Curve detection using a new clustering approach in the Hough space

Abstract

Detecting curves (straight lines, circles, ellipses, etc.) from an image is one of the basic tasks in computer vision. The Hough transform (HT) and its variants have been the commonly used curve detecting methods. However, quantization of the Hough space has serious problems ranging from loss of accuracy to detection of artifacts due to false alignments in the image. Researchers have applied cluster analysis to Hough space to tackle these problems; however, good clustering performance relies heavily on the correct number of prototypes for the curves in question. Since we do not usually have adequate prior knowledge on the input image, to achieve the required performance, we need a robust clustering algorithm that can explore the data structure adaptively during the learning process. We apply a new clustering approach to the Hough space to detect curves from a binary image. It starts from a single prototype in the Hough space. During the learning process, it splits adaptively into multiple prototypes that identify the regions of highest density. Our extensive experimental results show that our algorithm has great advantages, such as high accuracy, low storage and infinite parameter space.