Developing Genetic Programming based Image Denoising Systems

Abstract

During acquisition or transmission, the visual quality of digital images is deteriorated due to the occurrence of impulse and speckle noises. These noises adversely effect various applications in image processing, pattern recognition, computer vision and medical imaging. Due to emerging imaging applications, recent trend is to develop application specific denoising systems. In this thesis, genetic programming (GP) based various denoising systems are developed for impulse and speckle noises. The proposed GP based evolutionary systems have effectively developed the domain specific denoising models that select the optimal informative features from the corrupted images. In the first phase of research, the genetic programming based mixed impulse denoising (GP-MID) system is developed to improve the visual quality of corrupted digital images. In this system, GP has optimally/near-optimally selects suitable statistical features to remove noise. In the second phase, the genetic programming based multi-type impulse denoising (GP-MuID) system is developed for corrupted digital images. This system has successfully removed salt & pepper, uniform impulse, mixed impulse and impulse burst noises, simultaneously. In the third phase, an advanced version of multi-gene genetic programming (MGGP) based biomedical image denoising (MGGP-BmID) system is developed to improve the visual quality of biomedical images. In the last phase, the multi-gene genetic programming based ultrasound image denoising (MGGP-UsID) system is developed to denoise speckle from ultrasound images. The improved performance of the GP based systems is obtained for diverse types of natural and biomedical images. The comparative analysis with existing approaches highlights the effectiveness of the proposed GP based evolutionary denoising systems. The improved denoising performance is achieved by the proposed GP based systems. It is because, during evolutionary learning process, the useful statistical features and primitive functions from a wider solution space are optimally/near-optimally combined to develop GP based intelligent noise detectors and estimators for image denoising problems.