Locally Adaptive Selection of Parameters in Regularization-based Denoising Algorithms

K. Someya and K. Kameyama (Japan)


denoising, regularization, adaptive parameter selection, neural network, total variation, texture preserving


Noise removal from an observed signal is an impor tant problem in signal processing. PDE-based methods have been widely used because of their common prop erty being good at removing the noise while preserving the edges. These methods restore images by modifying towards cartoon-like images. Therefore, other important features such as textures and certain details tend to be de graded in the denoising process. In this work, we pro pose a modiļ¬ed variational denoising algorithm that adap tively selects the parameters according to the local nature of the image. In order to estimate the locally appropri ate parameters, neural network based learning is employed. This method can adaptively change the denoising level by changing the regularization parameter and the smoothing kernel according to the local image. The results of denois ing by the proposed method show both visual and objective improvements.

Important Links:

Go Back