ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2019, Vol. 56 ›› Issue (5): 1060-1070.doi: 10.7544/issn1000-1239.2019.20180185

• 人工智能 • 上一篇    下一篇

基于CNN噪声分离模型的噪声水平估计算法

徐少平,刘婷云,李崇禧,唐祎玲,胡凌燕   

  1. (南昌大学信息工程学院 南昌 330031) (xushaoping@ncu.edu.cn)
  • 出版日期: 2019-05-01
  • 基金资助: 
    国家自然科学基金项目(61662044,61163023,51765042);江西省自然科学基金项目(20171BAB202017)

Noise Level Estimation Algorithm Using Convolutional Neural Network-Based Noise Separation Model

Xu Shaoping, Liu Tingyun, Li Chongxi, Tang Yiling, Hu Lingyan   

  1. (School of Information Engineering, Nanchang University, Nanchang 330031)
  • Online: 2019-05-01

摘要: 现有的噪声水平估计(noise level estimation, NLE)算法通常采取先将图像内容信号与噪声信号分离,然后基于分离出的噪声信号估计出图像噪声水平值的实现策略.由于仅有噪声图像本身的信息可以利用,这些算法为保证噪声分离的准确性设计了各种复杂的处理过程,导致其执行效率偏低.为此,提出一种新的基于卷积神经网络噪声分离模型的NLE算法.首先,对大量原始无失真图像施加不同噪声水平的高斯噪声获得噪声图像集合,然后利用卷积神经网络构建一个专门从噪声图像中分离噪声信号获得噪声映射图(noise mapping)的预测模型.考虑到噪声映射图的系数值具有类高斯分布特性,利用广义高斯分布(generalized Gaussian distribution, GGD)模型对噪声映射图建模并以模型参数值作为反映图像噪声水平高低的特征值.最后,利用改进的BP神经网络将该特征值映射为最终的噪声水平预测值.大量实验数据表明:所提出的NLE算法在预测准确度和执行效率2个方面的综合性能优于现有的NLE算法,更具实用价值.

关键词: 噪声水平估计, 噪声分离, 卷积神经网络, 广义高斯模型, 噪声水平感知特征, 噪声水平值映射

Abstract: The existing noise level estimation (NLE) algorithms usually adopt the strategy that separates the noise signal from the content of an image to estimate its noise level. Since only a single noisy image can be exploited, these algorithms usually design a variety of complex processes to ensure the accuracy of noise separation, resulting in low execution efficiency. To this end, a novel NLE algorithm using convolutional neural network (CNN)-based noise separation model is proposed in this paper. Specifically, we first add Gaussian noise with different levels to a great amount of representative undistorted images to obtain a training database. Then, we train a CNN-based noise separation model on the training database to obtain the noise mapping from a given noisy image. Considering the fact that the coefficients of the noise mapping show Gaussian distribution behavior, we utilize the generalized Gaussian distribution (GGD) to model the coefficients of the noise mapping, and use two parameters (scale and shape) of the model as the noise level-aware features (NLAF) to describe the level of a noisy image. Finally, an improved back propagation (BP) neural network is used to map the NLAF features to the final noise level. Extensive experiments demonstrate that our method outperforms the most existing classical NLE algorithms in terms of both computational efficiency and estimation accuracy, which makes it more practical to use.

Key words: noise level estimation (NLE), noise separation, convolutional neural network (CNN), generalized Gaussian distribution (GGD), noise level-aware feature (NLAF), noise level mapping

中图分类号: