Research On Adaptive And Fast Blind Denoising Algorithm

Most existing denoising algorithms are non-blind denoising algorithms,and have better performance.They usually depend on the internal parameter(i.e.noise level),which can characterize the noise level on noisy image,to obtain the best denoising performance.As the preprocessing module of the denoising algorithm,the estimation accuracy and execution efficiency of the NLE algorithms are two important indicators reflecting the performance of the denoising algorithm.However,most existing NLE algorithms are single-image based(SNLE)lacking available prior.Therefore,they have to design complex processes to predict more accurate noise level of noisy image,which leads to low execution efficiency and affects the efficiency of the denoising algorithm.To solve the problem that the denoising effect of existing non-blind denoising algorithms are limited by the noise level parameter,corresponding research was carried out.First,inspired by the NLE algorithms based on the convolutional neural network(CNN),we proposed a CNN and multi-image based NLE algorithm(CNNMNLE),which can extract convolutional features as noise level aware features(NLAF).Specifically,we extract several output features from the first full connection layer of the CNN regression model as NLAF vetor.Then,we constructed a enhanced BP neural network with weaker ones.Finally,the extracted NLAF features vector was directly mapped to the corresponding noise level.Its execution efficiency and prediction accuracy have been greatly improved due to CNNMNLE making full use of the artificial neural network.However,it still needs further improvement.To further improve the performance of NLE,we proposed a fast multi-image based NLE algorithm(FMNLE)based on the principal component analysis(PCA)and the deep belief networks(DBN).Based on the first several eigenvalues(sorted in ascending order)of the covariance matrix of raw patches are significantly correlated with the noise level of the noisy image,and the computation of extracting these eigenvalues is very small,so FMNLE selects several eigenvalues as NLAF features from these eigenvalues,which have strong ability to characterize noise level.To map the selected NLAF features to the corresponding noise level accurately,we designed a DBN network as the final noise level prediction model,which has strong mapping ability.Benefits from the fast execution of proposed DBN and the small computation of extracting eigenvalues using PCA,the execution efficiency of FMNLE is significantly improved compared with CNNMNLE.In addition,because the selected NLAF can better characterize the noise level of noisy image,the prediction accuracy is also further improved.So far,the denoising algorithms based on deep learning are non-blind algorithms,and have great advantages on denoising effect and execution efficiency.However,the data dependencec is the common disadvantage of this kind of algorithm.The prediction accuracy and execution efficiency of the proposed FMNLE are greatly improved.Therefore,it can solve data dependence with applying the FMNLE to these existing denoising algorithms which have good denoising performance.To further improve the denoising effect,we improved the network after analyzing the advantages and disadvantages of DnCNN.Then,we proposed the multiscale denoising convolutional neural network(MDnCNN).The proposed MDnCNN used convolution feature extraction module with multiple convolution kernels with different sizes to capture more image features,and solve the problem that DnCNN can not capture rich image features with single size of convolution kernels.In addition,the MDnCNN was combined with FMNLE and we realized an adaptive fast blind denoising algorithm.To verify the performance of proposed CNNMNLE,FMNLE and MDnCNN,we verified them on three commonly-used image datasets,i.e.,Setl2,Berkeley Segmentation dataset(BSD)and Waterloo exploration database,respectively.For CNNMNLE and FMNLE,we compared them with the state-of-the-art NLE algorithms on prediction accuracy and execution efficiency.For MDnCNN,we compared it with the state-of-the-art denoising algorithms on denoising effect and execution efficiency.Experimental results show that,the proposed FMNLE has better prediction accuracy and execution efficiency.As the preprocessing module of non-blind denoising algorithms,it outperforms others on prediction accuracy and execution efficiency.The proposed MDnCNN has some improvement in the denoising effect,and can achieve adaptive fast blind denoising based on the proposed FMNLE.