Shape-Aware Ellipse Detection Via Parametric Correlation Learning

Abstract

Elliptical object detection, a pivotal challenge in computer vision and pattern recognition, has traditionally been addressed through parameter regression methods, overlooking the unique geometric characteristic and intricate parameter interactions of ellipses. Existing approaches often yield imprecise detections, especially for small or partially occluded ellipses. Our study proposes a novel shape-aware ellipse detection network (EDNet) that leverages the shape characteristics of elliptical objects to capture their fundamental geometric structure, moving beyond mere reliance on internal textures.EDNet incorporates a comprehensive analysis of ellipse parameters during training, refining the loss function to establish a robust relationship between the ellipse center and other parameters, thereby strengthening the model's constraints. Additionally, we introduce a LoG-like Edge Detection Module (LEDM) coupled with an Edge Guided Module (EGM), enabling the precise extraction of ellipse boundaries while augmenting multi-scale features for boundary details. Furthermore, to enhance the accuracy of occluded ellipse detections, we incorporate an auxiliary component dedicated to estimating ellipse vertexes.Through rigorous experiments on two widely recognized datasets, EDNet exhibits significant improvements, achieving an average increase of 6% and 10% in detection accuracy compared to leading state-of-the-art models.

Keywords: Ellipse detection, Object detection, Edge detection