Results

The impulsive noise classifier developed in this project was designed to be part of an impulsive noise removal filter—corrupt pixels in an image will be identified by a Support Vector Machine, and will be replaced with a value determined by a separate pixel interpolation component. During typical use, the Support Vector Machine will be used to identify noise in images other than the image on which it was trained. To determine how well a Support Vector Machine is able to detect noise in images other than the image used in training, a classifier was trained to identify noise in the image "Lena" and was used to identify noise in the remaining four images shown below.

The graph below shows the misclassification rate of a Support Vector Machine that has been trained on the image "Lena" when tested on noise detection in each of the five images above.

We were also interested in determining how well a Support Vector Machine would perform when the level of noise corruption in the test image differed to the level of noise in the training image. The graph below shows the misclassification rate of a Support Vector Machine that was trained on an image containing 2% impulsive noise, when tested on images containing between 1-20% impulsive noise.