Combining contextual and local edges for line segment extraction in cluttered images

TL;DR

This paper introduces a novel line segment extraction method combining contextual and local edges, improving robustness and completeness in cluttered images where traditional methods often fail.

Contribution

It proposes a new approach that combines large and small footprint image derivatives, using a statistical test for contextual edges, to enhance line segment detection in noisy, cluttered scenes.

Findings

Effective in extracting complete line segments of various lengths and widths.

Robust to noise and clutter, outperforming traditional methods.

Works well on both synthetic and real images.

Abstract

Automatic extraction methods typically assume that line segments are pronounced, thin, few and far between, do not cross each other, and are noise and clutter-free. Since these assumptions often fail in realistic scenarios, many line segments are not detected or are fragmented. In more severe cases, i.e., many who use the Hough Transform, extraction can fail entirely. In this paper, we propose a method that tackles these issues. Its key aspect is the combination of thresholded image derivatives obtained with filters of large and small footprints, which we denote as contextual and local edges, respectively. Contextual edges are robust to noise and we use them to select valid local edges, i.e., local edges that are of the same type as contextual ones: dark-to-bright transition of vice-versa. If the distance between valid local edges does not exceed a maximum distance threshold, we enforce connectivity by marking them and the pixels in between as edge points. This originates connected edge maps that are robust and well localized. We use a powerful two-sample statistical test to compute contextual edges, which we introduce briefly, as they are unfamiliar to the image processing community. Finally, we present experiments that illustrate, with synthetic and real images, how our method is efficient in extracting complete segments of all lengths and widths in several situations where current methods fail.