Learning-based geometric framework for noise discernment and denoising of 2D point sets

Abstract

Denoising, the recovery of ground-truth point sets from noisy inputs, is essential for real-world applications like sketch-to-vector conversion, sketch skeletonization/thinning, curve reconstruction from images, and scanned point clouds, etc. In practice, the nature of noise in these applications varies significantly. However, the widely used noise modeling for 2D pointsets falls under two categories, viz. (a) as a perturbed one, (b) as an offset-based approach. Hence, a single denoising method is often inadequate. In this paper, we propose a novel approach based on Delaunay triangulation (DT) for discernment as well as denoising of a point set. Using a developed dataset, a learning-based framework that derives features from DT is proposed for the discernment of a point set. Further, we propose different denoising approaches based on the classification. Experimental results show that our method effectively classifies and denoises diverse point sets, including real data, outperforming state-of-the-art techniques.