Poisson NL Means: Unsupervised Non Local Means For Poisson Noise
Charles-Alban Deledalle,Florence Tupin,Loic Denis, 17th International Coonference on Image Processing, Pg: 801-804,Sep 2010.
This paper focuses on images corrupted by Poisson noise . The proposed method denoises the image by using a pre-filtered image which adapts to the statistics of the poisson noise. Two filtering parameters α and β are tuned where an automatic setting of the parameters is proposed which is based on the estimator.
NL means algorithm reduces the noise as well as preserves the structures, where the similar pixels are averaged . This paper focuses on patch similarity defined by the euclidean distance. The setting of the two parameters is very critical for poisson noise , so a risk estimator is used for automatic setting of the parameters. SURE (Stein's Unbiased Risk Estimator) estimator is used for additive white gaussian noise and for poisson noise a PURE (Poisson Unbiased Risk Estimator) estimator is used. To reach the optimal parameters in few iterations, Newton's method is used to iteratively refine α and β. Newtons method in few iterations finds the best tradeoff between the information brought by the noisy image and the pre-estimated image to define the weights. If the value of β is high, then the pre-estimated iamge has poor quality, where the weights ll be determined from the noisy image. whereas, if the value of α is high, the pre-estimated image has a high quality and therefore the weights ll be determined from the pre-estimated image only. The pre-estimated image is obtained by the moving average filter.
The Poisson NL means filter provides good result with few processing artifacts.
Software:
NL means algorithm reduces the noise as well as preserves the structures, where the similar pixels are averaged . This paper focuses on patch similarity defined by the euclidean distance. The setting of the two parameters is very critical for poisson noise , so a risk estimator is used for automatic setting of the parameters. SURE (Stein's Unbiased Risk Estimator) estimator is used for additive white gaussian noise and for poisson noise a PURE (Poisson Unbiased Risk Estimator) estimator is used. To reach the optimal parameters in few iterations, Newton's method is used to iteratively refine α and β. Newtons method in few iterations finds the best tradeoff between the information brought by the noisy image and the pre-estimated image to define the weights. If the value of β is high, then the pre-estimated iamge has poor quality, where the weights ll be determined from the noisy image. whereas, if the value of α is high, the pre-estimated image has a high quality and therefore the weights ll be determined from the pre-estimated image only. The pre-estimated image is obtained by the moving average filter.
The Poisson NL means filter provides good result with few processing artifacts.
Software: