Randomized Circle Detection Performance Based on Image Difficulty Levels and Edge Filters

Abstract

Circle detection is one of the fundamental problems in image processing fields. There are many algorithms used for circle detection, the most known algorithm is Circular Hough Transform (CHT), Randomized Hough Transform (RHT) and Randomized Circle detection (RCD). These algorithms are performing differently as they are highly depending on edge images. In this paper, we investigate the performance of RCD algorithm, by testing the algorithm on three levels of images. Simple images, which contain few edges, image without noise. Moderate images, images contain more edges than simple images, more circles and overlapped circles. Complicated images, images contain a high number of edges and more shapes and non-regular circles and may contains other non circular shapes. In addition, we investigate the influence and sensitivity of different types of edge detection operators on the performance of RCD algorithm. We tested the RCD algorithm using Sobel, Canny, Laplacian of Gaussian, Roberts and Prewitt edge detection operators on the above mentioned image levels. After that, we compare the performance of RCD results based on, runtime performance and accuracy (number of detected circles) for each image level using each edge detector. Experimental results shows that RCD algorithm has a good performance on simple and moderate images, and it can detect all circles using any edge detector, for complicated images, RCD cannot detect the ball (while the ball in motion case) in ball tracking image. RCD can detect the circle on shapes image using Canny and Prewitt and Roberts's edge detectors, and cant using Sobel and LOG detectors. For the future work, RCD algorithm needs some modification to deals with non-regular circles. © Springer-Verlag Berlin Heidelberg 2013.