Optimisation for image processing

The main purpose of optimisation in image processing is to compensate for missing, corrupted image data, or to find good correspondences between input images. We note that image data essentially has infinite dimensionality that needs to be discretised at certain levels of resolution. Most image processing methods find a suboptimal solution, given the characteristics of the problem. While the general optimisation literature is vast, there does not seem to be an accepted universal method for all image problems. In this thesis, we consider three interrelated optimisation approaches to exploit problem structures of various relaxations to three common image processing problems: 1. The first approach to the image registration problem is based on the nonlinear programming model. Image registration is an ill-posed problem and suffers from many undesired local optima. In order to remove these unwanted solutions, certain regularisers or constraints are needed. In this thesis, prior knowledge of rigid structures of the images is included in the problem using linear and bilinear constraints. The aim is to match two images while maintaining the rigid structure of certain parts of the images. A sequential quadratic programming algorithm is used, employing dimensional reduction, to solve the resulting discretised constrained optimisation problem. We show that pre-processing of the constraints can reduce problem dimensionality. Experimental results demonstrate better performance of our proposed algorithm compare to the current methods. 2. The second approach is based on discrete Markov Random Fields (MRF). MRF has been successfully used in machine learning, artificial intelligence, image processing, including the image registration problem. In the discrete MRF model, the domain of the image problem is fixed (relaxed) to a certain range. Therefore, the optimal solution to the relaxed problem could be found in the predefined domain. The original discrete MRF is NP hard and relaxations are needed to obtain a suboptimal solution in polynomial time. One popular approach is the linear programming (LP) relaxation. However, the LP relaxation of MRF (LP-MRF) is excessively high dimensional and contains sophisticated constraints. Therefore, even one iteration of a standard LP solver (e.g. interior-point algorithm), may take too long to terminate. Dual decomposition technique has been used to formulate a convex-nondifferentiable dual LP-MRF that has geometrical advantages. This has led to the development of first order methods that take into account the MRF structure. The methods considered in this thesis for solving the dual LP-MRF are the projected subgradient and mirror descent using nonlinear weighted distance functions. An analysis of the convergence properties of the method is provided, along with improved convergence rate estimates. The experiments on synthetic data and an image segmentation problem show promising results. 3. The third approach employs a hierarchy of problem's models for computing the search directions. The first two approaches are specialised methods for image problems at a certain level of discretisation. As input images are infinite-dimensional, all computational methods require their discretisation at some levels. Clearly, high resolution images carry more information but they lead to very large scale and ill-posed optimisation problems. By contrast, although low level discretisation suffers from the loss of information, it benefits from low computational cost. In addition, a coarser representation of a fine image problem could be treated as a relaxation to the problem, i.e. the coarse problem is less ill-conditioned. Therefore, propagating a solution of a good coarse approximation to the fine problem could potentially improve the fine level. With the aim of utilising low level information within the high level process, we propose a multilevel optimisation method to solve the convex composite optimisation problem. This problem consists of the minimisation of the sum of a smooth convex function and a simple non-smooth convex function. The method iterates between fine and coarse levels of discretisation in the sense that the search direction is computed using information from either the gradient or a solution of the coarse model. We show that the proposed algorithm is a contraction on the optimal solution and demonstrate excellent performance on experiments with image restoration problems.Open Acces