Adaptive representations for image restoration

• Main Author: Yan, Ruomei
• Other Authors: Shao, Ling
• Published: University of Sheffield 2014
• Subjects: 621.3
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.619120

In the field of image processing, building good representation models for natural images is crucial for various applications, such as image restoration, sampling, segmentation, etc. Adaptive image representation models are designed for describing the intrinsic structures of natural images. In the classical Bayesian inference, this representation is often known as the prior of the intensity distribution of the input image. Early image priors have forms such as total variation norm, Markov Random Fields (MRF), and wavelets. Recently, image priors obtained from machine learning techniques tend to be more adaptive, which aims at capturing the natural image models via learning from larger databases. In this thesis, we study adaptive representations of natural images for image restoration. The purpose of image restoration is to remove the artifacts which degrade an image. The degradation comes in many forms such as image blurs, noises, and artifacts from the codec. Take image denoising for an example. There are several classic representation methods which can generate state-of-the-art results. The first one is the assumption of image self-similarity. However, this representation has the issue that sometimes the self-similarity assumption would fail because of high noise levels or unique image contents. The second one is the wavelet based nonlocal representation, which also has a problem in that the fixed basis function is not adaptive enough for any arbitrary type of input images. The third is the sparse coding using over-complete dictionaries, which does not have the hierarchical structure that is similar to the one in human visual system and is therefore prone to denoising artifacts. My research started from image denoising. Through the thorough review and evaluation of state-of-the-art denoising methods, it was found that the representation of images is substantially important for the denoising technique. At the same time, an improvement on one of the nonlocal denoising methods was proposed, which improves the representation of images by the integration of Gaussian blur, clustering and Rotationally Invariant Block Matching. Enlightened by the successful application of sparse coding in compressive sensing, we exploited the image self-similarity by using a sparse representation based on wavelet coefficients in a nonlocal and hierarchical way, which generates competitive results compared to the state-of-the-art denoising algorithms. Meanwhile, another adaptive local filter learned by Genetic Programming (GP) was proposed for efficient image denoising. In this work, we employed GP to find the optimal representations for local image patches through training on massive datasets, which yields competitive results compared to state-of-the-art local denoising filters. After successfully dealing with the denoising part, we moved to the parameter estimation for image degradation models. For instance, image blur identification uses deep learning, which has recently been proposed as a popular image representation approach. This work has also been extended to blur estimation based on the fact that the second step of the framework has been replaced with general regression neural network. In a word, in this thesis, spatial correlations, sparse coding, genetic programming, deep learning are explored as adaptive image representation models for both image restoration and parameter estimation. We conclude this thesis by considering methods based on machine learning to be the best adaptive representations for natural images. We have shown that they can generate better results than conventional representation models for the tasks of image denoising and deblurring.