Intelligent image noise types recognition and denoising system using deep learning / Khaw Hui Ying

Khaw , Hui Ying (2019) Intelligent image noise types recognition and denoising system using deep learning / Khaw Hui Ying. PhD thesis, University of Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (183Kb)
	PDF (Thesis PhD.) Restricted to Repository staff only until 31 December 2021. Download (6Mb)

Abstract

Digital images may be seriously contaminated by noise during acquisition and transmission. Image denoising is a fundamental role to recover original image by making an estimation to preserve important image information. It is equally, if not more crucial, to characterize the type and level of the noises present in the images based on its nature and distribution, so that appropriate denoising tools can be applied, achieving a restored image of higher quality. The first objective of this work is to present a method to recognize image noise of different types: impulse, gaussian, speckle and poison noises, and their combination at different noise levels as a first step, and to ultimately perform denoising to recover noisy images. The main purpose is to take the advantage of Convolutional Neural Network (CNN) capability in analyzing image characteristics. To classify image noise type, the CNN trained with Backpropagation (BP) algorithm and Stochastic Gradient Descent (SGD) optimization technique are implemented. In order to reduce the training time and computational cost of the algorithm, Principal Components Analysis (PCA) pretraining strategy is deployed to obtain data adaptive filter banks. The proposed CNN with PCA for Noise Types Recognition (CPNTR) model is semi- supervised, because the PCA kernels are generated in an unsupervised way while the classifier at the output layer is trained by supervised learning. The designed system is validated by using images treated with noise of single and combination of various types. The experiments conducted have proven the reliability of the proposed CPNTR model by having achieved an overall average accuracy of 99.9% while recognizing four classes of image noise types and of 99.3% while recognizing types and levels of noise. The model has also accurately identified 85.7% of images that are degraded by combinations of noise. These have shown that the proposed CPNTR structure provides an excellent image noise types classification solution. The second objective is to design an efficient CNN with Particle Swarm Optimization (PSO) model for high-density impulse noise removal. The proposed High-density Impulse Noise Detection and Removal (HINDR) model mainly consists of two parts: the impulse noise removal and impulse noisy pixel detection for restoration. Unlike traditional methods that usually start with detection and followed by denoising, the model initially leverages the powerful ability of deep CNN architecture to separate noise from noisy image, then adopts PSO to pinpoint the most optimized threshold values for detecting impulse noisy pixels. An ensemble of these algorithms is an intelligent and adaptive solution, producing a clean output, while preserving significant pixel information. Targeting on impulse noise density as high as 50%, 60%, 70%, 80% and 90%, the model has been trained with a massive collection of natural images and 14 standard testing images are used for validation purposes. Based on the final denoised images, the model has proven its reliability, in terms of both visual quality and quantitative evaluation.

Actions (For repository staff only : Login required)