Introduction

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. Noise may be introduced into an image at acquisition, due to a noisy sensor, or during transmission, due to errors in the communications channel.

Original noise-free image (left), and with 10% impulsive noise (right).

It is clearly desirable to remove impulsive noise corruption from images, which not only improves the visual appearance of the image, but also to excludes corrupt pixels form further image processing.

This project treated impulsive noise removal as a classification problem—the pixels of an image were classified into two categories: pixels that represented impulsive noise corruption, and pixels that were valid (uncorrupt). Following pixel classification, any noise identified in the image can be replaced with an estimate of the original (uncorrupt) pixel's value. This project considered only the problem of identifying the corrupt pixels in an image; the problem of determining the value with which to replace corrupt pixels is a separate interpolation problem.

Pixel classification in this project was performed by a Support Vector Machine. A Support Vector Machine is a tool for supervised classification: to learn a particular classification function, a Support Vector Machine is given labelled training data (in the case of this project, the training data consisted of a set of classified pixels). Following training, a Support Vector Machine is able to classify previously unseen examples (pixels) based on the contents of the training dataset.