Research On Fast Noise Level Estimation Algorithm Based On Deep Neural Network

The performance of most image denoising algorithms depends to varying degrees on the accurate estimation of noise level values in the noisy images to be processed.Recently,most noise level estimation algorithms were designed based on a single noisy image.Since a single noisy image can be used,the accuracy of noise level estimation is not high,especially for noisy images with rich texture details.In other words,some algorithms are designed to be more complex in order to achieve a more accurate estimation of the noise level in the noise image,making the execution time is too long,reducing the execution efficiency of the noise level estimation algorithm and further affecting the overall performance of the overall noise reduction algorithm.To overcome the shortcomings of the traditional noise level estimation algorithms,we proposed two fast noise level estimation algorithms based on deep neural network by adopting the implementation strategy based on multiple noisy images: a fast noise level estimation(FNLE)algorithm and an improved noise level estimation(I-FNLE)algorithm.The main difference between the FNLE algorithm and the I-FNLE algorithm lied in which feature value was extracted.The FNLE algorithm uses the Daubechies 9/7 wavelet basis to perform the wavelet transform on noise images in 3 directions and 3 directions.The 18-dimensional feature vector of 2 feature parameters on each subband coefficient value following Generalized Gaussian distribution(GGD)is extracted.To form the feature vector of an image,but the I-FNLE algorithm used the PCA technique to compute the covariance matrix on the set of raw patches and extracts the first 16-dimensional minimum eigenvalues of the redundant dimension to construct the feature vector describing the noise level of the image.Finally,both algorithms used DNN network technology to train the noise image in the training set to obtain the prediction model,and the eigenvalues extracted from the test image were directly mapped to the estimated values of the image noise level by the prediction model.Because in the process of feature extraction,I-FNLE algorithm was a feature vector that described the noise level of the image to be evaluated by directly selecting the first 16-dimensional minimum feature values of the redundant dimension of the original image block,so that the execution time of the I-FNLE algorithm was shorter and more efficient than that of the FNLE algorithm.To verify the noise estimation effect of the FNLE algorithm and the I-FNLE algorithm,comparative experiments were performed on the different image databases with the state-of-the-art algorithms,and to observe the actual use of FNLE algorithm and I-FNLE algorithm,the noise level estimation predicted by the FNLE algorithm and the I-FNLE algorithm was used as the parameter input of the BM3D(block-matching and 3D filtering)noise reduction algorithm,and eventually the two algorithms were used estimate the noise level.The noise reduction effect of the BM3 D algorithm with the noise level value was very close to the noise reduction effect of the BM3 D algorithm using the real noise value.Therefore,the problem that the BM3 D algorithm needed to manually determine and set the noise level parameter before the noise reduction processing can be solved.It also confirmed that the proposed algorithm had a certain degree of practicality.In short,a large number of experimental results show that the two fast noise level estimation algorithms based on deep neural network presented in this research are not only faster and more accurate than current classical noise level estimation algorithms,but also have certain practical value.