This thesis studies the detection of elliptical shapes, including circles, ellipses and frontal faces, in gray-scale images.
Object detection is one of the fundamental problems in computer vision. It includes the detection of elliptical arcs and human faces. Many circular objects are frequently found in the real world, such as in automatic inspection and assembly. The oblique projection of a circle onto an image plane is an ellipse. Detection of faces is indispensable for face recognition systems and useful for man-machine interfaces. This is particularly so for the frontal face, i.e., the front view of a face, which is oval or elliptical.
Hough transform (HT) is one of the most popular methods used to detect analytic curves such as lines, circles and ellipses. But, although it is robust to noises, occlusions, shape distortions, etc., it suffers from large memory requirements and high computational loads. To overcome these limitations, many variants of HT have been developed. In this direction, the original contributions of this thesis are based mainly on a new Hough transform called ERHT, applied to ellipse detection and frontal face detection, and a new variant of standard Hough transform, applied to circle detection.
The proposed ERHT-based ellipse detector can detect ellipses efficiently with small memory requirements. The proposed circle detector is a two-step algorithm based on the concept of intersecting chords. It can efficiently detect simple circles in an image with very complex background. Both the circle detector and the ellipse detector work on the edge image extracted from the input image.
Two methods to detect frontal faces are proposed. The first one is a combination of a GHT-based ellipse detector and a neural network-based face detector. The second one is a combination of an ERHT-based ellipse detector and a neural network-based face detector. Both GHT-based and ERHT-based ellipse detectors are applied to locate the face candidates quickly. The neural network-based face/non-face classifier then verifies each face candidate. The applicability and efficiency of the two proposed methods are experimentally verified using representative face databases.