AG3line: Active grouping and geometry-gradient combined validation for fast line segment extraction

Abstract

Line segment detectors based on local image domain passively fit a line segment from a set of pixels, but no constraint on line geometry is set in the grouping process. Therefore, unstable pixels, such as the pixels in grass, clouds, or weak gradient edges, may cause false positives and fractures. This paper proposes the detector named AG3line, which employs an efficient active grouping strategy. In AG3line, the pixel for the next grouping is calculated actively with the line geometry and it can even be accurate to one pixel. To reduce the fracture caused by unstable pixels, when the adjacent pixel cannot satisfy the grouping rules, the candidate pixels for the next grouping are expanded with the line geometry constraint. To furtherly control false positives, AG3line then validates and refines the line segments by exploiting both the line geometry and the alignment of gradient magnitude. When AG3line was evaluated utilizing the image dataset with the ground truth, it outperformed both the classical and the latest detectors.The implementation of AG3line is available at https://github.com/weidong-whu/AG3line.