Research On Fast Noise Level Estimation Algorithm Based On Noise Level-aware Features

So far,most image denoising algorithms suffer from the drawback that their performance are restricted to the accurate estimation of the noise level in the image to be denoised.Therefore,a robust noise level estimation(NLE)method that automatically provides the noise level for a given image is of great theoretical significance and practical value.As the preprocessing module of most denoising algorithms,the estimation accuracy and execution efficiency of the NLE algorithms are two important indicators reflecting its performance.Most of the existing NLE algorithms adopt the so-called single-image based strategy to estimate the noise level.Since limited information available,these algorithms design complex processes to ensure the estimation accuracy,especially for the images with highly textured content,resulting in low execution efficiency and restricted practical application.Therefore,to solve the problems of NLE algorithms in estimation accuracy and execution efficiency,corresponding research carried out.Since the single-image based NLE(SNLE)algorithms are of low efficiency in execution,in this paper,three novel fast NLE algorithms based on multiple image strategies were proposed by making full use of the regular information of known multiple noisy images.First,a multi-image based NLE(MNLE)algorithm based on prior knowledge of multiple images was proposed.Specifically,based on the some statistics of natural image disturbed by noise will change regularly with the noise level,we can extract these statistics as noise level-aware features(NLAFs).Then the NLAFs are mapped to corresponding noise level by weighted average approach.In this algorithm,the NLAFs are composed of two parameters from the generalized Gaussian distribution(GGD)model and four parameters from the asymmetric generalized Gaussian distribution(AGGD)in four directions,resulting in eighteen dimensional features.Compared with the existing NLE algorithms,the execution efficiency of the MNLE algorithm has been significantly increased,but there still remains room for further improvement in terms of estimation accuracy.The main reason is that the MNLE algorithm uses too much features,which are borrowed fromthe relevant features of image quality assessment(IQA)algorithms,and they cannot effectively characterize the noise level.To obtain the features that can better reflect the noise level,a fast noise level estimation algorithm(SFNLE)that applies principal component analysis(PCA)to the smallest eigenvalue of nonlinear rectification was proposed.The SFNLE algorithm is based on the fact that when natural image is disturbed by noise in varying degrees,the smallest eigenvalue of the covariance matrix of the raw patches after PCA transformation is significantly correlated with the noise level.Based on this,the smallest eigenvalue was extracted from the covariance matrix of the raw patches of noisy images with known noise levels,then constructing a nonlinear rectification model between the smallest eigenvalue and the noise level combining with polynomial regression technique.Compared with the MNLE algorithm,the SFNLE algorithm has the characteristics of high efficiency and improved estimation accuracy.However,it is still difficult to describe the distortion degree in noisy image steadily based on a single feature.Moreover,the nonlinear mapping ability of polynomial fitting function is not good enough.Therefore,we proposed a so-call fast NLE(TFNLE)algorithm based on two-stage AdaBoost to further improve its performance.Considering the fact that the first several eigenvalues of the covariance matrix of raw patches are significantly correlated with the noise level,the TFNLE algorithm uses AdaBoost technique to train estimation model on representative natural images,and then the first several eigenvalues of the covariance matrix are directly mapped to the noise level.To obtain higher estimation accuracy at low,medium and high noise levels,we adopt a two-stage AdaBoost strategy from coarse to fine.Specifically,we first estimated the noise level roughly with a coarse prediction model trained with AdaBoost technique;then,according to the preliminary estimation result,the final noise level was obtained with a more accurate prediction model that is specifically trained for low,medium,or high noise levels.The estimation accuracy of the proposed TFNLE algorithm has been further improved,and the execution efficiency is satisfied.To verify the performance of MNLE,SFNLE,and TFNLE algorithms,the proposed algorithms are compared with state-of-the-art NLE algorithms in terms of both estimation accuracy and execution efficiency on commonly used image set,BSD database and challenging Waterloo database.To further demonstrate the practical application of the proposed algorithm,we used the ground truths and the estimated noise levels of the three algorithms as the entry parameters of the BM3 D algorithm.Experimental results show that,compared with the existing NLE algorithms,the proposed TFNLE algorithm outperforms them in terms of both estimation accuracy and execution efficiency,when it is used as the preprocessing module of various image processing algorithms.Thus,the problem that the non-blind denoising algorithms must be manually fed with the accurate noise level to obtain better denoising results can be solved.