Research For Interactive Image Segmentation

Computer vision is an important branch of image signal processing, it is multi-disciplinary in that it involves multiple areas such as signal processing, computer science, mathematics and psychology, and it has important applications in military, manufacturing, medical industries as well as commercial image and graphics software. Traditional computer vision algorithms are simply regarded as an automatic black-box model, i.e. the algorithm automatically computes a solution based on the input data, without any human intervention during the whole process. However, for many problems with high-level semantics, e.g. image and video segmentation, a fully automatic algorithm cannot yet achieve satisfactory results since it is still quite difficult for a computer to understand those semantics. Therefore, interactive image processing was developed in recent years, and has been receiving more and more attention.The biggest advantage of interactive image segmentation, compared to tradition segmentation methods, is that with the guidance of the user, problems which previously were ambiguous in a semantic sense, now become clearly defined and solvable. With carefully designed algorithms to support the user interface (UI), we can efficiently and accurately solve the segmentation problem, on which traditional algorithms took a much longer time to produce a much coarser result.This paper extensively studies the basic principles of interactive image segmentation, its development and current deficiencies, and makes improvements over the current state-of-the-art algorithms in the following ways:1. Based on careful studies of the Lazy Snapping algorithm and Grab Cut algorithm, we propose a modified optimization method to tackle the limitations of the two methods.2. We study the geodesic distance theory, which in computer vision is applied in gray-scale image colorization and texture-less color image segmentation (the Grow Cut), and improve the Grow Cut algorithm to a substantial extent, so that the algorithm is more robust and more importantly, can tackle textures in interactive image segmentation.