Accurate Detection of Ellipses based on Co-clustering

Abstract

Ellipse detection plays a vital role in computer vision, primarily due to the absence of efficient techniques to incorporate information from arcs in a global view. In this paper, we present a novel co-clustering-based approach that captures the global compatibility of arc pairs. Specifically, we compute the probability distributions of arcs and measure their similarity using Jensen-Shannon divergence to construct a compatibility matrix. Co-clustering extracts low-rank substruc-tures, each representing an ellipse. Experiments on synthetic images show a 10% improvement in accuracy and 15% in recall compared to previous methods. We also demonstrate performance on real images containing occlusions and overlapping ellipses from the Prasad dataset, the Coins dataset, the Tableware dataset and the Iris V dataset. Compared to techniques relying solely on local information , our method achieves higher accuracy in detecting small and occluded ellipses owing to its global perspective. The proposed technique provides an efficient, robust solution to ellipse detection across diverse scenarios.