Contour segment grouping for overlapping convex object segmentation

Abstract

Segmentation of overlapping convex object has many real-world applications, including material analysis and morphological analysis of biological cells. The methods for convex object segmentation usually consist of the following stages: 1) edge detection, 2) edge segmentation, 3) segment grouping, and 4) estimating the full contours of the objects. This thesis presents an overview of methods and approaches that are used for convex object segmentation. This thesis is focused on the segment grouping part. The novel proposed method is based on estimating a minimal area shell which covers all segments of the group. The proposed segment grouping method was compared with a current state-of-art segment grouping method on two types of data: synthetic data and real data. The synthetic data consist of images with different triangle, quadrilateral, and ellipse particles. The real data consist of nanoparticles images captured using transmission electron microscopy and marked manually. The results of experiments shown that the proposed segment grouping method shows improved results on synthetic data. However, the proposed method is worse on real data.