Detail-preserving image information restoration guided by SVM based noise mapping

Abstract

In this paper, we propose a new method to solve the problem of fixed-valued impulsive noise reduction in images. Nonlinear filter like the median filter (MF) is useful for reducing random noise and periodical patterns, but direct median filtering have undesirable side effects such as smoothening of noise free regions, which results in loss of image detail and distortion of the signal. Impulse noise is suppressed by selectively filtering the contaminated signal regions only, thus minimizing distortion of clean passages and loss of high frequencies. In the first phase, support vector machines (SVM) are used to segment the set of pixels N that are likely to be contaminated by the mixed impulses. In the second phase, the image is restored by employing a combination of the best neighborhood match filter (BNM) and the modified multi-shell median filter (MMMF) to these segmented regions. This method combines the effectiveness of the best neighborhood matching (BNM) filter in suppression of the noise components while adapting itself to the local image structures, and the edge and finer image detail preserving characteristics of the MMMF. To support our proposed method, numerical results are also provided, which indicate that the filter is extremely useful for preserving edges or monotonic changes in trend, while eliminating short duration impulses of high density.