Shape matching for object detection using circular arcs and lines

Abstract

The detection of objects using shape information is an important problem in the computer vision field because such detection would be robust against variations in the illumination, color, and texture of images. In this dissertation, we propose a novel object detection algorithm using a single sketch. The proposed algorithm extracts circular arcs from image edges to describe shape information. As a building block of shape patterns, circular arcs are advantageous for efficient encoding, effective expression, and feasible reconstruction from damages. We designed a practical arc extraction algorithm using split-and-merge strategy. The arc extension robustly reconstructs the extracted curves that have been damaged or distorted by clutters and artifacts. It is possible to build arbitrary shape patterns using fewer circular arcs than line segments. Compared to random fragments and curves, circular arcs can be efficiently described and compared. Broken and poorly extracted circular arcs are fixed and refined within affordable computation cost. In addition to heuristic circular arc extraction algorithm, a biologically plausible circular arc extraction model was presented as a neural oscillator network with modified association field. The core of the proposed method is to overcome the limitation of conventional association field in detecting circular arcs by adding delicate inhibition and co-linearity constraint. An oscillatory network with the modified association field connectivity groups edge pixels on a circular arc by synchronized neural activity. The proposed network reliably detects circular arcs against partial missing of edge pixels. In addition, it can detect straight lines or salient curves by minor modifications. A simple but powerful shape description method in terms of circular arcs was proposed. The end point arc descriptor captures the spatial relation between end points of a circular arc segment. Two types of arc features were defined by combining...