This thesis presents the research of my PhD, which is the study of the mathematical image processing techniques and their application. The key point for this work is the automation of these techniques. We begin by introducing a traditional segmentation algorithm, the Otsu’s method, to study the automatic image thresholding technique, where we develop a new automatic global thresholding method for distinguishing different types of cell, and apply the method in drug development industry for high content screening. Starting from the traditional statistics-based method, we then investigate a more mathematical model, the total-variation (TV) method for image denoising and deblurring problems. In traditional TV-based models, it is not easy to systemically provide a choice of parameters for the system. Inspired by the ideas from machine learning, we design a new learning-based TV model where the parameters can be derived automatically via optimal control. However, only one optimal solution is given by this model. Finally we combine our model with a newly-invented evolutionary algorithm where allows us to study all the possible optimal solutions and compare the differences they bring to the output images of our model. The experimental results have shown effectiveness on image denosing and deblurring problems comparing with both traditional and existing learning-based TV methods.
|Date of Award||2018|
|Sponsors||China Scholarship Council|
|Supervisor||Ping Lin (Supervisor) & Irene Kyza (Supervisor)|