Abstract

Images are often corrupted by impulsive noise—a form of corruption in which a random number of pixels in the image are replaced with a uniformly-distributed random value. The problem of removing impulsive noise from an image has been the focus of much research. However, to date most of the existing techniques make no explicit distinction between noise and uncorrupt pixels, and affect the value of every pixel in the image, which results in image degradation. It has only been relatively recently that research has treated the problem of noise filtering as one of classification—identifying the pixels in an image that are noise, and only changing the values of these pixels. In this project, a supervised learning method based on Support Vector Machines is applied to the problem of impulsive noise detection in greyscale images. The generalisation performance and execution time of the Support Vector Machine classifier is improved by filtering the source image to remove background variation and to improve the separation of the noise and non-noise classes. A background-removal filter based on the median filter is presented, and a noise classifier that employs this filter is shown to perform as well in the detection of random-valued impulsive noise as existing techniques perform in the detection of two-valued salt-and-pepper noise.