摘要
大多数现有去噪方法只考虑了噪声输入图像的非局部自相似性先验方法(NSS),仅从已退化的输入图像中收集相似图像块,图像去噪的质量在很大程度上取决于输入图像本身。针对图像复原过程中的噪声去除问题,设计了一种基于卷积神经网络的组稀疏去噪模型。模型使用两种NSS先验(即噪声输入图像和预滤波图像的NSS先验),把图像去噪问题转化为组稀疏残差最小化问题。为了提高非局部相似块选择的准确性,使用了一种自适应块搜索的方法,并采用卷积神经网络进行预滤波,以获得对原始图像组稀疏系数的良好估计。实验结果表明:所提出的GSRC-CNN方法在客观和感知质量方面优于许多先进的去噪方法。
Most image denoising methods only consider the non-local self-similarity(NSS)of noise input images,and collect similar image blocks only from degraded input images.To some extent,the effects of image denoising depends on the input image.Aiming at the problem of denoising in image restoration,a group sparse denoising model based on convolutional neural network is established.Using two NSS priors(i.e.NSS priors for noise input images and pre-filtering images),the image denoising problem is transformed into group sparse residual minimization problem.Besides,to improve the accuracy of non-local similar blocks selecting,an adaptive block search method is used,group sparse coefficient of original image and convolutional neural network is used to pre-filter image,in order to estimate well.The experimental results show that the proposed GSRC-CNN method outperforms many state-of-the-art image denoising methods in terms of objective and perceived quality.
作者
陈梦雅
李润鑫
刘辉
尚振宏
CHEN Mengya;LI Runxin;LIU Hui;SHANG Zhenhong(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)
出处
《传感器与微系统》
CSCD
2020年第2期48-51,共4页
Transducer and Microsystem Technologies
基金
国家自然科学基金资助项目(11873027)
关键词
卷积神经网络
自适应块搜索
组稀疏残差约束
预滤波
convolutional neural network(CNN)
adaptive block search
group sparse residual constraint
pre-filtering
