Journal of Pattern Recognition Research

A direct inverse randomized Hough transform (DIRHT) is developed as a pre-processing procedure for incomplete ellipse detection in images with strong noise. A sample of five pixels is randomly drawn to solve for a set of five ellipse parameters. If there is a valid solution, a direct inverse transform maps the ellipse instance into a complete ellipse trace in an inverse image, by incrementing the intensities of all pixels on the ellipse. This process is repeated until sufficiently large number of valid samples is drawn. The DIRHT combines the advantages of both the inverse Hough transform and the randomized Hough transform (RHT), resulting in a number of desirable features. The intensities of all pixels on the underlying, complete ellipse trace were incremented so that the missing parts of the ellipse are restored. The DIRHT enhances the target ellipse by eliminating invalid samples (i.e. unrelated noise), and by restoring its missing parts. It converts a pixel-location classification problem into a pixel-intensity classification problem, which is essentially an image segmentation problem. This allows exclusion of noise pixels by using existing gray- level image segmentation techniques. These characteristics are demonstrated with both synthesized images and medical ultrasound images. Because the ellipse-restoration and ellipse-enhancement features, the DIRHT provides robustness against noise interference as a pre-processing procedure to RHT for incomplete ellipse detection. The final identified ellipses are highly accurate for both synthesized images and medical ultrasound images.