采用双通道卷积神经网络构建的随机脉冲噪声深度降噪模型  

A Dual-Channel Deep Convolutional Neural Network Model for Random-Valued Impulse Noise Removal

作　　者：徐少平[1] 林珍玉 崔燕 刘蕊蕊 杨晓辉[1] XU Shaoping;LIN Zhenyu;CUI Yan;LIU Ruirui;YANG Xiaohui(School of Information Engineering,Nanchang University,Nanchang 330031,China)

机构地区：[1]南昌大学信息工程学院,南昌330031

出　　处：《电子与信息学报》2020年第10期2541-2548,共8页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61662044,61163023);江西省自然科学基金(20171BAB202017)。

摘　　要：为提高对随机脉冲噪声(RVIN)图像的降噪效果,该文提出一种被称为双通道降噪卷积神经网络(D-DnCNN)的RVIN深度降噪模型。首先,提取多个不同阶对数差值排序(ROLD)统计值及1个边缘特征统计值构成描述图块中心像素点是否为RVIN噪声的噪声感知特征矢量。其次,利用预先训练好的深度置信网络(DBN)预测模型实现特征矢量到噪声标签的映射,完成对噪声图像中噪声点的检测。再次,在噪声检测标签的指示下采用Delaunay三角剖分插值算法快速修复噪声像素点从而获得初步复原图像。最后,将初步复原图像作为参考图像与噪声图像联接(concatenate)后输入D-DnCNN模型后获得残差图像,将参考图像减去残差图像即可获得降噪后图像。实验数据表明:D-DnCNN模型在各个噪声比例下的降噪效果均显著超过了现有的经典开关型RVIN降噪算法,与普通的单通道RVIN深度降噪模型相比也有较大幅度提升。A Dual-channel Denoising Convolutional Neural Network(D-DnCNN)model for the removal of Random-Valued Impulse Noise(RVIN)is proposed.To obtain the reference image quickly,several Rank-Ordered Logarithmic absolute Difference(ROLD)statistics and one edge feature statistic are first extracted from a local window to construct a RVIN-aware feature vector which can describe the central pixel of the patch is RVIN or not.Next,a noise detector based on Deep Belief Network(DBN)is trained to map the extracted feature vectors to their corresponding noise labels to detect all noise-like pixels in the observed image.Then,under the guidance of noise labels,the Delaunay triangulation-based interpolation algorithm is exploited to restore all detected noise-like pixels quickly and generate a preliminary restored image used as reference image.Finally,the reference image and the noisy image are simultaneously fed into the D-DnCNN model to output its corresponding residual image,and the final restored image can be obtained by subtracting the residual image from the noisy image.Extensive experimental results show that,the denoising effect of the proposed D-DnCNN denoising model outperforms the existing state-of-art switching ones across a range of noise ratios,and it also works better than the ordinary single-channel DnCNN model.

关 键 词：图像处理 随机脉冲噪声 双通道降噪卷积神经网络 参考图像 噪声感知特征 噪声检测 插值 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]