摘要
鉴于从噪声图像分解获得的原生图块集合的协方差矩阵前若干个特征值(按照升序排序)与图像噪声水平值具有强相关性,提出了一种基于主成分分析和深度神经网络的快速噪声水平估计算法.该算法首先选用原生图块集合协方差矩阵前若干个特征值构成刻画图像噪声水平高低的特征矢量,然后在大量有代表性且已标定噪声水平值的噪声图像集合上利用深度神经网络训练预测模型以实现将特征矢量直接映射为噪声水平值,最后为获得更高的预测准确性,采用粗精预测模型相结合的两步预测方式实现.实验表明:文中算法在各个噪声级别上都具有稳定的预测准确性,且执行效率非常高,作为降噪算法的前置预处理模块具有更好的综合优势.
Considering the fact that there exists the strong correlation between the first several eigenvalues( in ascending order) of the covariance matrix of the raw patches extracted from a noisy image and its noise level,we proposed a novel fast multiple image-based noise level estimation( FMNLE) algorithm using the principal component analysis( PCA) and the deep neural network( DNN). Specifically,we selected the first several eigenvalues of the raw patches to form a feature vector characterizing the noise level of an image. Then,we employed deep neural network to train an estimation model on a large number of representative natural images corrupted with known noise levels,by which the feature vector can be directly mapped into the corresponding noise level. To obtain higher estimation accuracy,a two-step estimation strategy was adopted.Extensive experiments show that, the estimation accuracy of the proposed algorithm is stable at each noise level with good efficiency,demonstrating a better comprehensive advantage as the pre-processing module for denoising algorithms.
引文
[1]Xu S P,Zeng X X,Jiang Y N,et al. A multiple image-based noise level estimation algorithm[J]. IEEE SignalProcessing Letters,2017,24(11):1701-1705.
[2]徐少平,张兴强,姜尹楠,等.局部均值噪声估计的盲3维滤波降噪算法[J].中国图象图形学报,2017,22(4):422-434.Xu Shao-ping,Zhang Xing-qiang,Jiang Yin-nan,et al.Noise level estimation based on local means and its appli-cation to the blind BM3D denoising algorithm[J]. Journalof Image and Graphics,2017,22(4):422-434.(in Chi-nese)
[3]周先春,吴婷,石兰芳,等.一种基于曲率变分正则化的小波变换图像去噪方法[J].电子学报,2018,46(3):621-628.Zhou Xian-chun,Wu Ting,Shi Lan-fang,et al. A kind ofw avelet transform image denoising method based on curva-ture variation regularization[J]. Acta Electronica Sinica,2018,46(3):621-628.(in Chinese)
[4]Dabov K,Foi A,Katkovnik V,et al. Image denoising withblock-matching and 3D filtering[A]. SPIE 6064,ImageProcessing:Algorithms and Systems,Neural Netw orks,andM achine Learning[C]. San Jose,CA,United States,2006,354-365.
[5]杨学志,叶铭,周芳,等.采用混合特征相似性的极化SAR图像降噪算法[J].电子学报,2016,44(11):2583-2591.Yang Xue-zhi,Ye M ing,Zhou Fang,et al. Speckle reduc-tion for polSAR images using hybrid features similarity[J]. Acta Electronica Sinica,2016,44(11):2583-2591.(in Chinese)
[6]Ghazi M M,Erdogan H. Image noise level estimation basedon higher-order statistics[J]. M ultimedia Tools and Appli-cations,2017,76(2):2379-2397.
[7]Xu S P,Hu L Y,Yang X H. Quality-aware features-basednoise level estimator for block matching and three-dimen-sional filtering algorithm[J]. Journal of Electronic Ima-ging,2016,25(1):Article No. 013029.
[8]Donoho D L,Johnstone I M. Adapting to unknown smooth-ness via w avelet shrinkage[J]. Journal of the AmericanStatistical Association,1995,90(432):1200-1224.
[9]Immerkaer J. Fast noise variance estimation[J]. ComputerVision and Image Understanding,1996,64(2):300-302.
[10]张旗,梁德群,樊鑫.基于小波域的图像噪声估计新方法[J].计算机工程,2004,30(8):37-39.Zhang Qi,Liang De-qun,Fan Xin. Estimating image noisebased on region segmentation in the w avelet domain[J].Computer Engineering,2004,30(8):37-39.(in Chi-nese)
[11] Zoran D,Weiss Y. Scale invariance and noise in naturalimages[A]. 2009 IEEE 12th International Conference onComputer Vision[C]. Kyoto,Japan:IEEE,2009,2209-2216.
[12]Liu W,Lin W S. Additive white Gaussian noise level esti-mation in SVD domain for images[J]. IEEE Transactionson Image Processing,2013,22(3):872-833.
[13] Pyatykh S,Hesser J,Zheng L. Image noise level estima-tion by principal component analysis[J]. IEEE Transac-tions on Image Processing,2013,22(2):687-699.
[14]Liu X H,Tanaka M,Okutomi M. Single-image noise levelestimation for blind denoising[J]. IEEE Transactions onImage Processing,2013,22(12):5226-5237.
[15] Cui G M,Feng H J,Xu Z H,et al. No-reference imagenoise estimation based on noise level accumulation[J].Optical Review,2016,23(2):208-219.
[16] Zhang L,Zhang L,Bovik A C. A feature-enriched com-pletely blind image quality evaluator[J]. IEEE Transac-tions on Image Processing,2015,24(8):2579-2591.
[17] Hinton G,Deng L,Dong Y,et al. Deep neural networksfor acoustic modeling in speech recognition:the sharedview s of four research groups[J]. IEEE Signal ProcessingM agazine,2012,29(6):82-97.
[18]Tanaka M,Okutomi M. A novel inference of a restrictedBoltzmann machine[A]. 2014 22nd International Confer-ence on Pattern Recognition[C]. Stockholm,Sw eden:IEEE,2014,1526-1531.
[19]Abdullah S M,Tischer P,Wijewickrema S,et al. Hierar-chical mutual nearest neighbour image segmentation[A].2016 International Conference on Digital Image Compu-ting:Techniques&Applications(DICTA)[C]. GoldCoast,QLD,Australia:IEEE,2016,1-8.
[20] Chen G Y,Zhu F Y,Heng P A. An efficient statisticalmethod for image noise level estimation[A]. 2015 IEEEInternational Conference on Computer Vision[C]. Santia-go,Chile:IEEE,2015,477-485.
[21]Yang S M,Tai S C. Fast and reliable image-noise estima-tion using a hybrid approach[J]. Journal of Electronic Im-aging,2010,19(3):033007.
[22]Zhang K,Zuo W M,Chen Y J,et al. Beyond a Gaussiandenoiser:residual learning of deep CNN for image denois-ing[J]. IEEE Transactions on Image Processing,2017,26(7):3142-3155.