Redução de ruído Poisson com o algoritmo BM3D utilizando distâncias estocásticas

Abstract

Noise is a problem constantly present in digital images incoming from multiple sources, being important to study noise filters to improve the quality of images. When dealing with radiographic images, Poisson noise is an inherent problem that is more noticeable in a proportion that is inverse to the radiation dosage used for the same energy. Radiation, in high dosages, can be harmful to the patient's health, therefore the radiation dose reduction through new technologies is a desirable application. There are several techniques capable of filtering noise in digital images, which are most effective when directed to a specific type of noise. BM3D is a non-local image noise filter that works on Wavelet transform domain. The BM3D works by extracting similar small fragments in the image, called patches, having their comparison made by the Euclidean distance and grouped in sets of 3 dimensions. These patches are used for double filtering using hardthresholding on Wavelet domain and a Wiener filter, reducing noise and reconstructing the image. The Euclidean distance is very effective in comparing patches for AWGN noise (Additive White Gaussian Noise), however, it is inappropriate for Poisson noise and less effective when compared to stochastic distances. Based on this strategy, this work presents a technique for the reduction of Poisson noise with BM3D algorithm, replacing the calculation of the Euclidean distance by stochastic distances.