A robust circle detector with regionalized radius aid

Abstract

The detection of circles in geometric shapes is highly valued in computer vision and pattern recognition. Conventionally, the least-squares fitting is sensitive to occlusion or noise and prone to false circles. Therefore, this paper proposes a novel algorithm for robust circle detection using the least-squares fitting method combined with regionalized radius aid on the arc and chord lengths. To reduce edge noise impact, we present an edge pruning method to prune non-curve edge branch ports. Furthermore, we extract arcs based on the inflection points and sharp corners of the approximate line segment. Next, curves that belong to the arcs obtain circle parameters according to the regionalized radius aided the least-squares method. Then, valid circles are obtained by considering two distance deviations to verify the candidate circles. Finally, valid arcs that belong to the same circle are combined and refitted, wherein the most representative arc is used as the basis for the refitting of all arcs. All experiments are conducted on real images from four publicly diverse datasets (one of them is the one we built). The detection results are compared with those of representative state-of-the-art algorithms, and the proposed algorithm has several advantages based on the comparison results: more robust to noise, effective rejection of false circles, and better performance.