MPG-LSD: A high-quality line segment detector based on multi-scale perceptual grouping

Abstract

Existing methods that rely on a single Gaussian scale to detect line segments could yield poor line continuity and inferior orientation and position accuracy due to insufficient suppression of quantization noise inherent in digital images. To address this fundamental issue, a novel multi-scale perceptual grouping-based line segment detector (MPG-LSD) is proposed in this paper. Our multi-scale perceptual grouping is developed to identify and aggregate collinear pixels that have similar gradient orientations for producing line segment candidates, not only over the input image but also across a set of Gaussian scales. Multi-scale line segment refinement and validation are then further developed and implemented to produce the final detection result, which delivers high quality in terms of line continuity, orientation and position accuracy. To enrich the evaluation of line segment detection performance, a new dataset consisting of high-resolution and natural noise-corrupted images with line segment annotations is constructed. Extensive experimental results show that our proposed MPG-LSD can outperform the current state-of-the-arts by a large margin.