Image Enhancement Under Gaussian Impulse Noise for Satellite and Medical Applications

Introduction

Gaussian noise is the most common type of noise present in the image acquisitions systems. Since modelling Gaussian noise is simple and convenient, most of the optimization methods design the criterion function under this noise assumption. However, errors due to transmission errors, malfunctioning detector elements and faulty memory locations replace the underlying pixel intensity levels with random values. Since the intensity levels of these corrupted pixels are very different from the neighboring ones, they do not possess any information about the original pixel intensity or Gaussian corrupted pixels. Applying an algorithm designed for Gaussian noise to these kinds of images results in sub-optimal results. Since this is the result of impulse noise which heavily corrupt a limited number of pixels randomly, this random error cannot be properly modelled using Gaussian distribution. Since image denoising is a low-level computer vision task, its successful application helps in the accurate delivery of high-level vision-tasks like image classification, object detection, target tracking and semantic segmentation.

This chapter starts with the discussion on two different imaging domains corrupted by mixed Gaussian-impulse noise; namely hyperspectral imaging (HSI) and magnetic resonance imaging (MRI). Since Gaussian-impulse noise is the dominant character of noise in these imaging domains, image formation model under this noise assumption is imperative to appreciate the advantages obtained by applying denoising schemes specific to Gaussian-impulse noise. The next section talks about the image denoising schemes under four different solution strategies with emphasis on the types of Gaussian-impulse noise (i.e., fixed-valued or random-valued) and their limitations. In the last section, rigorous experimental evaluation on the two imaging domains; namely HSI and MRI are performed; followed by a conclusion.