结合噪声掩模训练和最近邻搜索机制的轻量级椒盐去噪  

Lightweight salt-and-pepper denoising combining noise mask training and nearest searching mechanism

作　　者：黄成强 金星 HUANG Cheng-qiang;JIN Xing(School of Physics and Electronics Sciences,Zunyi Normal University,Zunyi 563006,China)

机构地区：[1]遵义师范学院物理与电子科学学院,贵州遵义563006

出　　处：《液晶与显示》2023年第9期1234-1247,共14页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.62164014);贵州省科技厅基础研究重点项目(No.黔科合基础[2018]1424);贵州省教育厅创新群体项目(No.黔教合KY字[2018]028);贵州省教育厅青年人才成长项目(No.黔教合KY字[2020]097);遵义市科技局联合基金(No.遵市科合HZ字[2021]210)。

摘　　要：随着图像处理应用在各新兴领域的不断扩展,高性能椒盐去噪仍然是一项具有挑战性的任务。本文提出了一种结合噪声掩模训练和最近邻搜索机制的椒盐去噪方法。首先,搭建一个包含9个卷积层的轻量级神经网络,用于生成高质量的噪声掩模。接着,根据该噪声掩模的噪点标记结果,正常像素不作处理,通过最近邻搜索机制寻找与噪点最相邻的正常像素灰阶替代噪点灰阶。本文提出了一种用于噪点标记的轻量级卷积神经网络。在降低网络深度的同时,在中间层采用深度可分离卷积代替常规卷积,这两个因素使得运算复杂度和参数量得到数量级的降低。另外,提出了一种基于最近邻搜索机制的去噪方法,提升了去噪性能。实验结果表明,所提出网络的运算复杂度比传统网络有数量级的降低,训练所得噪声掩模的误判率分别比极点标记、均值标记和极值图像块标记分别降低了94.79%、94.79%和83.65%。此外,去噪图像的峰值信噪比相比于传统卷积神经网络方法的处理结果提升了2.53%,信息损失降低了6.76%。本文首次将轻量级卷积神经网络应用于椒盐去噪,降低了网络的复杂度,提升了去噪性能。As the application of image process extends to each emerging field,high-performance salt-and-pepper denoising is still a challenging task.Therefore,a salt-and-pepper denoising method combining training of noise mask and nearest searching mechanism is proposed.Firstly,a lightweight neural network with 9 convolutional layers is constructed to generate a high-quality noise mask.Subsequently,according to the marking result of this mask,the normal pixel is not processed,while gray level of the noise pixel is replaced by that of the nearest normal pixels,which is found by using the nearest searching mechanism.In this paper,a lightweight convolutional neural network for noise labeling is proposed.While reducing the network depth,the conventional convolution for the middle layer is replaced by depth-separable convolution.These two factors reduce computational complexity and parameters number by orders of magnitude.And a denoising method based on the nearest searching mechanism is proposed,which will improve the denoising performance.The pixel units marked as normal points are not processed,and only noise points are processed.Experimental results show that the computational complexity of the proposed network is orders of magnitude lower than that of traditional networks,the misjudging rate for the trained noise mask is 94.79%,94.79%and 83.65%lower than that of the extreme marking,the extreme image block marking and the average marking,respectively.In addition,PSNR of image processed by the proposed method is 2.53%higher than traditional CNN method,and MSE is 6.76%lower.A lightweight convolutional neural network is applied to salt and pepper denoising for the first time,which reduces network complexity and improves denoising performance.

关 键 词：椒盐噪声 噪声掩模 轻量级卷积神经网络 最近邻搜索 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP751[自动化与计算机技术—计算机科学与技术]