Noise-robust line detection using normalized and adaptive second-order anisotropic Gaussian kernels

Abstract Line detection is an essential procedure for many tasks in computer vision. Although literature is rich in proposals for this topic, most of existing methods are vulnerable to noise, especially when lines are fairly narrow. In order to design a line detector that minimizes the impact of noise, regardless of the scale or direction of the lines, in this paper, we present a framework for multiscale line detection based on second-order anisotropic Gaussian kernels. Firstly, we model a line segment using a directional Gaussian function. Secondly, with the help of a newly proposed normalization method, we enable the second-order anisotropic Gaussian kernels to quantitatively measure the line prominence as well as the line scale. Subsequently, based on a noise-robustness analysis in terms of the signal-to-noise ratio, an adaptive anisotropy factor is proposed. By incorporating postprocessing techniques, an automated line detector using the normalized and adaptive second-order anisotropic Gaussian kernels is developed. The performance of the proposed method is quantitatively evaluated by comparing it with five competing methods on a publicly available dataset.