Abstract: Ellipse detection is a hot issue in the field of computer vision, and how to detect ellipses quickly and accurately, especially small ellipses with low resolution, is the main challenge of this problem. First, ellipses data sets of different sizes are constructed, and the region detection method in deep learning is introduced into the ellipse detection process. Furthermore, a two-stage arc filtering strategy is proposed based on quadratic curve constraints, which can effectively detect nested ellipses and reduce the number of missed and false positive detections. Finally, the bicubic interpolation method is employed to enlarge the area for small ellipses detection. The experimental results demonstrate the proposed method has a significant improvement in detection accuracy with competitive speed. Especially, on the small ellipse data set, the detection accuracy of the proposed method is more than two times of the traditional arc-based methods.