Background
Segmentation or contour estimation of overlapping objects is an image analysis task. This task is connected to the problem of analyzing 2D projections of 3D objects. It is widely used in the industry and biology. Usually, it is difficult to estimate inner contours of the overlapped object so the segmentation methods must relay just on visible parts of particles (see figure 2). To solve such problems one must estimate the full contour based on visible edge fragments and prior knowledge about the object shape [26].
This work focuses on segmentation of convex objects. The work continues earlier research where a framework to segment (estimate contours) of partially overlapping nanoparticles was developed [25]. The framework consists of three steps: 1) detecting of concave edge points, 2) grouping of the resulting edge segments to form contour evidence, and 3) estimating the full contours of the objects (see fig. 2) [24, 26].In order to be able to estimate full contours of objects with partially observed edges, all edge points or edge segments belonging to the same object need to be grouped. To do this, shape analysis of the resulting object is needed. This can be done by employing a grouping method that defines how likely two edge segments belong to the same object, that is how well the resulting object fits the prior information about the object shapes or contour model.
Objectives and delimitations
The aim of this master’s thesis is to develop an efficient grouping strategy that grouping the contour segments which belong to the same object.
The objectives are as follows:
1. Make an overview of existing methods and frameworks for overlapping object segmentation.
2. Propose the new segment grouping method, that improves the performance of segment grouping on images with shapes of different types.
3. Compare the proposed method with existing state-of-art segment grouping method on real and generated data.
Structure of the thesis
The rest of thesis has the following structure. Chapter 2 gives a brief overview of an existing contour segmentation methods. Chapter 3 presents a segmentation framework with a new proposal of a segment grouping method. Chapter 4 contains the information about experiments and validation of the proposed segment grouping method. Chapter 5 discusses the findings and describes goals of the further research. Chapter 6 concludes the thesis and give a brief overview of the problem, the solution, and the results.
In this work methods for segment grouping were studied and a new segment grouping framework was proposed. A survey of existing segmentation methods was made in order to understand which methods can be used for the tasks of segmentation and segment grouping. There are two main approaches for segmentation. The one is based on seed-points detection and the other is based on concave points extraction. The proposed framework is based on the Zafari BB algorithm [26] with a proposed cost function. The cost function is based on a hypothesis that objects have a convex form and the shape is similar to one of the next type of shapes: ellipse, quadrilateral, or triangle. The cost function was compared with the original BB cost function [26] on the real and the synthetic data. The results of the experiments showed that the proposed cost function outperforms the original cost function on the synthetic data. The experiments with the real data showed that the proposed cost function is worse than Zafari BB algorithm. The main reason for this is the fact that the most particles in the real images have elliptical a symmetrical shapes.
