Image Restoration using Soft Computing

Abstract

Image restoration is defined as a technique to bring back contents of initial image from the degraded one with or without having prior knowledge about degradation process. Image degradation may appear in any of the processing phases including: acquisition, reproduction, storage, transmission, compression, and/or pre-processing phase. In this thesis, research has been presented in the domain of image restoration. Image restoration offers exciting application and research opportunities for many application domains e.g. astronomical imaging, medical imaging, defense applications along with numerous other post-processing techniques. A degradation model needs to be developed first for image restoration. This degradation is usually a consequence of addition of blur or noise. Once the degradation has been modeled, images can be restored towards the approximation of their reality. However, in reality, a-priori information about the blur or noise is not known in many situations thereby making image restoration a more difficult task. In order to overcome this problem, estimation techniques are used on true image as well as degraded one to get a better approximation. Major focus of this thesis has been to study existing modern soft computing based techniques and to develop new image restoration techniques using soft computing. Research centers around restoration of corrupted images subjected to different types of impulse noise for grayscale as well as for color images. This restoration is achieved by making good tradeoff between two essential but contradictory characteristics of images i.e. smoothness and edge preservation. This dissertation makes the following contributions in the field of image restoration: (1) Machine Learning based Impulse Noise Detector is proposed which has high noise detection accuracy and very few false alarms for whole range of noise density. (2) Directional Weighted Switching Median filter is proposed which performs well even at high noise density and has outstanding detail preservation capability. (3) Fuzzy based filter using statistical estimators is proposed for random-valued impulse noise. (4) Color differences based fuzzy color filter is proposed which preserves the color differences of the image. (5) In the end, a novel technique is proposed for noise type identification making automatic image restoration techniques to be more generic and useful. Significant experimentation is performed to evaluate performance of proposed techniques with the impressive results. This experimentation is based upon both qualitative and quantitative error measures.