quality of image vision with the peak signal-to-noise ratios of 39.32 dB,39.52 dB and 37.36 dB,the structural similarities of 0.908,0.951 and 0.952 on the SIDD,DND and PolyU datasets,respectively,and the universal quality index values of 0.992 and 0.982 on the SIDD and PolyU datasets.
关键词
真实图像去噪/深度学习/卷积神经网络/密集残差网络/多尺度/注意力机制/深度卷积Key words
real-world denoising/deep learning/convolutional neural network/dense residual networks/multiscale/attention mechanisms/depth-wise convolution引用本文复制引用
基金项目
国家自然科学基金(62066047)
国家自然科学基金(61365001)
出版年
2024
NETL
NSTL
万方数据