Unsupervised Polygonal Reconstruction of Noisy Contours by a Discrete Irregular Approach

In this paper, we present an original algorithm to build a polygonal reconstruction of noisy digital contours. For this purpose, we first improve an algorithm devoted to the vectorization of discrete irregular isothetic objects. Afterwards we propose to use it to define a reconstruction process of noisy digital contours. More precisely, we use a local noise detector, introduced by Kerautret and Lachaud in IWCIA 2009, that builds a multi-scale representation of the digital contour, which is composed of pixels of various size depending of the local amount of noise. Finally, we compare our approach with previous works, by considering the Hausdorff distance and the error on tangent orientations of the computed line segments to the original perfect contour. Thanks to both synthetic and real noisy objects, we show that our approach has interesting performance, and could be integrated into document analysis systems.