基于双通道神经网络的图像脉冲噪声降噪  

Denoising process for image impulse noise based on dual-channel neural networks

作　　者：袁新颜[1] YUAN Xinyan(School of Electronics and Information,Jiangsu Vocational College of Business,Nantong 226011,China)

机构地区：[1]江苏商贸职业学院电子与信息学院,江苏南通226011

出　　处：《光学技术》2021年第4期507-512,共6页Optical Technique

基　　金：江苏省高等学校自然科学基金面上项目(18KJB520015)。

摘　　要：脉冲噪声是成像过程中的一个主要噪声源,传统的滤波器难以有效消除高密度的脉冲噪声。针对这一问题,提出了一种基于非对称并行神经网络的图像脉冲噪声降噪算法。算法利用隐写分析丰富模型提取图像的噪声卷积特征图;将原图像特征图与噪声卷积特征图分别送入两个相同的卷积神经网络进行处理;结合l_(1)损失与l_(2)损失作为神经网络的总代价函数,同时利用了l_(1)损失的高视觉效果与l_(2)损失的强收敛性。实验结果表明:提出的降噪算法在各密度下的降噪性能均优于基于滤波器的降噪算法,对于高密度脉冲噪声也具有明显优势。The impulse noise is one of the main noise sources during the imaging processes, meanwhile, it is hard to remove high density impulse noise with traditional filters. To solve this problem, an image denoising algorithm for impulse noise based on asymmetric neural networks is proposed. In this algorithm, the steganalysis rich model is used to extract the convoluted noise feature maps of the noisy image. then the feature maps of original image and noisy convoluted version are delivered to two identical convolutional neural networks. The l1 loss and l2 loss are combined as the total loss function of the neural networks, it takes advantages of high visual quality of l1 loss and strong convergence of l2 loss. Experimental results show that the proposed denoising method outperforms the filter based denoising algorithms on different densities of noise, it also performs better on high density impulse noise.

关 键 词：激光成像 图像噪声 脉冲噪声降噪 卷积神经网络 噪声滤波器 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]