Impulse Noise Classification Using Machine Learning Classifier and Robust Statistical Features

Abstract

Identifying corrupted pixels in an image helps the denoising algorithm to perform better. In this paper, I propose a machine learning (ML) based classification algorithm which uses random decision forest classifiers and classical pixel-wise statistical parameters. The ML algorithm is trained to identify impulse noise by taking the computed statistical parameters from the corrupted image as input. From the experiment, it is clear that the identification of corrupted pixels depend on both the robustness of the chosen image parameters and the accuracy of the trained classifier. Particularly, random forest classifier which is an ensemble of random decision trees is employed as it suits better for this application. The implemented decision tree which is limited to 10, shows a better classification performance in the images corrupted with impulse noise. The improvement in the classification is better for random impulse noises rather than salt and pepper noises. Also, its significance is visible in low and medium noise intensities rather than high densities and hence I limit the noises by 50% for the discussion.